Full text of Economic Policy Review (Federal Reserve Bank of New York) : Volume 16, Number 2

View original document

The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.

Rajashri Chakrabarti

Program Design, Incentives,
and Response: Evidence from
Educational Interventions
• In an effort to reform K-12 education,
policymakers have introduced vouchers in
some U.S. school districts, enabling students
to transfer from public to private schools.
• The different designs of two school voucher
programs—the Milwaukee and Florida
programs—have had different effects on
public school incentives and performance.
• In Milwaukee, vouchers were imposed from
the outset; in Florida, schools were first
threatened with vouchers and thus had
an incentive to avoid them.
• The Florida public schools’ efforts to avoid
vouchers resulted in performance effects that
far exceeded those of Milwaukee’s program.
• Program design is critical: Policies that
present failing public schools with functional
and credible sanctions are best suited to
provide the results intended by policymakers.

Rajashri Chakrabarti is an economist at the Federal Reserve Bank of New York.
rajashri.chakrabarti@ny.frb.org

1. Introduction

oncerns that U.S. students are not performing as well
as their counterparts in other developed countries on
international math and science tests have led to widespread
demands for the reform of K-12 education in the United States.
Of the various reforms under consideration, school voucher
reform is at the forefront.
Vouchers are scholarships that make students eligible to
transfer from public to private schools. A basic feature of all
publicly funded voucher programs in the United States is the
funding of vouchers by public school revenue, so that money
always “follows” students. In other words, schools that lose
students lose their corresponding funding. Schools therefore
recognize the financial implications of vouchers and have an
incentive to avoid being subject to voucher programs.
This article investigates the role of program design in the
context of two such educational interventions in the United
States—the Milwaukee and Florida school voucher
programs—and analyzes the effects of design on public school
incentives and performance.1 We demonstrate that variations
in program design have markedly different outcomes for public
schools affected by vouchers.
The Milwaukee program, introduced in 1990, was the first
voucher program in the country. Implemented in 1999, the

The author thanks Andrew Haughwout for helpful suggestions. Noah Schwartz
provided excellent research assistance. All errors are the author’s. The views
expressed are those of the author and do not necessarily reflect the position
of the Federal Reserve Bank of New York or the Federal Reserve System.

FRBNY Economic Policy Review / October 2010

Florida program was the nation’s third, following Cleveland’s.
The Milwaukee and Florida voucher programs share the basic
feature of funding by public school revenue. But there are
crucial differences. Milwaukee’s is a means-tested program
targeting low-income students while Florida’s embeds a
voucher program in a full-fledged accountability system.
Using test-score data from Milwaukee and Florida and
implementing a difference-in-differences estimation strategy,
our study estimates the impact of each program by comparing
the post-program results of the affected schools with a
comparable set of control schools. Controlling for potentially
confounding pre-program time trends and post-program
common shocks, we find that the performance effects of the
Florida program far exceed those of Milwaukee’s program.
These results are quite robust in that they hold after controlling
for other confounding factors, such as mean reversion and a
possible stigma effect; they also withstand several sensitivity
tests.
Our findings have important policy implications, which we
consider in the context of New York State’s federal, state, and
city accountability programs. These programs include New
York City’s accountability policy, known as the “Progress
Report” policy, and the federal No Child Left Behind (NCLB)
law, as implemented by New York State.
Our study is organized as follows. Section 2 describes the
Milwaukee and Florida voucher programs. In Section 3, we
discuss the incentives created by the programs and the
corresponding responses that might be expected from the
affected public schools. Our data and empirical strategy are
reviewed in Sections 4 and 5, respectively. Section 6 presents
our results, and Section 7 considers policy implications.

2. Institutional Details
The first publicly funded school voucher program in the
United States was established in Milwaukee, Wisconsin, in
1990. The Milwaukee Parental Choice Program made the city’s
entire low-income public school population eligible for
1
Our study focuses on the impact of alternative voucher designs on public
school performance. A growing body of literature analyzes the many issues
associated with school vouchers. Nechyba (1996, 1999, 2000, 2003) analyzes
distributional effects of alternative voucher policies in a general equilibrium
framework; Epple and Romano (1998, 2002) and Chakrabarti (2009)
investigate sorting attributable to vouchers; Manski (1992) considers the
impact of vouchers on public school expenditure and social mobility; and
McMillan (2004) and Chakrabarti (2008b) model the quality of public schools
facing vouchers. Hoxby (2003a, b) and Chakrabarti (2008a) study the effects of
the Milwaukee voucher program, while Greene (2001), Greene and Winters
(2003), Figlio and Rouse (2006), West and Peterson (2005), and Chakrabarti
(2007, 2008a) study the effects of the Florida program.

Program Design, Incentives, and Response

vouchers. Specifically, starting in the 1990-91 school year,
the program made all Milwaukee public school students with
family income at or below 175 percent of the poverty line
eligible for vouchers to attend nonsectarian private schools.
In contrast, the Florida Opportunity Scholarship Program,
introduced in 1999, can be looked upon as a “threat-ofvoucher” program. Here, failing public schools were
threatened with the imposition of vouchers, with vouchers
implemented only if schools failed to meet a governmentdesignated cutoff quality level. The institutional details of the
Milwaukee and Florida programs are summarized in Table 1.
The Florida Department of Education classified schools
according to five grades: A, B, C, D, or F. The state assigned
school grades based on Florida Comprehensive Assessment

The major difference in program design
between the Milwaukee and Florida
programs is that in Milwaukee vouchers
were imposed at the outset, whereas in
Florida failing schools were first threatened
with vouchers, with vouchers introduced
only if the schools failed to show adequate
improvement in performance.
Test (FCAT) reading, math, and writing scores. For FCAT
reading and math, it categorized students into five achievement
levels—1 lowest, 5 highest—that correspond to specific
ranges on the raw-score scale. Using current-year data, the
Department of Education assigned an “F” grade to a school if it
was below the minimum criteria in reading, math, and writing;
a “D” if it was below the minimum criteria in one or two of the
three subject areas; and a “C” if it was above the minimum
criteria in all three subjects, but below the higher performing
criteria in all three. In reading and math, at least 60 percent
(50 percent) of students had to score level 2 (3) and above;
in writing, at least 50 percent (67 percent) had to score 3 and
above to meet the minimum (high-performing) criteria in
that respective subject.2
Under the Florida Opportunity Scholarship Program,
all public school students became eligible for vouchers, or
2

In 1999, seventy-eight schools received an “F” grade. Students in two of those
schools became eligible for vouchers. In 2000, four elementary schools received
an “F,” although none became eligible for vouchers. In 2001, no schools
received an “F” grade. In 2002, sixty-four schools received an “F.” Students in
ten of those schools became eligible for vouchers. In 2003, students in nine
schools became eligible for vouchers; in 2004, the figure was twenty-one
schools.

Table 1

Comparison of Milwaukee and Florida Voucher Programs
Milwaukee Program

Florida Program

First U.S. voucher program

 Third U.S. voucher program

 Started in 1990-91 school year

 Started in 1998-99 school year

 Public school students with family income at or below 175 percent
of the poverty line eligible for vouchers to attend nonsectarian
private schools

 Vouchers contingent on school performance

 Private schools were not permitted, by law, to discriminate against
students who apply with vouchers:
– Had to accept all students unless oversubscribed
– If oversubscribed, had to choose students randomly
 Average voucher amount equaled the state aid per pupil, and vouchers
were financed by an equivalent reduction of state aid to the school district
 1990-91 and 1996-97:
– Average voucher amounts were $3,346
– Vouchers as a percentage of total revenue per pupil were 45.23 percent

 Schools classified according to five grades: A, B, C, D, F
(A-highest, F-lowest)
– Grades based on the Florida Comprehensive Assessment Test (FCAT)
reading, math, and writing scores
– F, if below the minimum criteria in reading, math, and writing
– D, if below the criteria in one or two of the three subjects
– C, if above the minimum criteria in all three subjects, but below
the higher performing criteria in all three
 Students categorized into five achievement levels in FCAT reading
and math (1-lowest, 5-highest)
 Minimum criteria:
– Reading and math: at least 60 percent must score level 2 and above
– Writing: at least 50 percent must score level 3 and above
 High-performing criteria:
– Reading and math: at least 50 percent must score level 3 and above
– Writing: at least 67 percent must score level 3 and above
 All students of a public school became eligible for vouchers if the school
received two “F” grades in a period of four years
 Private schools were not permitted, by law, to discriminate against
students who apply with vouchers
– Had to accept all students unless oversubscribed
– If oversubscribed, had to choose students randomly
 Average voucher amount equaled the state aid per pupil, and vouchers were
financed by an equivalent reduction of state aid to the school district
 1999-2000 and 2001-02:
– Average voucher amounts were $3,330
– Vouchers as a percentage of total revenue per pupil were 41.55 percent

Source: Information and data provided in various Florida Department of Education and Milwaukee Department of Public Instruction reports.

“opportunity scholarships,” if the school received two “F”
grades in a period of four years. Therefore, a school that
received an “F” for the first time was exposed to the threat of
vouchers, but did not face them unless and until it got a second
“F” within the next three years. Thus, the major difference in
program design between the Milwaukee and Florida programs
is that in Milwaukee vouchers were imposed at the outset,
whereas in Florida failing schools were first threatened with
vouchers, with vouchers introduced only if the schools failed
to show adequate improvement in performance.
Apart from the above differences, the design of the two
programs was strikingly similar. In both programs, private
schools could not, by law, discriminate against students who
applied with vouchers—the schools had to accept all students
unless they were oversubscribed, in which case they had to
choose students randomly. Indeed, the application form did
not ask questions about the student’s race, sex, parents’
education, past scores, or prior records (for example, truancy,
violence). The questions were specifically worded only to

ascertain whether the student was eligible for the program.3
The system of funding for the Milwaukee and Florida voucher
programs was also very similar. Under each program, the
average voucher amount was equal to the state aid per pupil,
and vouchers were financed by an equivalent reduction of state
aid to the school district. Thus, state funding was directly tied
to student enrollment, and enrollment losses due to vouchers
were reflected in a revenue loss for the public school.4 The
average voucher amounts under the Milwaukee (1990-91
through 1996-97) and Florida (1999-2000 through 2001-02)
programs were $3,346 and $3,330, respectively. During these
periods, vouchers as a percentage of total revenue per pupil
were 45.23 percent in Milwaukee and 41.55 percent in Florida.

3
While the schools could not employ any selection criteria for the voucher
students, this was not the case for nonvoucher students in the same school.
Also note that the private schools had the choice of whether to participate in
the program. However, if they decided to participate, they were required by law
to accept all students or to choose students randomly, if oversubscribed.

FRBNY Economic Policy Review / October 2010

3. Discussion: Effects of the
Programs on Public School
Incentives and Response

Chart 1

Analyzing the Effect of “Voucher Threat”
versus Vouchers
Panel A

What incentives would be created by the aforementioned
program rules, and how would one expect the affected public
schools to respond? Consider a public school subject to the
Florida program, a school that has just received its first “F”
grade (“F-school” hereafter). The school realizes that if it can
avoid another “F” grade in the next three years, it can escape
vouchers and the monetary loss and embarrassment associated
with them.5 Therefore, it would have an incentive to improve
its scores so as not to receive a second “F” grade. In contrast,
if the same school were subject to a Milwaukee-type voucher
program—in which vouchers have already been introduced—
the school could not avoid vouchers (and the revenue loss)

Net revenue
R*
RT
Rv
v=0
v>0

R1
R2

e v e* e T e
Effort
Panel B

In a Florida-type [voucher] program, the
threatened public schools . . . have more
of an incentive to respond in order to
improve their scores and escape vouchers.

Net revenue
R*
RT
Rv
v=0
v>0

altogether. In this case, improvements would serve to retain
or attract a few students, but the effect would be marginal
compared with that of a Florida-type program. In a Floridatype program, the threatened public schools (schools that have
received their first “F” grade) have more of an incentive to
respond in order to improve their scores and escape vouchers.6
Thus, the key difference between the two programs is that in
the Milwaukee program, vouchers have already been
implemented, whereas the Florida program first threatens the
schools and gives them a window to respond, and an adequate
response can preclude sanctions. Sanctions (vouchers) are
implemented only if the schools fail to attain the predesignated
standard.
4

We focus on the Milwaukee program up to 1996-97. The reason is that
following a 1998 Wisconsin Supreme Court ruling, there was a major shift
in the program when religious private schools were allowed to participate.
Moreover, the financing of the Milwaukee program underwent some crucial
changes, so that the voucher amounts and the revenue loss per student due
to vouchers were not comparable between the Florida and second-phase
Milwaukee programs. See Chakrabarti (2008a) for an analysis of how the shift
in the Milwaukee voucher program affected public school incentives and
performance as well as for a comparison of public school responses in the two
phases of the program. We focus on the Florida program up to 2001-02. This
period is chosen because schools that received an “F” grade in 1999 would
face the threat of vouchers only through 2002.
5
The loss of students due to vouchers leads to a decrease in both revenue and
costs for the school. But for a school operating at full capacity, the cost savings
due to the loss of students are marginal, while the loss in revenue is not. This
effect is a major reason why public schools do not support vouchers.
6
For a formal proof, see Chakrabarti (2008b).

Program Design, Incentives, and Response

R2
e* e v e T e
Effort

The intuition above is shown in Chart 1. Let R 1 illustrate
the initial net revenue function of the public school. The public
school chooses the effort to maximize net revenue. Let this
equilibrium effort be denoted by e and the corresponding net
revenue by R . Now assume that Milwaukee-type vouchers are
introduced. This leads to a downward shift of the net revenue
function—the new net revenue function is denoted by R 2 and
the corresponding optimum effort and net revenue by e V and
R V , respectively.7 Panel A of the chart illustrates the case in
which e V  e , and panel B the case in which e V  e . The chart
implies that any target effort in the range  e V , e  under a
threat-of-voucher regime will induce an effort strictly greater
than e V . For example, assume that the policymaker
implements a target effort, e T . Satisfying this target would lead
to a net revenue of R T while failing to satisfy it would lead to
the introduction of vouchers and corresponding revenue of
R V ( R T  . Therefore, the school has an incentive to implement
an effort of e T (  e V  .

For formal proofs, see Chakrabarti (2008b).

4. Data
The Florida data consist of school-level data on test scores,
grades, socioeconomic characteristics of schools, and school
finances; they are obtained from the Florida Department of
Education. School-level data on test scores are obtained from
the Florida Comprehensive Assessment Test. Mean scale scores
(on a scale of 100-500) on grade 4 reading and grade 5 math are
available for 1998-2002. Mean scale scores (on a scale of 1-6)
on the Florida grade 4 writing test are available for 1994-2002.
Data on socioeconomic characteristics include sex
composition (1994-2002), percentage of students eligible for
free or reduced-price lunch (1997-2002), and race composition
(1994-2002), and are obtained from the school indicators
database of the Florida Department of Education. This study
refers to school years by the calendar year of the spring
semester. School finance data consist of several measures of
school-level and district-level per-pupil expenditures, and are
obtained from the school indicators database and the Office
of Funding and Financial Reporting of the Florida Department
of Education.
The Wisconsin data consist of school-level data on test
scores, socioeconomic characteristics of schools, and per-pupil
expenditures (both at the school and district levels). The data
are obtained from the Wisconsin Department of Public
Instruction, the Milwaukee Public Schools, and the Common
Core of Data of the National Center for Education Statistics.
School-level data on test scores are obtained for 1) the Third
Grade Reading Test (renamed the Wisconsin Reading
Comprehension Test, or WRCT, in 1996) and 2) the grade 5
Iowa Test of Basic Skills (ITBS). School scores for the WRCT,
which was first administered in 1989, are reported in three
“performance standard categories”: percentage of students
below, percentage of students at, and percentage of students
above the standard.8 Data for these three categories are
available for 1989-97. School-level ITBS reading data (mean
scores) are available for 1987-93; ITBS math data (mean
scores) are available for 1987-97.

5. Empirical Strategy
5.1 Florida
In Florida, the schools that received an “F” grade in 1999 were
directly exposed to the threat of vouchers because all their
students would be eligible for vouchers if the school received
8

The method of reporting ITBS math and WRCT reading scores changed
in 1998. Therefore, we use pre-1998 scores.

another “F” in the next three years. These F-schools constitute
the group of treated schools. Schools that received a “D” grade
in 1999 were closest to the F-schools in terms of grade, but were
not directly treated by the program. These “D-schools”
constitute the group of control schools. The treatment and
control groups consist of 65 and 457 elementary schools,

If the F-schools and D-schools have
similar trends in scores in the pre-program
period, any shift of the F-schools
compared with the D-schools in the postprogram period can be attributed to the
program.
respectively.9 Because the program was announced in
June 1999 and the grades were based on tests held in February
1999, we classify schools into treatment and control groups on
the basis of their pre-program scores and grades.
The identifying assumption here is that if the F-schools and
D-schools have similar trends in scores in the pre-program
period, any shift of the F-schools compared with the D-schools
in the post-program period can be attributed to the program.
To test this assumption, we first run the following fixed-effects
regression (and its ordinary least squares [OLS] counterpart)
using only pre-program data:
(1)

sit = f i +  0 t +  1  F * t  +  2 X it +  it ,

where si t is the mean score of school i in year t, fi are schoolfixed effects, t denotes a time trend, F is a dummy variable
taking a value of 1 for F-schools and 0 for D-schools, F*t is an
interaction between the F dummy and trend, X it denotes the
set of school characteristics, and  it is a stochastic error term.
Scores considered in the Florida part of the analysis include
mean school scores in FCAT reading, FCAT math, and FCAT
writing. The pre-program difference in trend of the F-schools
is captured in  1 .
If F-schools and D-schools have similar pre-program
trends, we investigate whether the F-schools demonstrate a
higher improvement in test scores in the post-program era
using specification 2 below. If the treated F-schools
demonstrate a differential pre-program trend, then in addition
to estimating this specification, we estimate a modified version
in which we control for the pre-program differences in trends.
We estimate a completely unrestricted and nonlinear model
that includes year dummies to control for common year
effects and interactions of post-program year dummies with
9
We restrict our analysis to elementary schools because there were too few
middle and high schools that received an “F” grade in 1999 (seven and five,
respectively) to justify analysis.

FRBNY Economic Policy Review / October 2010

the F-school dummy to capture individual post-program year
effects:
2002

(2)

sit = f i +



 0j Dj +

j = 1999

2002



 1j  F * D j  +  2 X it +  it ,

j = 1999

where D j , j =  1999 2000 2001 2002  are year dummies
for 1999, 2000, 2001, and 2002, respectively. While the above
specification includes school-fixed effects, we also estimate an
OLS counterpart to it. OLS regressions corresponding to both
specifications 1 and 2 include a dummy for the treatment
group F. Note that this is absorbed in the fixed-effects
regressions because it is a time-invariant school effect.
Specification 2 does not constrain the post-program yearto-year gains of the F-schools to be equal and allows the
program effect to vary across years. The coefficients
 1t  i = 2000 2001 2002 represent the effect of one, two, and
three years into the program, respectively, for the F-schools.
Given the nature of the Florida program, the 1999 threatened
schools (that is, the schools that received an “F” grade in 1999)
would be exposed to the threat of vouchers for the next three
years only. Therefore, we track the performance of the
threatened schools (relative to the control schools) for three
years after the program—2000, 2001, and 2002—when the
threat of vouchers would be in effect.
The above specifications assume that the D-schools were
not affected by the program. Although the D-schools did not
face any direct threat from the program, they might have faced
an indirect threat because they were close to receiving an “F”
grade.10 Therefore, we next allow the F-schools and D-schools
to be different treated groups (with varying intensities of
treatment) and compare their post-program improvements, if
any, with 1999 “C-schools,” which are the next grade up in the
scale using the above specifications after adjusting for another
treatment group. It should be noted that since D-schools and
C-schools may face the threat to some extent, our estimates
may be underestimates (lower bounds), but not overestimates.

5.2 Milwaukee
Our strategy is based on and is similar to that of Hoxby
(2003b). Since students in the Milwaukee Public Schools
eligible for free or reduced-price lunch were also eligible for
vouchers, the extent of treatment of the Milwaukee schools
depended on the percentage of students eligible for free or
reduced-price lunch.11 Using this information, Hoxby
10
In fact, there is some anecdotal evidence that D-schools may have responded
to the program. The superintendent of Hillsborough County, which had no
F-schools in 1999, announced that he would take a 5 percent pay cut if any of
his thirty-seven D-schools received an “F” grade on the next school report card.
For more information, see Innerst (2000).

Program Design, Incentives, and Response

classifies the Milwaukee schools into two treatment groups
based on the percentages of students eligible for free or
reduced-price lunch—“most treated” (at least two-thirds of
students eligible in the pre-program period) and “somewhat
treated” (less than two-thirds of students eligible in the preprogram period).
We classify the schools into three treatment groups (in
contrast to Hoxby’s two) based on their pre-program (1989-90
school year) percentage of students eligible for free or reducedprice lunch. Thus, our treatment groups are more homogenous
as well as starker from each other. Additionally, to test the

Since students in the Milwaukee Public
Schools eligible for free or reduced-price
lunch were also eligible for vouchers,
the extent of treatment of the Milwaukee
schools depended on the percentage
of students eligible for free or reducedprice lunch.
robustness of our results, we consider alternative samples
obtained by varying the cutoffs that separate the different
treatment groups, departing from the Hoxby approach.
The 60-47 (66-47) sample classifies schools that have at least
60 percent (66 percent) of students eligible for free or reducedprice lunch as “more treated,” schools with such population
between 60 percent (66 percent) and 47 percent as “somewhat
treated,” and schools with such population less than 47 percent
as “less treated.” We also consider alternative classifications,
such as “66” and “60” samples, where there are two treatment
groups—schools that have at least 66 percent (60 percent) of
students eligible for free or reduced-price lunch are designated
as more treated schools, and schools with such population
below 66 percent (60 percent) as somewhat treated schools.
Since there were very few middle and high schools in the
Milwaukee Public Schools and student participation in the
Milwaukee Parental Choice Program was mostly in the
elementary grades, we restrict our analysis to elementary
schools.

11
Under the Milwaukee program, all households at or below 175 percent of
the poverty line are eligible to apply for vouchers. Households at or below
185 percent of the poverty line are eligible for free or reduced-price lunch.
However, the cutoff of 175 percent is not strictly enforced (Hoxby 2003a),
and households within this 10 percent margin are often permitted to apply.
In addition, there were very few students who fell in the 175 percent185 percent range, while in fact 90 percent of students eligible for free or
reduced-price lunch qualified for free lunch (Witte 2000). Students below
135 percent of the poverty line qualified for free lunch.

Our control group criteria are also based on Hoxby (2003b).
Since all schools in Milwaukee were potentially affected by the
program, Hoxby constructs a control group that consists of
Wisconsin schools outside Milwaukee that satisfy the following
criteria in the pre-program period that: 1) had at least 25 percent of their population eligible for free or reduced-price lunch,
2) had black students who make up at least 15 percent of the
population, and 3) were urban. Her control group consists
of twelve schools.
For our control schools, we designate schools that are
located outside Milwaukee but within Wisconsin, satisfy
Hoxby’s first two criteria, and have locales as similar as possible
to the Milwaukee schools. Note that all of these characteristics
pertain to the pre-program school year 1989-90.12
Using each sample, we investigate how the different
treatment groups in Milwaukee responded to the “voucher
shock” program. Using specification 3 below, we first test
whether the pre-program trends of the untreated and the
different treated groups were the same. We then estimate OLS
and fixed-effects versions of specification 4 below. If we
observe differences in pre-existing trends between the different
treated groups of schools, then in addition to estimating
specification 4, we estimate modified versions of the
specification that control for pre-existing differences in trends:
(3)

 1k  Ik * t  + 2 Xit + it

s it = fi +  0 t +

Table 2

Pre-Program Demographic Characteristics of Florida
and Wisconsin More Treated and Control Schools
Percent
Panel A: More Treated Schools
Wisconsin

Black
Hispanic
White
Male
Free or reducedprice lunch

Hispanic
White
Male

(4)

sit = f i +

2007

j = 1989

j = 1989x

  0j Dj + 

 1k j  I k * D j  + 2 X it +  it ,

where sit denotes scores of school i in period t; D j ,
j =  1989  2007  are year dummies for 1989 through
2007, respectively; k   MT ST LT  for the WRCT and
k   MT ST  for the ITBS, where MT denotes “more treated,”
ST denotes “somewhat treated,” and LT denotes “less treated.”
The scores considered are mean scores in ITBS reading and
ITBS math as well as percentages of students above the
standard in WRCT reading.

6. Results
Table 2 presents baseline characteristics of treated and control
groups in Florida and Wisconsin. It shows that the more
treated schools in Florida were indeed similar to the more
12
The more treated and control group characteristics are presented in Table 2.
In the 66-47 sample, the somewhat treated (less treated) group had an average
of 55.4 percent (37.17 percent) of students eligible for free or reduced-price
lunch, 50.99 percent (45.37 percent) who were black, and 4.09 percent
(3.83 percent) who were Hispanic.

Florida

66-47

60-47

66-47

60-47

62.79
(28.23)
18.95
(23.40)
17.18
(19.54)
51.38
(4.84)
85.80
(9.95)

66.55
(32.22)
18.07
(24.54)
10.21
(10.68)
52.25
(2.60)
84.5
(6.48)

62.90
(29.58)
14.81
(21.86)
17.38
(16.55)
52.33
(2.58)
82.9
(9.04)

-3.76
[0.56]
0.88
[0.87]
6.97
[0.07]
-0.87
[0.34]
1.3
[0.50]

-0.10
[0.99]
4.14
[0.36]
-0.20
[0.96]
-0.95
[0.22]
2.9
[0.12]

Panel B: Control Schools
Florida
Black

Free or reducedprice lunch

Florida–Wisconsin

18.12
(14.17)
15.49
(21.23)
63.59
(22.33)
51.38
(4.84)
50.14
(17.51)

Wisconsin
22.37
(12.93)
14.84
(6.02)
60.85
(12.80)
50.63
(2.29)
44.95
(11.66)

Florida–Wisconsin
-4.25
[0.10]
0.17
[0.86]
2.73
[0.49]
0.76
[0.43]
5.19
[0.10]

Source: Author’s calculations.
Notes: The group of Florida more treated and control schools is composed
of F-schools and C-schools, respectively. Samples 66-47 and 60-47 are
described in Section 5.2 of the article. Standard deviations are in
parentheses; p-values are in brackets.

treated schools in Wisconsin and, except in one case, the
differences between them were not statistically significant.
Similarly, the control schools in Florida were similar to the
control schools in Wisconsin, and the differences between
them were not statistically significant.
However, the treated schools were somewhat different
from the control schools within each state. The reason is that
Wisconsin schools outside Milwaukee were considerably more
advantaged than schools in Milwaukee. We arrived at this
control group despite using the strategy (following Hoxby
[2003a, b]) of selecting control schools as similar as possible to
Milwaukee’s more treated schools in terms of pre-program
characteristics.
It is important that both the more treated schools and the
control groups be similar across the two programs in terms of
pre-program characteristics as well as across the two locations.
As a result, for purposes of comparing effects across the two

FRBNY Economic Policy Review / October 2010

programs, we use the C-schools in Florida as the control group.
Noticeably, the control group in Wisconsin was very similar to
the C-schools in Florida and was not statistically different from
them in terms of any characteristics (Table 2). Still another
reason for selecting the C-schools as the control group in Florida
was that while the D-schools were more similar to the more
treated F-schools in terms of grade and demographics, they were
very close to receiving an “F” grade; hence, to some extent they
perceived an indirect threat and to some extent were treated by
the program.
Because of differences between the treated and control
schools, one might argue that in the absence of the program,
the control group would have evolved differently from the
more treated group. However, multiple years of pre-program
data allow us to check (and control) for any differences in
pre-program trends of these groups. In this way, we can dispose
of any level differences between the treated and control groups
as well as control for differences in pre-program trends, if any.
It seems likely that once we control for differences in trends as
well as in levels, any remaining differences between the treated
and control groups will be minimal. In other words, our
identifying assumption is that if the treated schools followed
the same trends as the control schools in the immediate preprogram period, they would have evolved similarly in the
immediate post-program period in the absence of the program.
We also control for time-varying observable characteristics.
School-fixed effects remove any time-invariant unobservable
characteristics. Note that while time-varying unobserved
characteristics cannot be directly controlled for, they did not
drive the results as long as the F-schools did not experience a
differential shock in unobserved characteristics that coincided
with the timing of the program.

6.1 Florida
Considerable anecdotal evidence suggests that F-schools have
responded to the voucher program. Just after the program’s
inception, Escambia County implemented a 210-day extended
school year in its F-schools (the typical duration was 180 days),
introduced an extended school day at least twice a week, and
added small-group tutoring on afternoons and Saturdays and
longer time blocks for writing and math instruction. To curb
absenteeism, the county started an automated phone system to
contact parents when a child is absent. Miami-Dade County
hired 210 additional teachers for its twenty-six F-schools,
switched to phonics instruction, and encouraged parents (many
of whom were dropouts) to go back to school for a high-school-

Program Design, Incentives, and Response

equivalency diploma. Broward County reduced its class size to
eighteen to twenty students in its low-performing schools and
increased services for children whose primary language is not
English. Palm Beach County targeted its fourth-grade teachers
for coaching and began more frequent and closer observation of
teachers in its F-schools (Innerst 2000). Carmen Varela-Russo,
Associate Superintendent of Technology, Strategic Planning,

Considerable anecdotal evidence
suggests that F-schools have responded
to [Florida’s] voucher program.
and Accountability at Broward County Public Schools, described
the situation this way: “People get lulled into complacency . . . the
possibility of losing children to private schools or other districts
was a strong message to the whole community” (Innerst 2000).
The analysis below investigates whether the data in Florida
support this behavior.
Chart 2, which depicts trends in reading, math, and writing
scores in F-schools and D-schools, shows that 1999 was the
watershed year. In both reading and math, the F-schools had
similar trends before the program. However, the F-schools
showed improvement relative to the D-schools after the
program, and the gap between F- and D-schools narrowed. In
writing, while the F-schools were deteriorating relative to the
D-schools before the program, this pattern changed after it.
The F-schools showed improvement relative to the D-schools
to the extent that they successfully closed the “F” to “D” gap
after the program.
We now turn to our estimation results. All regressions
control for ethnicity (the percentage of students in different
racial categories in a school), the percentage of male students,
the percentage of students eligible for free or reduced-price
lunch, and real per-pupil expenditures. Table 3 presents preprogram trends in reading, math, and writing. It reveals that
F-schools have no significant differences in trend compared
with D-schools in reading and math, although they exhibit a
small, negative differential trend in writing. Compared with
C-schools, F-schools exhibit a negative differential trend in
reading and writing, but no significant differential trend in
math. D-schools exhibit a negative trend in reading and a
positive trend in math and writing compared with C-schools.
Whenever there is a difference in pre-program trends, our
reported regressions control for these differences by including
interactions between trend and the respective treated
dummies.

D-schools are considered as an additional treatment group
in Table 4, columns 4-6. Here, we see how the program affects
F-schools (more treated) and D-schools (less treated)
compared with C-schools. All columns control for differences
in pre-existing trends between groups. The results show
positive, significant year effects in reading, math, and writing
for F-schools in each of the years after the program’s
implementation. Although many of the D-school effects are
also positive and significant, the F-school shifts are statistically
larger in each year.15 The F-school effects are economically
meaningful as well. In reading, relative to the base year,
F-schools showed a 3.6 percent improvement in the first year
after the program, a 4.2 percent improvement after the second
year, and a 6.3 percent improvement after the third year.
In math, F-schools showed a 3.4 percent, 4.2 percent, and

Chart 2

Florida “Threat-of-Voucher” Program
Score

290

FCAT Reading

280

D-schools

270
F-schools

260
250
1998
300

FCAT Math

D-schools

290

Our results show considerable
improvement in the F-schools after the
program compared with the control
schools.

F-schools
280

270
260
1998
3.5

FCAT Writing

3.0

D-schools
F-schools

2.5

2.0
1.5
1994

Source: Author’s calculations.
Note: FCAT is the Florida Comprehensive Assessment Test.

Table 4, columns 1-3 present the effects of the Florida
threat-of-voucher program on F-school reading, math, and
writing scores compared with those for D-schools. All models
reported include school-fixed effects. The results from our OLS
estimation are similar to the fixed-effects estimates and hence
are not reported. The regressions for writing include
interactions of the “F” dummy with trend to control for
differences in pre-program trends seen above.13,14 The table
shows economically large, positive, and statistically significant
effects in each subject area and year.

4.5 percent improvement in the first, second, and third years,
respectively, after implementation of the program. In writing,
the percentage improvement was around 15 percent. At the
end of 2002 (three years after program implementation), the
pre-program gap between F-schools and C-schools was closed
by 37.08 percent in reading, 30.31 percent in math, and around
75 percent in writing.
In summary, based on different samples, different subjects,
and both OLS and fixed-effects estimates, our results show
considerable improvement in the F-schools after the program
compared with the control schools. Although D-schools show
non-negligible improvement (at least in reading and writing),
their improvement is considerably less than and statistically
different from that of the F-schools.

Note that the table reports only the coefficients that reflect program effects;
therefore, the coefficient corresponding to this interaction term (which
captures the differential pre-existing trend) is not reported. Pre-existing
trends are reported in Table 3.
14
The regressions for reading and math (columns 1 and 2) do not include this
interaction term because there is no evidence of differential pre-program
trends in reading and math for F-schools and D-schools (Table 3). Note that
the results with inclusion of this term remain very similar.
15
Here, we test whether the F-school effects are statistically different from
the D-school effects against the null hypothesis that they are equal.

FRBNY Economic Policy Review / October 2010

Table 3

Pre-Program Trend of F-, D-, and C-Schools in Florida
Sample of F- and D-Schools
FCAT Reading

Trend
F * trend

FCAT Math
OLS
(3)

FE
(4)

Sample of F-, D-, and C-Schools
FCAT Writing
OLS
(5)

FE
(6)

OLS
(7)

FE
(8)

FCAT Math
OLS
(9)

FE
(10)

FCAT Writing

OLS
(1)

FE
(2)

0.41
(0.56)
-1.78
(2.47)

-0.05
(0.47)
-2.01
(1.46)

13.20***
(0.55)
-0.98
(1.44)

13.02**
(0.61)
-0.72
(1.48)

0.20**
(0.008)
-0.05***
(0.011)

0.21**
(0.003)
-0.04***
(0.007)

2.66**
(0.57)
-3.80
(2.29)
-2.29***
(0.66)

2.70
(0.36)
-4.77***
(1.41)
-2.69***
(0.57)

9.79***
(0.53)
2.46
(1.51)
3.46***
(0.60)

10.20***
(0.38)
1.96
(1.43)
2.79***
(0.67)

0.18***
(0.01)
-0.03***
(0.01)
0.02**
(0.007)

0.19***
(0.002)
-0.03***
(0.01)
0.02***
(0.003)

Y
1,013
0.58

Y
1,013
0.93

Y
1,006
0.59

Y
1,006
0.90

Y
2,948
0.64

Y
2,948
0.80

Y
2,386
0.76

Y
2,386
0.95

Y
2,377
0.74

Y
2,377
0.93

Y
6,982
0.65

Y
6,982
0.82

D * trend

Controls
Observations
R2

FCAT Reading

OLS
(11)

FE
(12)

Source: Author’s calculations.
Notes: FCAT is the Florida Comprehensive Assessment Test; OLS is ordinary least squares regression; FE is fixed-effects regression. Controls include race,
sex, percentage of students eligible for free or reduced-price lunch, and real per-pupil expenditure. Huber-White standard errors are in parentheses.
All regressions are weighted by the number of students tested.
***Statistically significant at the 1 percent level.
***Statistically significant at the 5 percent level.
***Statistically significant at the 10 percent level.

Program Design, Incentives, and Response

Table 4

Effect of “Threatened Status” on FCAT Reading, Math, and Writing Scores
Sample of Treated F- and Control D-Schools in Florida
Reading
FE
(1)
Treated * one year after program
Treated * two years after program
Treated * three years after program

4.85***
(1.68)
3.30*
(1.71)
7.08***
(1.78)

Math
FE
(2)
6.78***
(1.63)
7.25***
(1.82)
5.35***
(2.00)

Writing
FE
(3)

Less treated * two years after program
Less treated * three years after program

More treated * one year after program
More treated * two years after program
More treated * three years after program

Y
Y
Y
2,550
0.77
0.00

Y
Y
Y
2,524
0.76
0.00

Math
FE
(5)

Writing
FE
(6)

0.35***
(0.04)
0.37***
(0.04)
0.43
(0.05)

Less treated * one year after program

School-fixed effects
Year dummies
Controls
Observations
R2
p-valuec

Reading
FE
(4)

Y
Y
Y
4,476
0.85
0.00

3.53***
(0.76)
5.52***
(0.80)
7.94***

0.97
(0.85)
2.54***
(0.94)
3.47***

0.05**
(0.02)
0.00
(0.02)
-0.03

(0.87)
9.32b***
(1.87)
10.75a***
(1.87)
16.03b***
(1.91)

(0.92)
8.96b***
(1.59)
11.00b***
(1.77)
11.94b***
(1.95)

(0.02)
0.39b***
(0.04)
0.37a***
(0.04)
0.39a***
(0.05)

Y
Y
Y
5,933
0.86
0.00

Y
Y
Y
5,909
0.83
0.00

Y
Y
Y
10,587
0.86
0.00

Source: Author’s calculations.
Notes: FCAT is the Florida Comprehensive Assessment Test. FCAT scores for reading and math are for the period 1998-2000; FCAT scores for writing are
for the period 1994-2002. FE is fixed-effects regression. Huber-White standard errors are in parentheses. Controls include race, sex, percentage of students
eligible for free or reduced-price lunch, and real per-pupil expenditure. All regressions are weighted by the number of students tested.
a

More treated significantly different from less treated at 5 percent level.
More treated significantly different from less treated at 1 percent level.
c
p-value of F-test of the program effect on treated schools.
b

***Statistically significant at the 1 percent level.
***Statistically significant at the 5 percent level.
***Statistically significant at the 10 percent level.

FRBNY Economic Policy Review / October 2010

Table 5

Pre-Program Trend of More Treated, Somewhat Treated, and Less Treated Schools in Milwaukee
WRCT (Percentage above)
OLS
(1)
Trend
More treated * trend
Somewhat treated * trend
Less treated * trend

Observations
R2

ITBS Reading

FE
(2)

ITBS Math

OLS
(3)

FE
(4)

OLS
(5)

FE
(6)

-3.84
(2.33)
-3.08
(3.41)
-4.41
(3.01)
-2.33
(3.61)

-4.34**
(2.16)
-2.03
(3.35)
-3.61
(2.67)
-3.23
(3.10)

-4.09
(4.11)
4.01
(3.69)
3.14
(4.05)

-3.45
(3.42)
-1.88
(2.73)
2.12
(3.17)

-3.04*
(1.66)
0.56
(1.97)
0.73
(1.83)

2.52**
(0.98)
0.32
(1.40)
0.31
(1.21)

242
0.50

242
0.87

411
0.30

411
0.56

410
0.30

410
0.71

Source: Author’s calculations.
Notes: WRCT is the Wisconsin Reading Comprehension Test; ITBS is the Iowa Test of Basic Skills; OLS is ordinary least squares regression; FE is fixed-effects
regression. Controls include race, sex, and percentage of students eligible for free or reduced-price lunch.
***Statistically significant at the 1 percent level.
***Statistically significant at the 5 percent level.
***Statistically significant at the 10 percent level.

6.2 Milwaukee

These results seem to be robust in that they are replicated in the
analysis with other samples.17 Chart 3 presents the trends in

The Milwaukee analysis uses the 66-47 sample. Estimation
results for pre-program trends are presented in Table 5.
The results show no statistical difference in trends between
the various treated and control groups in any subject area.
Table 6 examines the effect of the Milwaukee “voucher
shock” program on the WRCT (the percentage above), ITBS
reading, and ITBS math scores of different treated groups.
Except for the positive and statistically significant effect in
WRCT reading in the test’s second year, there is no statistically
significant evidence of any effect of the program. Although
the second year’s somewhat treated effect in ITBS math is
statistically significant, it is more than the corresponding
more treated effect.16
Thus, the results in Milwaukee are mixed. The program
seems to have had a positive and significant effect in the second
year after the program’s implementation, at least in the WRCT.
16

Since the ITBS was administered in Milwaukee as a district assessment
program, we do not have data on non–Milwaukee, Wisconsin, schools for this
test. As a result, our comparison group is the less treated group of schools.
Since the comparison group is also treated to some extent, we expect our
estimates for the ITBS to be lower bounds.

Program Design, Incentives, and Response

The results show no statistical difference
in trends between the various treated and
control groups in any subject area. . . .
Except for the positive and statistically
significant effect in [Wisconsin Reading
Comprehension Test] reading in the test’s
second year, there is no statistically
significant evidence of any effect of the
program. . . . Thus, the results in
Milwaukee are mixed.
ITBS scores for the various groups. As expected, there is no
evidence of any program effect.

These results are not reported here, but are available from the author.

Table 6

Chart 3

Effect of the Milwaukee “Voucher Shock” Program

Milwaukee “Voucher-Shock” Program

WRCT
(1)
Somewhat treated *
one year after program

ITBS Reading
(2)

ITBS Math
(3)

More treated * three years
after program

2.03
(2.81)
5.38**
(2.43)
5.01
(3.03)
-0.92
(3.33)
6.06*
(3.14)
5.69
(3.98)

4.15
(4.49)
7.83
(5.17)
6.78
(5.31)
1.12
(3.86)
6.59
(5.15)
2.85
(5.18)

-1.35
(2.94)
6.14*
(3.38)
2.47
(3.31)
-4.02
(3.26)
4.36
(3.83)
-2.22
(3.54)

School-fixed effects
Year dummies
Controls
Observations
R2
a
p-value

Y
Y
Y
1,195
0.58
0.11

Y
Y
Y
717
0.55
0.62

Y
Y
Y
1,127
0.60
0.27

Somewhat treated *
two years after program
Somewhat treated *
three years after program
More treated * one year
after program
More treated * two years
after program

Score

ITBS Reading

Control

50
Somewhat treated schools

40
30

More treated schools
20
1987
70

ITBS Math

60
Control
50
Somewhat treated schools
40
More treated schools

Source: Author’s calculations.
Notes: WRCT is the Wisconsin Reading Comprehension Test; ITBS is the
Iowa Test of Basic Skills. Huber-White standard errors are in parentheses.
All regressions include school-fixed effects and control for race, sex,
percentage of students eligible for free or reduced-price lunch, and
real per-pupil expenditure.

30
1987 88

Source: Author’s calculations.
Note: ITBS is the Iowa Test of Basic Skills.

p-value of the F-test of joint significance of more treated shift
coefficients.
***Statistically significant at the 1 percent level.
***Statistically significant at the 5 percent level.
***Statistically significant at the 10 percent level.

FRBNY Economic Policy Review / October 2010

7. Robustness Checks

Table 7

7.1 Mean Reversion

Panel A: 98F- and 98D-Schools

Mean Reversion of 98F-Schools Compared
with 98D- and 98C-Schools, 1998-99

Dependent Variable: FCAT Score, 1998-99

Several factors might bias the results; we consider each factor
and its potential solutions. First is the issue of mean reversion.
Mean reversion is the statistical tendency whereby highor low-scoring schools tend to score closer to the mean
subsequently. Because the F-schools scored low in 1999, a
natural question would be whether the improvement in Florida
is driven by mean reversion rather than the voucher program.
Since we conduct a difference-in-differences analysis, our
estimates will be tainted by mean reversion only if F-schools
mean-revert to a greater extent than do the D-schools or the
C-schools, or both.
To investigate mean reversion, we examine whether and by
how much schools that received an “F” grade in 1998 improved
during the 1998-99 academic year compared with those that
received a “D” (or “C”) grade in 1998. Since these years fall
within the pre-program period, the gain can be taken to
approximate the mean-reversion effect and can be subtracted
from the post-program gain of F-schools compared with
D-schools (or C-schools) to get at the mean-reversioncorrected program effect.
The accountability system of assigning letter grades to
schools began in 1999. The pre-1999 accountability system
classified schools into four groups, designated 1 (low) to 4
(high). However, using the state grading criteria and data on
the percentage of students in different achievement levels in
each FCAT reading, math, and writing, we assigned letter
grades to schools in 1998 and implemented the above strategy.
Schools receiving “F,” “D,” and “C” grades in 1998 using this
procedure are referred to as “98F-schools,” “98D-schools,” and
“98C-schools,” respectively.
Using Florida data for 1998 and 1999, we demonstrate in
Table 7, panel A, that when compared with the 98D-schools,
the 98F-schools show no evidence of mean reversion either in
reading or math, although there is mean reversion in writing.
Compared with the 98C-schools (panel B), there is no evidence
of mean reversion in reading; both 98D-schools and 98Fschools show comparable amounts of mean reversion in math;
only 98F-schools show mean reversion in writing.

Program Design, Incentives, and Response

Reading
FE
(1)
Trend
98F * trend
Observations
R2

2.01***
(0.43)
-0.65
(1.14)
1,353
0.93

Math
FE
(2)
14.02***
(0.49)
1.17
(1.19)
1,354
0.91

Writing
FE
(3)
0.04***
(0.01)
0.14***
(0.02)
1,355
0.85

Panel B: 98F-, 98D-, and 98C-Schools
Dependent Variable: FCAT Score, 1998-99
Reading
FE
(1)
Trend
98F * trend
98D * trend
Observations
R2

1.76***
(0.35)
-0.55
(1.12)
0.16
(0.54)
2,605
0.96

Math
FE
(2)
9.71***
(0.36)
4.63***
(1.16)
4.22***
(0.58)
2,608
0.94

Writing
FE
(3)
0.03***
(0.01)
0.14***
(0.02)
0.01
(0.01)
2,608
0.87

Source: Author’s calculations.
Notes: FCAT is the Florida Comprehensive Assessment Test; FE is fixedeffects regression. All regressions control for race, sex, percentage of
students eligible for free or reduced-price lunch, and real per-pupil
expenditure. The ordinary least squares regressions include 98F- and
98D-school dummies. In the sample of 98F- and 98D-schools, the
standard deviations of FCAT reading, math, and writing are 18.9, 18.05,
and 0.30, respectively. In the sample of 98F-, 98D-, and 98C-schools, the
standard deviations of FCAT reading, math, and writing are 21.16, 21.56,
and 0.31, respectively.
***Statistically significant at the 1 percent level.
***Statistically significant at the 5 percent level.
**Statistically significant at the 10 percent level.

Table 8

Is There a Stigma Effect of Getting the Lowest Performing Grade?
Effect of Being Categorized in Group 1 on FCAT Writing Scores
Using FCAT Writing Scores, 1997-98
Sample: Group 1, 2 Schools
OLS
(1)
Trend
Group 1 * trend

0.52***
(0.04)
-0.01
(0.08)

FE
(2)
0.52***
(0.03)
-0.02
(0.06)

Sample: Group 1, 2, 3 Schools
FE
(3)
0.48***
(0.04)
-0.02
(0.06)

Group 2 * trend

Controls
Observations
R2

N
314
0.49

N
314
0.84

Y
314
0.85

OLS
(4)

FE
(5)

FE
(6)

0.48***
(0.02)
0.03
(0.07)
0.03
(0.04)

0.48***
(0.01)
0.01
(0.05)
0.04
(0.03)

0.46***
(0.02)
0.02
(0.05)
0.04
(0.03)

N
1,361
0.52

N
1,361
0.87

Y
1,358
0.87

Source: Author’s calculations.
Notes: FCAT is the Florida Comprehensive Assessment Test; OLS is ordinary least squares regression; FE is fixed-effects regression. Huber-White standard
errors are in parentheses. All regressions are weighted by the number of students tested; controls include race, sex, percentage of students eligible for free
or reduced-price lunch, and real per-pupil expenditure. The OLS regressions include group 1 and group 2 dummies.
***Statistically significant at the 1 percent level.
***Statistically significant at the 5 percent level.
***Statistically significant at the 10 percent level.

7.2 Stigma Effect of Getting the Lowest
Performing Grade
A second concern in Florida is the potential stigma effect of
receiving a performance grade of “F.” If there is such a stigma,
the F-schools will try to improve only to avoid this stigma
rather than in response to the program. We use several

If there is [a low-performance] stigma,
the F-schools will try to improve only to
avoid this stigma, rather than in response
to the program.
alternative strategies to investigate this possibility. First,
although the system of assigning letter grades to schools started
in 1999, Florida had an accountability system in the pre-1999
period when schools were categorized into four groups,
designated 1 (low) to 4 (high), based on FCAT writing and
reading and math norm-referenced test scores. Using FCAT
writing data for two years (1997 and 1998), we investigate

whether the schools, which were categorized in group 1 in
1997, improved in relation to the 1997 group 2 and group 3
schools in 1997-98.18 Our rationale is that if a stigma effect
is associated with getting the lowest performing grade, the
group 1 schools should improve relative to the group 2 and 3
schools, even in the absence of the threat-of-voucher program.
Table 8, using pre-program FCAT writing scores, shows that
no such stigma effect exists—group 1 schools display no
improvement relative to the group 2 or group 3 schools.
Second, all the schools that received an “F” grade in 1999
received higher grades in 2000, 2001, and 2002. Therefore,
although the stigma effect on F-schools may be operative in
2000, this is not likely to be the case in 2001 or 2002 since none
of the F-schools received an “F” grade in the preceding year
(2000 or 2001, respectively). However, the F-schools would
face the threat of vouchers until 2002, so any improvement in
18
We do not use the pre-1999 reading and math norm-referenced test
(NRT) scores because different districts used different NRTs during this
period, which varied in content and norms. Also, districts often chose
different NRTs in different years. Thus, these NRTs were not comparable
across districts and across time. Moreover, since districts could choose the
specific NRT to administer each year, the choice was likely related to timevarying (and also time-invariant) district-unobservable characteristics that
also affected test scores.

FRBNY Economic Policy Review / October 2010

2001 and 2002 would provide evidence in favor of the threatof-voucher effect and against the stigma effect. F-schools
showed strong gains in both 2001 and 2002—a result that
provides further support for the threat-of-voucher effect and
against the stigma effect.

Table 9

Effect of Milwaukee Program on Demographic
Composition of Schools
Percent
Black
(1)
Less treated * program

7.3 Sorting
Another factor relates to sorting in the context of the
Milwaukee voucher program. Vouchers affect public school
quality not only through direct public school response but also
through changes in student composition and peer quality
brought about by sorting. These three factors are then reflected
in the public school scores.19 This issue is important in
Milwaukee because over the years students have left the city’s
public schools with vouchers. In contrast, no Florida school
became eligible for vouchers in 2000 or 2001. Therefore, the
program effects (for each of the years 2000, 2001, and 2002) are
not likely to be tainted by this factor.20 Moreover, as we discuss
shortly, the demographic compositions of the different groups
of schools remained very similar across the years under
consideration.
We also examine whether the demographic composition
of the different Milwaukee treated groups changed over the
years (Table 9). No such evidence is found. Only a few of the
coefficients are statistically significant, and they are always very

Vouchers affect public school quality
not only through direct public school
response but also through changes in
student composition and peer quality
brought about by sorting.
small in magnitude. They imply changes of less than 1 percent,
more precisely, ranging between 0.22 percent and 0.65 percent.
This result suggests that sorting was not an important factor.
Note that we conducted the same exercise for Florida as well
and found no evidence of any relative shift of the demographic
composition of the F-schools compared with the D-schools
or C-schools.

See Hsieh and Urquiola (2006) for a discussion.
This does not mean that the Florida program was not credible. Ten schools
received a second “F” grade in 2002, nine schools in 2003, and twenty-one in
2004; all of these students became eligible for vouchers.
20

Program Design, Incentives, and Response

Hispanic
(2)

Asian
(3)

Somewhat treated *
program * trend
More treated * program
* trend

0.90
(1.59)
-0.25
(1.35)
-1.0
(1.34)
0.22
(0.32)
0.70
(0.25)
0.08
(0.23)

0.40
(0.83)
1.06
(0.63)
1.57
(0.81)
0.16
(0.15)
-0.12
(0.13)
-0.39***
(0.14)

0.04
(0.37)
0.53
(0.37)
0.65*
(0.37)
0.24***
(0.07)
0.29***
(0.07)
-0.22***
(0.07)

Observations
R2

1,228
0.95

1,226
0.97

1,216
0.91

Somewhat treated *
program
More treated * program
Less treated * program
* trend

Source: Author’s calculations.
Notes: Huber-White standard errors are in parentheses. All regressions
are weighted by the number of students tested. All columns include a
time trend, a program dummy that takes a value of 1 after the program,
and an interaction between program dummy and trend.
***Statistically significant at the 1 percent level.
***Statistically significant at the 5 percent level.
***Statistically significant at the 10 percent level.

A Comparison of Program Effects in Florida
and Milwaukee
Since Florida and Milwaukee are in different regions, we argue
that our comparison of the effects of the two programs is fair
and reasonable. First, as mentioned earlier, apart from the
crucial design differences between the two programs, the other
features of the programs were very similar. In both programs,
private schools could not discriminate against voucher
applicants. Also, the method of funding for the two programs,
the average voucher amounts, and the per-pupil revenue losses
from vouchers were very similar. Second, state and local
revenues constituted very similar proportions of total revenue
during the relevant periods—the percentages of revenue from
state and local sources were 51 percent and 41 percent,
respectively, in Florida, and 55 percent and 36 percent,
respectively, in Milwaukee. Third, the demographic
characteristics of the more treated and control schools in
Florida were very similar, both economically and statistically,
to those of the more treated and control schools in Milwaukee

Table 10

Comparison of Results from Florida “Threat-of-Voucher” and Milwaukee “Voucher-Shock” Programs
Using Standardized Reading and Math Scores
Corrected for Mean Reversion
Reading
Wisconsin
WRCT
(1)
More treated * one year
after program
More treated * two years
after program
More treated * three years
after program

Math
Florida
FCAT
(2)

-0.06

Wisconsin
ITBS
(3)

Reading
Florida
FCAT
(4)

0.47***

-0.24

0.45***

0.38*

0.50***

0.26

0.55***

0.35

0.80***

-0.13

0.60***

Wisconsin
WRCT
(5)
-0.06

Math
Florida
FCAT
(6)

Wisconsin
ITBS
(7)

Florida
FCAT
(8)

0.47***

-0.24

0.24***

0.38*

0.50***

0.26

0.34***

0.36

0.80***

-0.13

0.39***

Source: Author’s calculations.
Notes: Reading test scores are from the Wisconsin Reading Comprehensive Test (WRCT), 1989-97, and the Florida Comprehensive Assessment Test (FCAT)
Reading, 1998-2002. Math test scores are from the Iowa Test of Basic Skills (ITBS) Math, 1986-97, and the FCAT Math, 1998-2002. All figures are respective
sample standard deviations. All figures are obtained from regressions that contain school-fixed effects, year dummies, interactions of year dummies with the
respective treatment dummies, race, sex, percentage of students eligible for free or reduced-price lunch, and real per-pupil expenditure. Standard deviation
of FCAT reading scores = 20; standard deviation of FCAT math scores = 20; standard deviation of WRCT (percentage above) reading scores = 16; standard
deviation of ITBS reading scores = 18.45; standard deviation of ITBS math scores = 16.71. For standard deviations corresponding to the mean reversion
sample, see the notes to Table 4.
***Statistically significant at the 1 percent level.
***Statistically significant at the 5 percent level.
***Statistically significant at the 10 percent level.

(Table 2). Fourth, we repeat our analysis by comparing the
improvement in Milwaukee with that of a large urban district
in Florida: Miami-Dade County (the state’s largest school
district). The results are very similar and hence are not reported
here. Finally, and perhaps most importantly, since we follow a
difference-in-differences strategy in trends, any level or even
trend differences between the two regions (that are common to
schools in that region) are differenced out. It is unlikely that
any remaining difference, which differentially affects the trends
in the two regions only in the post-program period, will be
large.
Table 10 compares the effects of the Florida and Milwaukee
programs on their respective more treated schools both before
and after correcting for mean reversion. Figures are based on
data in Tables 4 and 6, and all numbers are expressed in terms
of their respective sample standard deviations. Columns 1-4
present results before correcting for mean reversion; columns
5-8 present results corrected for mean reversion. Precorrection results show positive and significant effect sizes in
each of the years and subject areas in Florida, which always
exceed the corresponding Milwaukee effect sizes (which are not

significant, except in second-year reading). Mean-reversioncorrected effect sizes are obtained by subtracting the effect size
attributed to mean reversion (obtained from expressing the
relevant coefficients in Table 7, panel B, in terms of respective
standard deviations) from the F-school effect sizes (obtained
from expressing the more treated coefficients in Table 4,
columns 4-6, in terms of respective sample standard
deviations) in each of the three years after the program. The
estimates in reading are the same as those described earlier.
In math, although the effect sizes fall in Florida, they are still
positive and considerably larger than those in Milwaukee. In
reading (math), relative to the control schools, the F-schools
show an improvement of 0.47 (0.24) standard deviations in the
first year after the program, 0.5 (0.34) standard deviations after
the second year, and 0.8 (0.39) standard deviations after the
third year. Mean-reversion-corrected effect sizes in writing are
0.29, 0.25, and 0.29 in the first, second, and third years,
respectively, after the program. Note that since none of the
F-schools received an “F” grade in either 2000 or 2001, the
mean-reversion-corrected effect sizes attributed to the Florida
program in the second and third years may be underestimates.

FRBNY Economic Policy Review / October 2010

8. Lessons for New York City
Our analysis of school voucher programs implies that policies
that threaten underperforming public schools (or other agents)
with functional and credible sanctions can induce them to
respond in a way intended or desired by the policymaker. This
finding has important implications for some educational
policies in New York City. These include New York City’s own
accountability policy, also known as the “Progress Report”
policy, and the federal No Child Left Behind law, as
implemented by New York State.
The Progress Report policy was introduced in New York
City in 2007. It rates schools on a scale of A to F, with grades
based on three components: school environment, student
performance, and student progress. A school’s environment

As in Florida’s voucher program, public
schools in New York face valid sanctions if
they fail to perform. Therefore, incentives
faced by New York’s low-performing
schools are similar to those faced by the
F-schools in Florida, and one would
expect a similar response from them.
score is based on attendance rates and responses from surveys
given to teachers, parents, and students. The other two scores
are based on student performance in state math and English
Language Arts (ELA) examinations. While student
performance measures rely on level scores, student progress
measures rely on growth or changes in student scores over
years. The program attaches consequences to the letter grades.
Higher grade (A) schools are eligible for increases in per-pupil
funding and bonuses for principals. Schools receiving “F” or
“D” grades are required to implement “school improvement
measures and target setting.” Low-performing (F- and Dschools) are also threatened with changes in their principal,
and possible restructuring and closure if they continue to
receive poor grades. The program also makes students in
F-schools eligible to transfer to better performing schools.
Although the Progress Report program does not have a
voucher element, it is in many ways similar to the Florida
voucher program; indeed, its design was based on the Florida
program. Like the Florida program, it embeds sanctions in an
accountability framework with consequences/sanctions
imposed on low-performing schools if they fail to improve.
Additionally, the criteria of the New York City program that

Program Design, Incentives, and Response

make students in low-performing schools eligible to transfer to
other higher performing schools are similar to those of Florida’s
program. The only distinction is that in New York, students can
transfer to public schools only—not to private schools, as in the
Florida program. The threat of removal of the principal and the
possibility of restructuring are sanctions imposed over and
above the transfer option. These sanctions are credible and pose
a valid threat to administrators. For example, as reported in
Rockoff (2008), “Seven schools receiving an F and two schools
receiving a D were told in December of 2007 that they would be
closed immediately or phased out after the school year 200708. . . . Additionally, 17 percent of the remaining F-school
principals (and 12 percent of the D-school principals) did not
return in the school year 2008-09, relative to 9 percent of
principals receiving a C, B, or A grade.”
Thus, as in Florida’s voucher program, public schools in
New York face valid sanctions if they fail to perform. Therefore,
incentives faced by New York’s low-performing schools are
similar to those faced by the F-schools in Florida, and one
would expect a similar response from them. Accordingly, the
above analysis would indicate that low-performing schools
under the Progress Report program would have an incentive to
improve. In fact, there is some evidence in favor of such
improvement. In 2009, 82 percent of students passed in math
and 69 percent in English, up from 42 percent and 38 percent,
respectively, in 2002. Earlier, all five boroughs of New York
City ranked toward the bottom in the state; now Queens and
Staten Island rank toward the top in elementary-school math
scores. The racial achievement gap in passing rates has been
closed by half in some tests. (Statistics are from Elissa Gootman
and Robert Gebeloff, New York Times, August 4, 2009.)
Gootman and Gebeloff also report:
At Public School 398 in Brownsville, Brooklyn,
77 percent of students passed the math tests this
year and 60 percent passed English, up from 56
and 43 percent last year. Gene McCarthy, a fifthgrade teacher, attributed the improvement to
“a tremendous amount of test prep,” but said that
with a little creativity on his part, “ultimately I think
it’s learning.” The principal, Diane Danay-Caban,
said at P.S. 398, which had struggled for years with
low scores and discipline problems, she has come
to feel that the push to raise scores has brought
genuine gains in knowledge.
Rockoff and Turner (2008) find that schools labeled “F”
improved their performance in both ELA and math, with larger
effects in math. Winters (2008), analyzing the same program,
finds improvement of F-schools in math, although he finds
no such effect in ELA.

NCLB, a major reform of the Elementary and Secondary
Education Act, was signed into law on January 8, 2002. The
states, including New York, implemented it soon thereafter.
In compliance with the law, New York established Adequate
Yearly Progress (AYP) targets, and all schools were graded
on the basis of the targets. AYP is determined based on each
school’s progress toward meeting the state proficiency level
for all students in English language arts, mathematics, science,
as well as the high-school graduation rate. Schools are held
accountable for the achievement of students of different races
and ethnic groups, students with disabilities, students with
limited English proficiency, and students of low-income
families. Schools must also have an average over two years
of 95 percent of their students participating in state tests. If
a school does not meet requirements in any one of these
categories, it is said to miss AYP. Schools that receive Title I
money are subject to NCLB sanctions if they miss AYP in two
consecutive years. A school missing AYP for two consecutive
years is required to provide public school choice to students.
A school missing AYP for three consecutive years is required to
provide supplemental educational services (such as tutoring)
in addition to the above sanctions. Missing AYP for four
consecutive years leads to corrective action in addition to

Only a fraction of eligible students took
advantage of the transfer option in New
York as well as in the nation as a whole.
This result is attributable mainly to two
factors: the absence of an adequate
number of spaces in nearby schools and
the lack of adequate information.
the above sanctions; for five consecutive years, it results in
restructuring in addition to the above sanctions. Thus, sanctions
start with two years of missed AYP and escalate from there.
While NCLB does not have any voucher component, the
accountability-sanctions component is similar in spirit to that
of Florida’s voucher program. In fact, the design of NCLB was
based on that program. As in the Florida program, NCLB first
threatens failing schools with sanctions, and sanctions are
introduced only if the schools fail to meet the predesignated
targets in the following years.21 Therefore, one would expect
similar incentives to be created by NCLB and threatened
21
Note, though, that while under NCLB all low-performing schools face stigma
(embarrassment) due to public reporting of scores and grades, only Title I
schools (schools that receive Title I money) are subject to sanctions.

schools to respond in a way similar to the F-schools under the
Florida program. In other words, one would expect schools
threatened by the NCLB sanctions to improve their
performance in an effort to make AYP. However, it should be
emphasized that these incentives and responses would be

The challenge to policymakers in
[accountability] programs is to establish—
and enforce—credible sanctions that
function as valid threats to the agents
(here, public schools).
applicable only if the sanctions are credible and pose a valid
threat to the affected schools. Under NCLB, though,
implementation of the sanctions has been largely limited. For
example, only a fraction of eligible students took advantage of
the transfer option in New York as well as in the nation as a
whole. This result is attributable mainly to two factors: the
absence of an adequate number of spaces in nearby schools and
the lack of adequate information. For example, as reported in
the New York Daily News, “Some parents of kids in failing
schools told the Daily News they weren’t even aware they could
transfer out, and some were turned away from better schools
that are already overcrowded” (February 3, 2008).
In summary, both New York City’s Progress Report
program and NCLB have the potential to induce improvement
from threatened schools, but the incentives and response
ultimately depend on how functional and credible the threats
under consideration are. The challenge to policymakers in such
programs is to establish—and enforce—credible sanctions that
function as valid threats to the agents (here, public schools).
Only in such cases would the agents have an incentive to
respond in the direction intended or deemed appropriate by
the policymakers.

9. Conclusion
This article examines the role of program design in the context
of two educational interventions in the United States—the
Florida and Milwaukee school voucher programs. Even though
both programs involve vouchers, their designs are quite
different: the Milwaukee program makes low-income
Milwaukee public school students eligible for vouchers, while
the Florida system ties vouchers to low school performance.
Specifically, Florida students become eligible for vouchers if

FRBNY Economic Policy Review / October 2010

and only if their school receives two “F” grades in a period of
four years. This study shows that program design matters;
indeed, the design differences have had very different incentive
and performance effects on schools subject to the two
programs. Specifically, the Florida program led to considerably
larger improvements from the threatened schools compared
with corresponding schools under the Milwaukee program.
These findings are robust to several sensitivity checks.
The lessons drawn from our analysis are applicable to
some of New York City’s educational policies. These policies
include the No Child Left Behind Act, as implemented by the
state, and New York City’s “Progress Report” policy. While

Program Design, Incentives, and Response

neither of these programs has voucher components, both are
accountability programs that have consequences for schools
that fail to perform. In that sense, one would expect the
incentives and responses generated by these programs to be
similar to those created by the Florida program. Hence, the
threatened schools could be expected to improve in an effort
to avoid the sanctions. In fact, there is some evidence of such
improvement in the affected schools, especially in schools
treated by New York City’s Progress Report program.
However, the extent of the responses and the performance
effects ultimately depends on the credibility of the sanctions
and the validity of the threat posed to the affected schools.

References

Chakrabarti, R. 2007. “Vouchers, Public School Response, and the
Role of Incentives: Evidence from Florida.” Federal Reserve Bank
of New York Staff Reports, no. 306, October.
———. 2008a. “Can Increasing Private School Participation and
Monetary Loss in a Voucher Program Affect Public School
Performance? Evidence from Milwaukee.” Journal of Public
Economics 92, no. 5-6 (June): 1371-93.
———. 2008b. “Impact of Voucher Design on Public School
Performance: Evidence from Florida and Milwaukee Voucher
Programs.” Federal Reserve Bank of New York Staff Reports,
no. 315, January.
———. 2009. “Do Vouchers Lead to Sorting under Random PrivateSchool Selection? Evidence from the Milwaukee Voucher
Program.” Federal Reserve Bank of New York Staff Reports,
no. 379, July.
Epple, D., and R. E. Romano. 1998. “Competition between Private and
Public Schools, Vouchers, and Peer Group Effects.” American
Economic Review 88, no. 1 (March): 33-62.
———. 2002. “Educational Vouchers and Cream Skimming.”
International Economics Review 49, no. 4 (November):
1395-1435.
Figlio, D. N., and C. E. Rouse. 2006. “Do Accountability and Voucher
Threats Improve Low-Performing Schools?” Journal of Public
Economics 90, no. 1-2 (January): 239-55.
Greene, J. P. 2001. “An Evaluation of the Florida A-Plus Accountability
and School Choice Program.” Manhattan Institute for Policy
Research Civic Report (with Florida State University and the
Program on Education Policy and Governance at Harvard
University), February.

———. 2003b. “School Choice and School Productivity: Could
School Choice Be a Tide that Lifts All Boats?” In C. M. Hoxby, ed.,
The Economics of School Choice, 287-342. National Bureau of
Economic Research Conference Report. Chicago: University of
Chicago Press.
Hsieh, C., and M. Urquiola. 2006. “The Effects of Generalized School
Choice on Achievement and Stratification: Evidence from Chile’s
Voucher Program.” Journal of Public Economics 90, no. 8-9
(September): 1477-1503.
Innerst, C. 2000. “Competing to Win: How Florida’s A+ Plan Has
Triggered Public School Reform.” Available at http://
www.edreform.com/published_pdf/Competing_To_Win_
How_Floridas_A_Plus_Plan_Has_Triggered_Public_
School_Reform.pdf.
Manski, C. F. 1992. “Educational Choice (Vouchers) and Social
Mobility.” Economics of Education Review 11, no. 4
(December): 351-69.
McMillan, R. 2004. “Competition, Incentives, and Public School
Productivity.” Journal of Public Economics 88, no. 9-10
(August): 1871-92.
Nechyba, T. J. 1996. “Public School Finance in a General Equilibrium
Tiebout World: Equalization Programs, Peer Effects, and Private
School Vouchers.” National Bureau of Economic Research
Working Paper no. 5642, June.
———. 1999. “School-Finance-Induced Migration and Stratification
Patterns: The Impact of Private School Vouchers.” Journal of
Public Economic Theory 1, no. 1 (January): 5-50.
———. 2000. “Mobility, Targeting, and Private-School Vouchers.”
American Economic Review 90, no. 1 (March): 130-46.

Greene, J. P., and M. A. Winters. 2003. “When Schools Compete:
The Effects of Vouchers on Florida Public School Achievement.”
Manhattan Institute for Policy Research Education Working
Paper no. 2, August.

———. 2003. “Introducing School Choice into Multidistrict Public
School Systems.” In C. M. Hoxby, ed., The Economics of School
Choice, 145-94. National Bureau of Economic Research
Conference Report. Chicago: University of Chicago Press.

Hoxby, C. M. 2003a. “School Choice and School Competition:
Evidence from the United States.” Swedish Economic Policy
Review 10, no. 2: 9-65.

Rockoff, J. E., and L. J. Turner. 2008. “Short-Run Impacts of
Accountability on School Quality.” National Bureau of
Economic Research Working Paper no. 14564, December.

FRBNY Economic Policy Review / October 2010

References (Continued)

West, M. R., and P. E. Peterson. 2005. “The Efficacy of Choice Threats
within School Accountability Systems: Results from Legislatively
Induced Experiments.” Harvard University Program on Education
Policy and Governance, PEPG no. 05-01.

Witte, J. F. 2000. The Market Approach to Education:
An Analysis of America’s First Voucher Program.
Princeton, N.J.: Princeton University Press.

Winters, M. A. 2008. “Grading New York: An Evaluation of New York
City’s Progress Report Program.” Manhattan Institute for Policy
Research Civic Report no. 55, November.

The views expressed are those of the author and do not necessarily reflect the position of the Federal Reserve Bank of New York
or the Federal Reserve System. The Federal Reserve Bank of New York provides no warranty, express or implied, as to the
accuracy, timeliness, completeness, merchantability, or fitness for any particular purpose of any information contained in
documents produced and provided by the Federal Reserve Bank of New York in any form or manner whatsoever.
22

Program Design, Incentives, and Response

Argia M. Sbordone, Andrea Tambalotti, Krishna Rao, and Kieran Walsh

Policy Analysis Using DSGE
Models: An Introduction
• Dynamic stochastic general equilibrium
models are playing an important role in the
formulation and communication of monetary
policy at many of the world’s central banks.

• These models, which emphasize the
dependence of current choices on expected
future outcomes, have moved from academic
circles to the policymaking community—but
they are not well known to the general public.

• This study adds to the understanding of the
DSGE framework by using a small-scale
model to show how to address specific
monetary policy questions; the authors
focus on the causes of the sudden pickup
in inflation in the first half of 2004.

• An important lesson derived from the exercise
is that the management of expectations can
be a more effective tool for stabilizing inflation
than actual movements in the policy rate;
this result is consistent with the increasing
focus on central bankers’ pronouncements
of their future actions.

Argia M. Sbordone is an assistant vice president and Andrea Tambalotti a
senior economist at the Federal Reserve Bank of New York; Krishna Rao is
a graduate student at Stanford University; Kieran Walsh is a graduate student
at Yale University.
Correspondence: argia.sbordone@ny.frb.org, andrea.tambalotti@ny.frb.org

1. Introduction

n recent years, there has been a significant evolution in
the formulation and communication of monetary policy
at a number of central banks around the world. Many of
these banks now present their economic outlook and policy
strategies to the public in a more formal way, a process
accompanied by the introduction of modern analytical tools
and advanced econometric methods in forecasting and policy
simulations. Official publications by central banks that
formally adopt a monetary policy strategy of inflation
targeting—such as the Inflation Report issued by the Bank
of England and the monetary policy reports issued by the
Riksbank and Norges Bank—have progressively introduced
into the policy process the language and methodologies
developed in the modern dynamic macroeconomic literature.1
The development of medium-scale DSGE (dynamic
stochastic general equilibrium) models has played a key role
in this process.2 These models are built on microeconomic
foundations and emphasize agents’ intertemporal choice.
The dependence of current choices on future uncertain
1

The Bank of England has published a quarterly Inflation Report since 1993.
The report sets out the detailed economic analysis and inflation projections on
which the Bank’s Monetary Policy Committee bases its interest rate decisions.
The Riksbank and Norges Bank each publish monetary policy reports three
times a year. These reports contain forecasts for the economy and an
assessment of the interest rate outlook for the medium term.
2
A simple exposition of this class of models can be found in Galí and Gertler
(2007). Woodford (2003) provides an exhaustive textbook treatment.

The authors thank Ariel Zetlin-Jones for his contribution to the early stages
of this work. The views expressed are those of the authors and do not
necessarily reflect the position of the Federal Reserve Bank of New York
or the Federal Reserve System.

FRBNY Economic Policy Review / October 2010

outcomes makes the models dynamic and assigns a central
role to agents’ expectations in the determination of current
macroeconomic outcomes. In addition, the models’ general
equilibrium nature captures the interaction between policy
actions and agents’ behavior. Furthermore, a more detailed
specification of the stochastic shocks that give rise to economic
fluctuations allows one to trace more clearly the shocks’
transmission to the economy.
The use of DSGE models as a potential tool for policy
analysis has contributed to their diffusion from academic to
policymaking circles. However, the models remain less wellknown to the general public. To broaden the understanding of
these models, this article offers a simple illustration of how an
estimated model in this class can be used to answer specific
monetary policy questions. To that end, we introduce the
structure of DSGE models by presenting a simple model,
meant to flesh out their distinctive features. Before proceeding
to a formal description of the optimization problems solved by
firms and consumers, we use a simple diagram to illustrate
the interactions among the main agents in the economy. With
the theoretical structure in place, we discuss the features of the
estimated model and the extent to which it approximates the
volatility and comovement of economic time series. We also
discuss important outcomes of the estimation—namely, the

This article offers a simple illustration of
how an estimated [DSGE] model . . . can
be used to answer specific monetary
policy questions.
possibility of recovering the structural shocks that drive
economic fluctuations as well as the historical behavior of
variables that are relevant for policy but are not directly
observable. We conclude by applying the DSGE tool to study
the role of monetary policy in a recent episode of an increase in
inflation. The lesson we emphasize is that, while they are a very
stylized representation of the real economy, DSGE models
provide a disciplined way of thinking about the economic
outlook and its interaction with policy.3
We work with a small model in order to make the
transmission mechanism of monetary policy, whose basic
contours our model shares with most DSGE specifications,
as transparent as possible. Therefore, the model focuses on
the behavior of only three major macroeconomic variables:
inflation, GDP growth, and the short-term interest rate.
3

Adolfson et al. (2007) offer a more extended illustration of how DSGE models
can be used to address questions that policymakers confront in practice. Erceg,
Guerrieri, and Gust (2006) illustrate policy simulations with an open-economy
DSGE model.

Policy Analysis Using DSGE Models: An Introduction

However, the basic framework that we present could easily
be enriched to provide more details on the structure of the
economy. In fact, a key advantage of DSGE models is that they
share core assumptions on the behavior of households and
firms, which makes them easily scalable to include details that
are relevant to address the question at hand. Indeed, several

A key advantage of DSGE models is that
they share core assumptions on the
behavior of households and firms, which
makes them easily scalable to include
details that are relevant to address the
question at hand.
extensions of the basic framework presented here have been
developed in the literature, including the introduction of wage
stickiness and frictions in the capital accumulation process
(see the popular model of Smets and Wouters [2007]) and a
treatment of wage bargaining and labor market search (Gertler,
Sala, and Trigari 2008).4 Recently, the 2008 financial crisis has
highlighted one key area where DSGE models must develop:
the inclusion of a more sophisticated financial intermediation
sector. There is a large body of work under way to model
financial frictions within the baseline DSGE framework—work
that is very promising for the study of financial intermediation
as a source and conduit of shocks as well as for its implications
for monetary policy. However, this last generation of models
has not yet been subjected to extensive empirical analysis.
Our study is organized as follows. Section 2 describes the
general structure of our model while Section 3 illustrates its
construction from microeconomic foundations. Section 4
briefly describes our approach to estimation and presents some
of the model’s empirical properties. In Section 5, we use the
model to analyze the inflationary episode of the first half of
2004. Section 6 concludes.

2. DSGE Models and Their
Basic Structure
Dynamic stochastic general equilibrium models used for policy
analysis share a fairly simple structure, built around three
interrelated blocks: a demand block, a supply block, and a
4

Some of these larger DSGE models inform policy analysis at central banks
around the world: Smets and Wouters (2007) of the European Central Bank;
Edge, Kiley, and Laforte (2008) of the Federal Reserve System; and Adolfson
et al. (2008) of the Riksbank.

The Basic Structure of DSGE Models
Mark-up
shocks

Demand
shocks

Y = f YY(Y e, i - π ee,...)
Demand

Productivity
shocks

π = f π(π e,Y,...)
Supply

Y e,π e
Expectations

i = f i(π − π *,Y,...)
Monetary policy

Policy
shocks

monetary policy equation. Formally, the equations that define
these blocks derive from microfoundations: explicit
assumptions about the behavior of the main economic actors
in the economy—households, firms, and the government.
These agents interact in markets that clear every period, which
leads to the “general equilibrium” feature of the models.
Section 3 presents the microfoundations of a simple DSGE
model and derives the equations that define its equilibrium.
But first, we begin by introducing the basic components
common to most DSGE models with the aid of a diagram.
In the diagram, the three interrelated blocks are depicted as
rectangles. The demand block determines real activity  Y  as
a function of the ex ante real interest rate—the nominal rate
e
minus expected inflation  i –   —and of expectations about
e
future real activity  Y  . This block captures the idea that, when
real interest rates are temporarily high, people and firms would
rather save than consume or invest. At the same time, people
are willing to spend more when future prospects are promising
e
( Y is high), regardless of the level of interest rates.
The line connecting the demand block to the supply block
shows that the level of activity  Y  emerging from the demand
block is a key input in the determination of inflation   ,
e
together with expectations of future inflation   . In
prosperous times, when the level of activity is high, firms must
increase wages to induce employees to work longer hours.
Higher wages increase marginal costs, putting pressure on
prices and generating inflation. Moreover, the higher inflation
is expected to be in the future, the higher is this increase in
prices, thus contributing to a rise in inflation today.
The determination of output and inflation from the
demand and supply blocks feeds into the monetary policy
block, as indicated by the dashed lines. The equation in that
block describes how the central bank sets the nominal interest

rate, usually as a function of inflation and real activity. This
reflects the tendency of central banks to raise the short-term
interest rate when the economy is overheating as well as when
inflation rises and to lower it in the presence of economic slack.
By adjusting the nominal interest rate, monetary policy in turn
affects real activity and through it inflation, as represented by
the line flowing from the monetary policy block to the demand
block and then to the supply block. The policy rule therefore
closes the circle, giving us a complete model of the relationship
between three key endogenous variables: output  Y , inflation
  , and the nominal interest rate  i .
While this brief description appears static, one of the
fundamental features of DSGE models is the dynamic
interaction between the blocks—hence, the “dynamic” aspect
of the DSGE label—in the sense that expectations about the
future are a crucial determinant of today’s outcomes. These
expectations are pinned down by the same mechanism that
generates outcomes today. Therefore, output and inflation

One of the fundamental features of
DSGE models is the dynamic interaction
between the [three interrelated] blocks—
hence, the “dynamic” aspect of the DSGE
label—in the sense that expectations
about the future are a crucial determinant
of today’s outcomes.
tomorrow, and thus their expectations as of today, depend on
monetary policy tomorrow in the same way as they do today—
of course, taking into account what will happen from then
on into the infinite future.
The diagram highlights the role of expectations and the
dynamic connections between the blocks that they create.
The influence of expectations on the economy is represented by
the arrows, which flow from monetary policy to the demand
and then the supply block, where output and inflation are
determined. This is to emphasize that the conduct of monetary
policy has a large influence on the formation of expectations.
In fact, in DSGE models, expectations are the main channel
through which policy affects the economy, a feature that is
consistent with the close attention paid by financial markets
and the public to the pronouncements of central banks on their
likely course of action.
The last component of DSGE models captured in the
diagram is their stochastic nature. Every period, random
exogenous events perturb the equilibrium conditions in each
block, injecting uncertainty in the evolution of the economy

FRBNY Economic Policy Review / October 2010

and thus generating economic fluctuations. Without these
shocks, the economy would evolve along a perfectly predictable
path, with neither booms nor recessions. We represent these
shocks as triangles, with arrows pointing toward the
equilibrium conditions on which they directly impinge. Markup and productivity shocks, for example, affect the pricing and
production decisions of firms that underlie the supply block,
while demand shocks capture changes in the willingness of
households to purchase the goods produced by those firms.

3. Microfoundations of a Simple
DSGE Model
We present the microfoundations of a small DSGE model that
is simple enough to fit closely into the stylized structure
outlined in our diagram. Our objective is to describe the basic
components of DSGE models from a more formal perspective,
using the mathematical language of economists, but avoiding
unnecessary technical details. Despite its simplicity, our model
is rich enough to provide a satisfactory empirical account of the
evolution of output, inflation, and the nominal interest rate in
the United States in the last twenty years, as we discuss in the
next section.
Given the constraints we impose on this treatment for the
sake of simplicity, our model lacks many features that are
standard in the DSGE models that central banks typically use.
For example, we ignore the process of capital accumulation,

Despite its simplicity, our model is rich
enough to provide a satisfactory empirical
account of the evolution of output,
inflation, and the nominal interest rate in
the United States in the last twenty years.

which would add another dimension—investment decisions
by firms—to the economy’s demand block. Nor do we attempt
to model the labor market in detail: for example, we make no
distinction between the number of hours worked by each
employee and the number of people at work, an issue that is
hard to overlook in a period with unemployment close to
10 percent. Finally, we exclude any impediment to the smooth
functioning of financial markets and assume that the central
bank can perfectly control the short-term interest rate—the
only relevant rate of return in the economy. The 2008 financial
crisis has proved that this set of assumptions can fail miserably

Policy Analysis Using DSGE Models: An Introduction

in some circumstances and has highlighted the need for a
more nuanced view of financial markets within the current
generation of DSGE models, as we observe in our introduction.

3.1. The Model Economy
Our model economy is populated by four classes of agents:
a representative household, a representative final-goodproducing firm (f-firm), a continuum of intermediate firms
(i-firms) indexed by i   0 1, and a monetary authority. The
household consumes the final good and works for the i-firms.
Each of these firms is a monopolist in the production of a
particular intermediate good i , for which it is thus able to

Our model economy is populated by
four classes of agents: a representative
household, a representative final-goodproducing firm, . . . a continuum of
intermediate firms, . . . and a monetary
authority.

set the price. The f-firm packages the differentiated goods
produced by the i-firms and sells the product to households in
a competitive market. The monetary authority sets the nominal
interest rate.
The remainder of this section describes the problem faced
by each economic agent, shows the corresponding optimization conditions, and interprets the shocks that perturb these
conditions. These optimization conditions result in dynamic
relationships among macroeconomic variables that define the
three model blocks described above. Together with market
clearing conditions, these relationships completely characterize
the equilibrium behavior of the model economy.

3.2. Households and the Aggregate
Demand Block
At the core of the demand side of virtually all DSGE models
is a negative relationship between the real interest rate and
desired spending. In our simple model, the only source of
spending is consumption. Therefore, the negative relationship
between the interest rate and demand emerges from the
consumption decision of households.

We model this decision as stemming from the optimal
choice of a very large representative household—the entire
U.S. population—which maximizes its expected discounted
lifetime utility, looking forward from an arbitrary date t 0

s
Et
  bt + s log  Ct + s
Max
0
0
0

s = 0 s = 0 
 Bt + s Ct + s  Ht + s  i  

To find the solution to the optimal problem above, we form
the Lagrangian
 
s
 bt 0 + s  log  Ct + s –  Ct + s – 1 
L = Et
0
0
0

s=0



– Ct

i   0 1
1

0+s–1

0

 – v  Ht

0+s

B
Pt C t + -----t  B t – 1 +
Rt

–t


i  di 


subject to the sequence of budget constraints

– Bt

0+s

 Pt

0+s–1

0+s

–

 0 Wt

0+s

i  di

+ Bt

0 +s

–1

0+s

 ,



i  Ht + s i  di

0 +s
0

with first-order conditions

 0 wt  i Ht  i  di ,

for t = t 0  t 0 + 1  ,, and given Bt – 1. The members of this
0
household, we call them “Americans,” like consumption, Ct ,
but dislike the number of hours they spend at work, Ht , to
an extent described by the convex function  . The utility
flow from consumption depends on current as well as past
consumption, with a coefficient  . As a result of this “habit,”
consumers are unhappy if their current consumption is low,
but also if it falls much below the level of their consumption in
the recent past. To afford consumption, Americans work a
certain amount of hours Ht i  in each of the i-firms, where they
earn an hourly nominal wage Wt i  which they take as given
when deciding how much to work.5 With the income thus
earned, they can purchase the final good at price Pt or save by
accumulating one-period discount government bonds Bt ,
whose gross rate of return between t and t + 1 is Rt .
From the perspective of time t , the household discounts
utility in period t + 1 by a time-varying factor  bt + 1  bt , where
bt + 1  bt is an exogenous stochastic process. Changes in bt + 1  bt
represent shocks to the household’s impatience. When bt + 1
increases relative to bt , for example, the household cares more
about the future and thus wishes to save more and consume
less today, everything else equal. In this respect, bt + 1  bt acts as
a traditional demand shock, which affects desired consumption
and saving exogenously. A persistent increase in bt + 1  bt is one
way of interpreting the current macroeconomic situation in the
United States, in which households have curtailed their
consumption—partly to build their savings. Of course, in
reality there are many complex reasons behind this observed
change in behavior, and an increase in people’s concern about
the future is surely one of them. For simplicity, the model
focuses on this one reason exclusively.
5

0

– v  Ht

In equilibrium, the wage—and thus the number of hours worked—will settle
at the level at which the supply of labor by the household equals the demand of
labor by firms. This labor demand in turn is determined by the need of firms
to hire enough workers to satisfy the demand for their products, as we describe
in the next section.

(3.1a)
(3.1b)

L------:  t =  Et   t + 1 R t
Bt

 bt + 1  bt
L 
1
-------- : -----t P = -------------------------- .
–  Et -------------------------C
Ct bt t
Ct –  Ct – 1
t + 1 –  Ct

for t = t 0 t 0 + 1 , and
(3.2)

  Ht  i  
L
--------------- : --------------------- = Wt  i 
Ht  i 
t  bt

for t = t 0 t 0 + 1 , and i   0 1  , together with the
sequence of budget constraints.
These conditions yield a fully state-contingent plan for the
household’s choice variables—how much to work, consume,
and save in the form of bonds—looking forward from the
planning date t 0 and into the foreseeable future. At any point
in time, the household is obviously uncertain about the way
in which this future will unfold. However, we assume that the
household is aware of the kind of random external events, or
shocks, that might affect its decisions and, crucially, that it
knows the probability with which these shocks might occur.
Therefore, the household can form expectations about future
outcomes, which are one of the inputs in its current choices.
We assume that these expectations are rational, meaning that
they are based on the same knowledge of the economy and of
the shocks that buffet it as that of the economist constructing
the model. We use the notation Et  X t + s  to denote
expectations formed at time t of any future variable X t + s ,
as in equations 3.1, for example. The optimal plan, then,
is a series of instructions on how to behave in response to the
realization of each shock, given expectations about the future,
rather than a one-time decision on exactly how much to
work, consume, and save on each future date.
Together, the optimality conditions in equations 3.1
establish the negative relationship between the interest rate and
desired consumption that defines the demand side of the
model. The nature of this relationship is more transparent in
the special case of no habit in consumption ( = 0 ), when we

FRBNY Economic Policy Review / October 2010

can combine the two equations to obtain
(3.3)

Rt
 bt + 1 1
1
- .
- ----------- ----------------------- = E t ------------bt C t + 1 Pt + 1  Pt
Ct

According to this so-called Euler equation, desired
consumption decreases when the (gross) real interest rate
Rt 
 ------------------- increases, when expected future consumption
 Pt + 1  Pt 
decreases, and when households become more patient
( bt + 1 rises).
A log-linear approximation of the Euler equation (3.3),
after some manipulation, gives
(3.4)

yt = Et yt + 1 –  i t – Et  t + 1  –  t ,

where  t  log Pt  Pt – 1 is the quarterly inflation rate, i t  log Rt
is the continuously compounded nominal interest rate,
 t  Et log   bt + 1  bt  is a transformation of the demand shock,
and y t  log Yt is the logarithm of total output. In this
expression, we can substitute consumption of the final good Ct
with its output Yt because in our model consumption is the
only source of demand for the final good. Therefore, market
clearing implies Yt = Ct .
In this framework, equation 3.4 is similar to a traditional IS
equation, since it describes the relationship between aggregate
activity y t and the ex ante real interest rate i t – Et  t + 1 , which
must hold for the final-good market to clear. Unlike a
traditional IS relationship, though, this equation is dynamic
and forward looking, as it involves current and future expected
variables. In particular, it establishes a link between current
output and the entire future expected path of real interest rates,
as we see by solving the equation forward


(3.5)

yt = – Et

  it + s – t + s + 1 – t + s  .
s=0

Through this channel, expectations of future monetary policy
directly affect current economic conditions. In fact, this
equation shows that future interest rates are just as important
to determine today’s output as the current level of the shortterm rate, as we describe in our discussion of the role of policy
expectations.
In our full model, the Euler equation is somewhat more
complicated than in equation 3.4 because of the consumption
habit (  0), which is a source of richer, and more realistic,
output dynamics in response to changes in the interest rate.
Nevertheless, these more intricate dynamics do not change the
qualitative nature of the relationship between real rates and
demand.
The third first-order condition of the household
optimization problem, equation 3.2, represents the labor
supply decision. It says that Americans are willing to work
more hours in firms that pay a higher wage and at times when

Policy Analysis Using DSGE Models: An Introduction

wages are higher, at least for differences in wages modest
enough to have no significant effect on their income.6 Large
wage changes, in fact, would trigger an income effect and
lead the now richer workers to curtail their labor supply.
Mathematically, workers with higher income could afford
more consumption, which would lead to a drop in the marginal
utility  t and thus to a decrease in labor supply at any given
wage level.
We can think of the labor supply schedule (equation 3.2)
as a relationship determining the wage that firms must pay
to induce Americans to work a certain number of hours.
In prosperous times, when demand is high and firms are
producing much, firms require their labor force to work long
hours and they must correspondingly pay higher hourly wages.
This is an important consideration in the production and
pricing decisions of firms, as we discuss in the next section.

3.3. Firms and the Aggregate Supply Block
The supply block of a DSGE model describes how firms set
their prices as a function of the level of demand they face. Recall
that in prosperous times, demand is high and firms must pay
their workers higher wages. As a result, their costs increase as
do their prices. In the aggregate, this generates a positive
relationship between inflation and real activity.
In terms of microfoundations, establishing this relationship
requires some work, since firms must have some monopoly
power to set prices. This is why our production structure
includes a set of monopolistic i-firms, which set prices, as well
as an f-firm, which simply aggregates the output of the i-firm
into the final consumption good. Because all the pricing action
occurs within the i-firms, we focus on their problem and omit
that of the f-firm.
Intermediate firm i hires Ht  i  units of labor of type i on
a competitive market to produce Yt  i  units of intermediate
good i with the technology
(3.6)

Yt  i  = At Ht  i  ,

where At represents the overall efficiency of the production
process. We assume that At follows an exogenous stochastic
process, whose random fluctuations over time capture the
unexpected changes in productivity often experienced by
modern economies—for example, the productivity boom in
the mid-1990s that followed the mass adoption of information
technology. We call this process an aggregate productivity shock,
since it is common to all firms.
Labor supply is upward sloping because   is an increasing function, as  is
convex. In other words, people dislike working an extra hour more intensely
when they are already working a lot rather than when they are working little.

The market for intermediate goods is monopolistically
competitive, as in Dixit and Stiglitz (1977), so firms set prices
subject to the requirement that they satisfy the demand for
their good. This demand comes from the f-firm and takes
the form
Pt  i  –t
(3.7)
,
Yt  i  = Yt  ----------Pt 
where Pt  i  is the price of good i and  t is the elasticity of
demand. When the relative price of good i increases, its
demand falls relative to aggregate demand by an amount
that depends on  t .
Moreover, we assume that firms change their prices only
infrequently. The fact that firms do not adjust prices
continuously, but leave them unchanged in some cases for long
periods of time, should be familiar from everyday experience
and is well established in the economic literature (for example,
Bils and Klenow [2004]; Nakamura and Steinsson [2008]). To
model this fact, we follow Calvo (1983) and assume that in
every period only a fraction 1 –  of firms is free to reset its
price while the remaining fraction maintains its old price.7
The subset of firms that are able to set an optimal price at t ,
call it  t   0 1  , maximize the discounted stream of expected
future profits, taking into account that s periods from now
s
there is a probability  that they will be forced to retain the
price chosen today. The objective function of each of these
firms is therefore
s

s  t + s
----------------  Pt  i Yt + s  i  – Wt + s  i  Ht + s  i  

Max E t
t
P
Pt  i 



s=0

for all i   t , subject to the production function 3.6 and to the
additional constraint that they must satisfy the demand for
their product at every point in time
Pt  i   –t + s
(3.8)
,
Yt + s  i  = Yt + s  ---------- Pt + s 
for s = 0 , 1, , . Profits, which are given by total revenue
at the price chosen today, Pt  i  Yt + s  i  minus total costs
s
Wt + s  i  Ht + s  i  , are discounted by the multiplier   t + s   t ,
sometimes called a stochastic discount factor, which translates
dollar profits in the future into a current dollar value.
The first-order condition of this optimization problem is

t + s – 1
Wt + s i 
s
- = 0,
   t + s Yt + s Pt + s
Pt   i  –  t + s -----------------(3.9) E t
At + s



s=0

for all i   t , where Pt  i  denotes the optimal price chosen by
firm i , Wt + s i   At + s is the firm’s nominal marginal cost at time
t + s –1
t + s , and  t + s = ---------------- is its desired mark-up—the mark-up
t + s
In our estimated model in Section 5, we actually assume that the  fraction
of firms that cannot choose their price freely can in fact adjust it in part to catch
up with recent inflation. This assumption improves the model’s ability to fit the
data on inflation, but it would complicate our presentation of the model’s
microfoundations without altering its basic message.

it would charge if prices were flexible. As rational monopolists,
optimizing firms set their price as a mark-up over their
marginal cost. However, this relationship holds in expected
present discounted value, rather than every period, since a
s
price chosen at time t will still be in effect with probability 
in period t + s.
We can rewrite the marginal cost of a firm that at time t + s
is still forced to retain the price Pt i  as
(3.10)

Wt + s  i    H t + s  i   1
S t + s  i   ------------------ = ----------------------------- ---------- t + s  bt + s At + s
At + s
P i  –t + s 
Y

A t + s Pt + s
= --------------------------------------------------,
A t + s  t + s  bt + s
t + s t
- -----------
  ---------


where we use the labor supply relation 3.2 to substitute for the
wage as well as the production function 3.6 and the demand
function 3.8 to give us an expression for the labor demand
Ht + s i .8 Inserting this expression into the first-order
condition 3.9, we see that the solution to the optimal pricing
problem is the same for all firms in the set  t , since it depends

only on the aggregate variables  Yt + s , At + s , Pt + s ,  t + s  s = 0 .
We denote this common optimal price as Pt .
t + s –1
The equation for the desired mark-up—  t + s = ---------------- , also
t + s
known as Lerner’s formula—says that monopolists facing a
more rigid demand optimally charge a higher mark-up, and
thus higher prices, since their clients are less sensitive to
changes in the latter. We assume that this sensitivity—the
elasticity of demand—and thus the desired mark-up, follows
an exogenous stochastic process. Positive realizations of this
desired mark-up shock, which correspond to a fall in the
elasticity of demand, represent an increase in firms’ market
power, to which they respond by increasing their price.
Equation 3.9, together with the definition of the aggregate
price level as a function of newly set prices Pt  and of the past
price index Pt – 1
1
Pt    1 –   Pt 

 1 – t 

------------1 – t 1 – t

+  Pt

–1



yields an approximate New Keynesian Phillips curve—a
relationship between current inflation, future expected
inflation, and real marginal cost—of the form
(3.11)

 t =  st +  Et  t + 1 + ut ,

where ut =  log ut is a transformation of the mark-up shock
and st  log  St  Pt  is the logarithm of the real marginal cost.9
The sensitivity of inflation to changes in the marginal cost,  ,
depends on the frequency of price adjustment  , as well as on

These substitutions are equivalent to “solving” for equilibrium in the labor
market.
9
Variables are in logarithms, because equation 3.11, like equation 3.4, is
obtained by a log-linear approximation.

FRBNY Economic Policy Review / October 2010

1 –    1 –   ,
other structural parameters, according to   -------------------------------------  1 +  
 H
where   ----------- is the elasticity of the marginal disutility of

work, while  is the average value of the elasticity of demand  t .
This Phillips curve, together with the expression for
marginal costs (3.10), provides the relationship between
inflation and real activity that defines the supply block of the
model. In fact, we see from equation 3.10 that marginal cost
depends on the level of aggregate activity, among other factors.
Higher economic activity leads to higher wages and marginal
costs. Thus, firms increase their prices, boosting aggregate
inflation.
Another important feature of the Phillips curve is that it is
forward looking, just as the Euler equation in the previous
section is. As in that case, therefore, we can iterate equation
3.11 forward to obtain


+ y  yt – yt   +t ,
e

where rt ,  t, and yt are the baselines for the real interest rate,
4Q
inflation, and output, respectively, and  t   log Pt  Pt – 4  is
the rate of inflation over the previous four quarters. The
i
monetary policy shock  t , a random variable with mean zero,
captures any deviation of the observed nominal interest rate
from the value suggested by the rule. This rule implies that,
if inflation and output rise above their baseline levels, the
nominal interest rate is lifted over time above its own baseline,
e
rt +  t, by amounts dictated by the parameters  and  y and
at a speed that depends on the coefficient  . The higher policy
rate, which is expected to persist even after output and inflation
have returned to normal, exerts a restraining force on the

 t = E t     st + s + u t + s  ,
s

When output is at its efficient level,
inflation is not stable, as policymakers
would like it to be, but fluctuates because
of the presence of mark-up shocks. This
is the essence of the monetary policy
trade-offs in the economy.

s=0

which highlights how inflation today really depends on the
entire future expected path of marginal costs, and through
those, of real activity. But this path depends in turn on
expectations about interest rates, and thus on the entire future
course of monetary policy, as equation 3.5 shows. Hence,
we have the crucial role of policy expectations for the
determination of current economic outcomes in this model,
a feature we discuss in Section 2.

Monetary Policy
Recall that when the interest rate—current and expected—is
low, people demand more consumption goods (equation 3.5).
But if demand is high, firms’ marginal costs increase and so do
their prices. The end result is inflation. The opposite is true
when the interest rate is high. But where does the interest rate
come from? In DSGE models, as in the real world, the shortterm interest rate is set by monetary policy. In practice, this is a
decision made by a committee (the Federal Reserve’s Federal
Open Market Committee, or FOMC) using various inputs:
large data sets, projections from several models, and the
judgment of policymakers. Despite the apparent complexity of
the process, Taylor (1993) famously demonstrated that it can
be reasonably well approximated by assuming that the Federal
Reserve raises the federal funds rate when inflation and/or
output is “high” with respect to some baseline. This behavior is
assumed in almost all variants of DSGE models, although the
definition of the appropriate baselines is somewhat
controversial.
In our model, we assume that interest rates are set according
to the policy rule

i t =  i t – 1 +  1 –    r t +  t +    t –  t 

(3.12)

Policy Analysis Using DSGE Models: An Introduction

economy—curbing demand, marginal costs, and inflation. In
e
this respect,  t and yt can be regarded as targets of monetary
policy—the levels of inflation and output that the central bank
considers consistent with its mandate—and therefore do not
elicit either a restrictive or a stimulative policy.
In equation 3.12, we denote the central bank’s objective in
e
terms of production as yt , the “efficient” level of output. This
unobserved variable can be derived from the microfoundations
of the model.10 It represents the level of output that would
prevail in the economy if we could eliminate at once all
distortions—namely, force the i-firms to behave competitively
rather than as monopolists and allow them to change their
prices freely. The level of activity that would result from such
a situation is ideal from the perspective of the representative
household in the model, as its name suggests. This is what
makes it a suitable target for monetary policy. However,
when output is at its efficient level, inflation is not stable, as
policymakers would like it to be, but fluctuates because of
the presence of mark-up shocks. This is the essence of the
monetary policy trade-offs in the economy. Achieving the
10

The precise mathematical definition of efficient output in the model is
irrelevant for our purposes. We present in Section 4 an estimate of the behavior
of this variable over the last twenty years.

efficient level of output requires undesirable movements in
inflation. In contrast, a stable inflation implies deviations from
the efficient level of output. The two objectives cannot be
reconciled, but must be traded off of each other.
Related to the efficient level of output is the efficient real
e
interest rate, rt , which is the rate of return we would observe
in the efficient economy described above. This definition
implies that, when the actual real interest rate is at its efficient
level and is expected to remain there in the future, output will
e
also be at its efficient level. This is why we include rt in our
definition
of the baseline interest rate.
The other component of this baseline rate is the inflation
target  t. We allow this target to vary slowly over time to
accommodate the fact that inflation hovered at about 4 percent
for a few years around 1990 before declining to nearly 2 percent
after the recession that ended in 1991. Nominal interest rates
were correspondingly higher in the first period, thus implying
a stable average for the real interest rate. We now present our
estimate of the evolution of the inflation target.

4. Empirical Approach
and Estimation Results
We estimate our model using data on the growth rate of real
GDP to measure output growth,  yt , the growth rate of the
personal consumption expenditures chain price index
excluding food and energy (core PCE) to measure inflation,  t ,
and the quarterly average of the monthly effective federal funds
rate to measure the nominal interest rate, i t . We measure
inflation by core PCE, rather than by a more comprehensive
measure, because the monetary policy debate in the United
States tends to focus on this index.
Our data span the period 1984:1 to 2007:4 (Chart 1). This is
the longest possible data set over which it is reasonable to argue
that U.S. monetary policy can be represented by a stable
interest rate rule. It follows the period of extremely high
interest rates in the early 1980s that brought inflation under
control. However, in the first few years of this sample, inflation
and the federal funds rate were still relatively high, with a fairly
abrupt reduction taking place around the 1991 recession. We
capture this low-frequency movement in inflation and the
nominal interest rate by including the slow-moving inflation
target  t in the policy rule.
We use Bayesian methods to characterize the posterior
distribution of the parameters of the model. This distribution
combines the model’s likelihood function with prior
information on the parameters, using techniques surveyed, for

Chart 1

Observable Variables
Percent, annualized
12
Federal funds rate

10
8
6
4
2
0

Inflation
-2

GDP growth

-4
1984 86

Sources: Bureau of Economic Analysis; Board of Governors
of the Federal Reserve System.

example, by An and Schorfheide (2007).11 A discussion of these
methods is beyond the scope of this article. Instead, we focus on
the implications of these estimates for some key properties of
the model. Our objective is to show that the estimated model
provides a good fit to the data across many dimensions, but
also to highlight some of the model’s most notable
shortcomings.

4.1. Moment Comparisons
We compare the second moments implied by the estimated
model with those measured in the data. Table 1 presents the
standard deviations of the observable variables, reported as
annualized percentages. In the model column, we report the
median and 5th and 95th percentiles of the standard deviations
across draws from the model’s posterior. This interval reflects
the uncertainty on the structural parameters—and thus on
the model-implied moments—encoded in the parameters’
posterior distribution. In the data column, we report the
median and 5th and 95th percentiles of the empirical standard
deviations in the data. This interval represents the uncertainty
on the true empirical moments because of the small sample
available for their estimation.
Our model does a very good job replicating the volatilities in
the data. It captures the standard deviation of output growth
and replicates quite closely the volatility of the nominal interest
rate, although it overestimates the standard deviation of
11

The technical appendix provides information on the priors for the
parameters.

FRBNY Economic Policy Review / October 2010

Table 1

Model-Implied and Empirical Standard Deviations
Percent
Variable
GDP growth
Core PCE inflation
Federal funds rate

Model

Data

2.03
[1.79, 2.37]
1.41
[0.98, 2.40]
2.23
[1.61, 3.56]

2.03
[1.74, 2.27]
1.15
[0.67, 1.38]
2.46
[1.55, 2.94]

Source: Authors’ calculations.
Notes: The table reports the standard deviations of the observable
variables. The model-implied standard deviations are medians across
draws from the posterior; the 5th and 95th posterior percentiles across
those same draws are in brackets. The empirical standard deviations are
medians across bootstrap replications of a VAR(4) fit to the data; the
5th and 95th percentiles across those same replications are in brackets.
PCE is personal consumption expenditures.

inflation. The ability of the model to accurately reproduce the
volatility of the observable variables is not a preordained
conclusion, even if we freely estimate the standard deviations
of the shocks. The reason is that a likelihood-based estimator
tries to match the entire autocovariance function of the data,
and thus must strike a balance between matching standard
deviations and all the other second moments—namely,
autocorrelations and cross-correlations.
These other moments are displayed in Chart 2. The black
line represents the model-implied correlation, with the shaded
area representing a 90 percent posterior interval. The solid blue

Our model does a very good job
replicating the volatilities in the data.
It captures the standard deviation of
output growth and replicates quite closely
the volatility of the nominal interest rate,
although it overestimates the standard
deviation of inflation.
line is instead the correlation measured in the data, with a
90 percent bootstrap interval around this estimate represented
by the dashed lines. The serial correlation of output growth in
the model is very close to its empirical counterpart and well
within the data-uncertainty band. For inflation and the interest
rate, the model serial correlations are on the high end of the

Policy Analysis Using DSGE Models: An Introduction

band. This excessive persistence is a result of the low-frequency
component in both variables associated with the inflation
target, as we can infer from the variance decomposition
in Table 2.
According to the model, shocks to the inflation target
account for 85 percent of the unconditional variance of
inflation and 38 percent of that of the nominal interest rate.
Although we do not calculate a variance decomposition by
frequency, we know that the contribution of the inflation target
shock is concentrated at low frequencies, since this shock is
very persistent (the posterior mean of its autocorrelation
coefficient is 0.98). This finding suggests that the model faces
a trade-off between accommodating the downward drift in
inflation in the first part of our sample and providing a more
balanced account of the sources of inflation variability.
The rest of the variance decomposition accords well with
conventional wisdom. The productivity shock plays an

According to the model, shocks to the
inflation target account for 85 percent of
the unconditional variance of inflation and
38 percent of that of the nominal interest
rate. This finding suggests that the model
faces a trade-off between accommodating
the downward drift in inflation in the first
part of our sample and providing a more
balanced account of the sources of
inflation variability.
important role in accounting for the volatility of output
growth, although the demand shock and the monetary policy
shocks (interest rate plus inflation target) are also nonnegligible. Moreover, the demand shock accounts for more
than half of the variance of the nominal interest rate. Finally,
mark-up shocks play a minor role as sources of volatility.
This finding has potentially important policy implications,
since in the model mark-up shocks are the only source of
the aforementioned policy trade-off between inflation and
real activity.
Returning now to the cross-correlations in Chart 2, we see
that the model is quite successful in capturing the lead-lag
relationships in the data. In our sample, there is no statistically
significant predictability of future inflation through current
output growth, a pattern that is reproduced by the model. The

Chart 2

Correlations
Corr. (Δ yt , π t+k )

Corr. ( Δyt , Δ yt-k )
1.0

0.4

0.8

0.3
0.2

0.6

0.1

0.4

0
0.2

-0.1

-0.2

-0.3
-0.4

-0.4
Corr. ( π t , π t-k )

Corr. ( π t , it+k )

1.0

0.8

0.6

0.4

0.2

-0.2
Corr. ( it , Δ yt+k )

Corr. ( it , i t-k )
1.0

0.6

0.8

0.4

0.6

0.2
0.4

0
0.2

-0.2

0
-0.2

4
k

-0.4
-8 -7 -6 -5 -4 -3 -2 -1

0
k

Source: Authors’ calculations.
Notes: The black line represents the median model-implied correlation across draws from the posterior; the shaded area represents the interval between
the 5th and 95th percentiles across those same draws. The solid blue line represents the median autocorrelation across bootstrap replications of a VAR(4)
fit to the data; the dashed blue lines represent the interval between the 5th and 95th percentiles across those same replications. Each statistic is calculated
at horizons k = 0, . . . 8 for autocorrelations and at horizons k = -8, . . . , 0, . . . 8 for cross-correlations.

model also reproduces the positive correlation between
inflation and the nominal interest rate present in the data both
in the leads and in the lags (the middle right panel of the chart).
The positive correlation between current interest rates and
future inflation might seem puzzling at first. We would expect
higher interest rates to bring inflation down over time, which
should make the correlation negative. However, over our

sample, this negative relationship is confounded by the lowfrequency positive comovement between inflation and the
nominal interest rate induced by the Fisher effect. Recall that
inflation and the nominal interest rate in fact are persistently
above their unconditional sample average over the first third
of the sample and are persistently below it after about 1992.
The bottom right panel of Chart 2 reports the dynamic

FRBNY Economic Policy Review / October 2010

Table 2

Variance Decomposition
Percent
Shocks
Variable
GDP growth
Core PCE inflation
Federal funds rate

Demand

Productivity

Mark-Up

Interest Rate

Inflation Target

0.20
[.13, .28]
0.04
[.01, .06]
0.54
[.32, .76]

0.56
[.45, .67]
0.01
[.00, .02]
0.06
[.00, .13]

0.07
[.04, .10]
0.10
[.04, .17]
0.01
[.00, .02]

0.04
[.02, .06]
0.00
[.00, .01]
0.01
[.01, .02]

0.13
[.07, .19]
0.85
[.76, .94]
0.38
[.16, .61]

Source: Authors’ calculations.
Notes: The table reports the share of the unconditional variance of each observable variable contributed by each shock. The point estimates are medians
across draws from the posterior; the 5th and 95th posterior percentiles across those same draws are in brackets. PCE is personal consumption expenditures.

correlation between output growth and the nominal interest
rate. In the data, high growth rates of output predict high
nominal interest rates one to two years ahead, but this
predictability is much less pronounced in the model.
Moreover, this discrepancy is statistically significant in the
sense that the model-implied median autocorrelation lies
outside the 90 percent bootstrap interval computed from the
data. This failure to match the data highlights the main
empirical weakness of our model: its demand-side
specification. As in most of the DSGE literature, our demand
block consists of the Euler equation of a representative
consumer. Standard specifications of a Euler equation of the
type adopted here provide an inaccurate description of the
observed relationship between the growth rate of consumption
(or output, as in our case) and financial returns, including
interest rates, as first documented by Hansen and Singleton
(1982, 1983) and subsequently confirmed by many others (see
Campbell [2003] for a review). Improving the performance of
the current generation of DSGE models in this dimension
would be an important priority for future research.
We now report our estimates of a few of the variables that
play an important role in the model, but that are not directly
observable. We focus on the three latent variables that enter
the interest rate rule: the inflation target  t, the output gap
e
e
yt – yt , and the efficient real interest rate rt (Chart 3). As in
Charts 1 and 2, the black line is the median estimate across
draws from the model’s posterior and the shaded area
represents a 90 percent posterior probability interval.
Starting from the top panel, we note that the estimated
inflation target captures well the step-down in inflation from
a local mean above 4 percent between 1984 and 1991 to an
average value of around 2 percent since 1994. This permanent
reduction in inflation represents the last stage of the

Policy Analysis Using DSGE Models: An Introduction

disinflation process initiated by Federal Reserve Chairman
Volcker in 1979, which became known as an example of
opportunistic disinflation (Orphanides and Wilcox 2002).
Needless to say, the estimated target is not completely smooth,
but it also displays some higher frequency variation. For
example, it reaches a minimum of around 1 percent at the
beginning of 2003, but moves closer to 2 percent over 2004.
(The next section studies in more detail the implications of
these movements.)
The middle and bottom panels of Chart 3 report estimates
of the output gap—the percentage deviation of output from its
efficient level—and of the efficient real interest rate. Several
features of the estimated output gap are noteworthy. First, its
two deepest negative troughs correspond to the two recessions
in our sample. In this respect, our model-based output gap
conforms well with more conventional measures of this
variable, such as the one produced by the Congressional
Budget Office (CBO). However, the shortfall of output from its
efficient level is never larger than 0.7 percent, even in these
recessionary episodes. By comparison, the CBO output gap is
as low as -3.2 percent in 1991. The amplitude of the business
cycle fluctuations in our estimated output gap is small because
the efficient level of output is a function of all the shocks in the
model and therefore it tracks actual output quite closely. The
last notable feature of the efficient output gap is that it displays
a very pronounced volatility at frequencies higher than the
business cycle. During the 1990s expansion, for example, it
crosses the zero line about a dozen times.
Compared with the output gap, the efficient real rate is
significantly smoother. Although some high-frequency
variation remains, the behavior of the efficient real rate is
dominated by swings at the business cycle frequency. The rate
spikes and then plunges for some time before the onset of

Chart 3

Kalman Smoother Estimates of Latent Variables
π *t
7
6

To show how our model can be used to address specific policy
questions, we examine a particular historical episode: the
puzzling pickup in inflation in the first half of 2004. This
exercise allows us to illustrate how we use the model’s forecasts
to construct alternative scenarios for counterfactual policy
analysis. Moreover, our analysis offers potentially interesting
lessons for the current situation—although inflation has
recently been quite low, there has been some concern that
it might accelerate in the near future.
After approaching levels close to 1 percent between 2002
and 2003, core PCE inflation started moving higher in mid2003. This pickup accelerated significantly in the first half

5
4
3
2
1
0
-1
0.8
0.6

5. The Model at Work: The Pickup
in Inflation in the First Half
of 2004

yt - yte

0.4
0.2
0
-0.2

-0.4
-0.6
-0.8
-1.0
e

rt
10
8
6
4
2
0
-2

-4
1984 86

Source: Authors’ calculations.
Notes: The black line represents the Kalman smoother estimate of
the relevant latent variable conditional on the posterior mean of the
parameters; the shaded area represents the interval between the 5th
and 95th percentiles of the Kalman smoother estimates across draws
from the posterior. The vertical bands indicate NBER recessions.

recessions and recovers a few quarters into the expansions. It is
interesting to note that the efficient real rate was negative for
the entire period between 2001 and 2004—a time when the
FOMC was concerned about the possibility that the U.S.
economy would fall into a liquidity trap.12 A negative efficient
real interest rate is a necessary condition for the zero bound on
nominal interest rates to become binding, and hence for the
liquidity trap to become a problem.
12

A liquidity trap describes a situation in which nominal interest rates have
reached their zero lower bound, as in Japan in the 1990s, and therefore cannot
be lowered any further.

of 2004, when (year-over-year) core inflation moved from
about 1.5 percent to more than 2 percent, where it remained
until the end of 2008. We use our DSGE model to analyze
the sources of this unusually rapid and persistent step-up in
the level of inflation.
We organize our discussion around three questions. First,
was the surge in inflation forecastable? As we will see, the
answer to this question is no, at least from the perspective of
our model. Second, what accounts for the discrepancy between
the model’s forecast and the observed paths of inflation, output
growth, and the federal funds rate? Third, could monetary
policy have achieved a smooth transition to inflation rates
below 2 percent and, if so, at what cost in terms of added
volatility in output and the interest rate?
Chart 4 presents forecasts of quarterly core PCE inflation,
real GDP growth, and the federal funds rate from the DSGE
model. The forecast starts in 2003:1, when quarterly inflation
reached 1.1 percent (at an annual rate)—its lowest level
following the 2001 recession—and extends through the
beginning of 2005. In each panel, the dashed line represents

FRBNY Economic Policy Review / October 2010

assessed only about one in ten chances (12.5 percent) of
inflation being as high as it was in that period.
From an economic perspective, it is interesting to note that
these sizable forecast errors for inflation roughly correspond to
the “considerable period” era that extended from June 2003 to
June 2004. At that time, the FOMC kept the federal funds rate
constant at 1 percent to guard against the risk of deflation,
while indicating in its statement that “policy accommodation
can be maintained for a considerable period.”13 According to
the model’s projection, this path for the federal funds rate
represents a deviation from the policy stance historically
maintained by the Federal Reserve in similar macroeconomic
circumstances. Based on the estimated interest rate rule, in fact,

Chart 4

Forecasts of Observable Variables
3.5

Core PCE Inflation

2.5

1.5

0.5
-0.5
8

Real GDP Growth

The DSGE forecast is just a description of
what would happen to the variables of
interest if we allowed the model economy
to “run” from its initial condition, without
introducing any innovations. Any observed
deviation from the forecast, therefore,
must be attributable to the realization of a
particular combination of such innovations.

4
2
0
-2
5

Federal Funds Rate

4
3
2
1
0
-1
2002

2003

2004

Source: Authors’ calculations.
Notes: The dashed line represents the forecast of the relevant variable
conditional on the posterior mean of the parameters; the solid line
represents the observed realization. The shaded areas represent
50 (light blue), 75 (medium blue), and 90 percent (dark blue)
symmetric probability intervals for the forecast at each horizon.
PCE is personal consumption expenditure.

the expected value of the forecast, while the bands show the
50 (light blue), 75 (medium blue), and 90 (dark blue) percent
probability intervals. The solid line shows the realized data.
The model performs well in its forecast of output and the
federal funds rate, especially in the medium term. Inflation,
by comparison, is close to the mean forecast in 2003, but is
well above it in 2004 and beyond. Moreover, the probability
intervals for the forecast suggest that this realization of
inflation was quite unusual, as we see from the fact that the
solid line borders the 75 percent probability interval in the first
half of 2004. This means that in 2003:1, the model would have

Policy Analysis Using DSGE Models: An Introduction

the DSGE predicts a slow rise in the interest rate over 2003 and
2004. Instead, the FOMC maintained the federal funds rate at
1 percent through the first half of 2004.
However, the pickup in inflation over this period is
significantly more “unusual” than the deviation of the federal
funds rate from the historical norm. Actual inflation in 2004 is
mostly outside the 50 percent probability interval of the model
forecast (the light blue band), while the actual federal funds
rate remains well within it. Moreover, the acceleration in
inflation is not accompanied by unexpectedly high real growth,
suggesting that it cannot be fully explained by the traditional
channel of transmission from an overheated economy to
higher inflation.
What else, then, accounts for the unexpected and unlikely
deviation of inflation from the model’s forecast over 2004?
The DSGE framework provides a particularly useful way of
addressing this question. As we discuss in Section 2, the
economic outcomes predicted by the model—the levels of
inflation, output, and the interest rate—are the result of the
endogenous responses of the agents in the economy to the
13

This formulation was maintained in the FOMC statement from August 2003
to December 2003, and was later substituted with “policy accommodation can
be removed at a pace that is likely to be measured.”

Chart 5

Forecasts of Shocks
2

Demand

Productivity

0
-2
-1
-4

-2
-3

-6
1.0

2.0

Mark-Up

Inflation Target

1.5

0.5

1.0
0
0.5
-0.5

0
-0.5

-1.0
2002

2003

2004

2002

2003

2004

Source: Authors’ calculations.
Notes: The dashed line represents the forecast of the relevant shock conditional on the posterior mean of the parameters while the solid line
represents an estimate of the realization based on the Kalman smoother. The shaded area represents the 75 percent symmetric probability
interval for the forecast at each horizon.

realization of a set of exogenous processes, such as productivity
or desired mark-ups. The innovations to these driving
processes account for the deviations of the data from the
model’s forecast. In fact, the DSGE forecast is just a description
of what would happen to the variables of interest if we allowed
the model economy to “run” from its initial condition, without
introducing any innovations. Any observed deviation from the
forecast, therefore, must be attributable to the realization of a
particular combination of such innovations.14
Chart 5 depicts the combinations of exogenous driving
processes that, according to the estimated DSGE model, are
responsible for the observed path of inflation, output, and the
interest rate over the period we analyze. In each panel, the
dashed line represents the evolution of the shock associated
with the mean forecast while the solid line represents the
sequence of shocks corresponding to the actual realization of
the observable variables. As in Chart 4, the medium blue band
denotes the 75 percent probability interval for the forecast.
14

In this study, we distinguish between exogenous driving processes—shocks,
for short—and innovations. Driving processes can be autocorrelated, and thus
forecastable, while their innovations are i.i.d.

The contribution of three shocks stands out. First, the
demand shock recovers from almost -4 percent to around
-1 percent. This movement is particularly pronounced during
2004, when inflation was picking up. However, this profile
is broadly consistent with the shock’s expected evolution,
represented by the dashed line. The productivity shock is also
broadly in line with expectations, with the exception of 2003:3;
this spike in productivity accounts for the corresponding spike
in output growth in that quarter.
However, the most significant and direct contribution to the
surge in inflation comes from a sizable upward movement in
the inflation target,  t. According to our estimates, this target
moves by about 1 percentage point, from less than 1 percent to
close to 2 percent. Moreover, this movement is at the edge of
the 75 percent probability interval for the forecast, suggesting
that the realization of this driving process is indeed quite
unusual.
To quantify more directly the effect on inflation of the
unexpected increase in the implicit inflation target, we depict
what would have happened to core PCE inflation in the
absence of such an increase (Chart 6). Here, the solid line

FRBNY Economic Policy Review / October 2010

Chart 6

Conditional Forecast of Inflation
3

Core PCE Inflation

0
2002

2003

2004

Source: Authors’ calculations.
Notes: The dashed line represents a forecast of inflation conditional on
the Kalman smoother estimates of all shocks except for those to the
inflation target; the solid line represents the observed realization.
The shaded areas represent 50 (light blue), 75 (medium blue), and
90 percent (dark blue) symmetric probability intervals for this
conditional forecast. Therefore, they represent uncertainty stemming
from future realizations of the inflation target shock alone. PCE is
personal consumption expenditure.

is realized inflation. The dashed line represents the counterfactual path of inflation predicted by the model in the absence
of shocks to the inflation target. In other words, this is a
forecast for inflation, conditional on the estimated path of all
but the inflation target shock. The bands therefore represent
the usual probability intervals, but in this case they are
computed around this conditional forecast.
The chart confirms our conclusion on the role of
innovations to the inflation target in accounting for the
observed evolution of inflation. According to the model, core
inflation would not have increased to above 2 percent, as it did
for most of 2004, without the steady increase in the inflation
target over the same period. In fact, inflation would have
remained within the “comfort zone” of 1 to 2 percent.
Moreover, note that the solid line of realized inflation is mostly
inside the area associated with the 90 percent probability
interval for the conditional forecast. This suggests that the
share of the forecast error in inflation accounted for by the
innovations in the inflation target in this episode is unusually
large compared with the historical average. This is just a more
formal way of saying that the increase in the inflation target is
disproportionately responsible for the observed increase in
inflation that we examine.
The estimated rise in the implicit inflation target provides
the missing link for a unified explanation of the pickup in
inflation, the “considerable period” monetary policy, and the
absence of a concomitant acceleration in output growth. In the
model, the inflation target is the main driver of movements

Policy Analysis Using DSGE Models: An Introduction

in inflation expectations, which are a key determinant of
firms’ pricing behavior together with the amount of slack in
the economy. According to the DSGE model, therefore, a
significant fraction of the inflation acceleration in 2003-04 can
be attributed to a change in inflation expectations, driven by an
increase in the Federal Reserve’s implicit inflation target as
perceived by the private sector. This increase in the perceived
target in turn is consistent with the unusually loose stance
of monetary policy maintained by the FOMC during the
“considerable period” era.
This brings us to the third question: If the DSGE model
is correct, and the pickup in inflation in 2004 is attributable
to an increase in the implicit inflation target perceived by
the public, could the Federal Reserve have prevented this
development?
Charts 7 and 8 show the results of this counterfactual
analysis. Both charts display the data (solid line) along with the
counterfactual outcomes for the economy predicted by the
model under a policy consistent with the stabilization of core
inflation at 1.6 percent through 2004. The way in which this

According to the DSGE model . . .
a significant fraction of the inflation
acceleration in 2003-04 can be attributed
to a change in inflation expectations,
driven by an increase in the Federal
Reserve’s implicit inflation target as
perceived by the private sector.
policy is implemented, however, is different in the two cases. In
Chart 7, we present the outcomes associated with what we call
a “no-communication” monetary strategy (dashed line) while
in Chart 8 we compare these results with those that would
emerge under a “full-communication” strategy (blue line).
Under the no-communication strategy, the path for the
interest rate compatible with the desired evolution of inflation
is achieved each period through “surprise” departures from the
historical rule. In contrast, under the full-communication
strategy, the Federal Reserve implements the same path for
inflation by announcing an inflation target that is consistent
with it.15
15

Technically, in both cases we choose shocks to the monetary policy rule
that are compatible with the desired evolution of inflation, conditional on
the smoothed value of all other disturbances. Under the no-communication
i
strategy, the shocks we choose are the i.i.d. monetary shocks,  t . Under
the full-communication strategy, our chosen shocks are to the inflation

target,  t .

Chart 7

Chart 8

“No-Communication” Counterfactual

“Full-Communication” Counterfactual
3

Core PCE Inflation

Real GDP Growth

-2

Federal Funds Rate

-1

Core PCE Inflation

Real GDP Growth

Federal Funds Rate

-2

-2
2002

2003

2004

2002

2003

2004

Source: Authors’ calculations.

Notes: The dashed line represents the counterfactual evolution of the
economy predicted by our model had monetary policy been set to
achieve the path for inflation depicted in the top panel. This counterfactual is conditional on the posterior mean of the parameters. Under
the “no-communication” scenario, the desired path for inflation is
achieved by the choice of the interest rate shock while all other shocks
are set at their Kalman smoother estimate. The shaded area represents the
75 percent symmetric probability interval for the unconditional forecast,
which is the same as in Chart 4. The black line represents the observed
realization of each series. PCE is personal consumption expenditure.

Notes: The blue line represents the counterfactual evolution of the
economy predicted by our model had monetary policy been set to
achieve the path for inflation depicted in the top panel. This counterfactual is conditional on the posterior mean of the parameters. Under
the “full-communication” scenario, the desired path for inflation is
achieved by the choice of the inflation target while all other shocks are
set at their Kalman smoother estimate. The shaded area represents the
75 percent symmetric probability interval for the unconditional forecast,
which is the same as in Chart 4. The black line represents the observed
realization of each series. The dashed line is the conditional forecast
under the “no-communication” scenario. PCE is personal
consumption expenditure.

The crucial difference between the results obtained under
the two scenarios stems from the key role that expectations play
in the DSGE model. Under the full-communication strategy,
inflation expectations are immediately affected by the
announcement of an inflation target. These expectations in
turn have a direct effect on actual inflation without requiring a
contraction in real activity to force businesses to contain their

price increases. Under the no-communication strategy,
inflation expectations remain at their historical level. As a
result, inflation can be controlled only by increasing interest
rates to contain GDP growth.
The way in which we model the full-communication
scenario is quite stark. In practice, expectations would be
unlikely to adjust instantaneously, even if the Federal Reserve

FRBNY Economic Policy Review / October 2010

were completely transparent about its inflation target. Nevertheless, the differences between the results of the two policy
strategies are striking. In the no-communication case, inflation
can be stabilized only through wild movements in the federal
funds rate. As a result, GDP growth is also extremely volatile:
it falls below zero in 2004:1, but then recovers to a quarterly
(annualized) growth rate of 10 percent and ends the period

Reserve around 2003 was to minimize the U.S. economy’s
likelihood of falling into such a trap.16 From this perspective,
our analysis might be interpreted as supportive of the policy
stance adopted by the central bank in 2003-04 as part of a
successful preemptive strike against a liquidity trap.

6. Conclusion
Our analysis might be interpreted as
supportive of the policy stance adopted
by the central bank in 2003-04 as part of
a successful preemptive strike against
a liquidity trap.

at zero. These movements in output are indeed extreme. They
lie well outside the 75 percent forecast probability interval
reported in the chart. In fact, the quantitative details of the
evolution of output and the interest rate under the counterfactual simulations should not be taken literally, since they
depend significantly on the details of the model and on the
assumption that the central bank insists on perfectly stabilizing
current inflation. However, the qualitative pattern of higher
volatility under the no-communication strategy is a robust
feature of models in which expectations matter.
Under the full-communication strategy, in contrast, the
desired path for inflation can be achieved with much less
pronounced fluctuations in real growth and an almost
unchanged policy relative to the actual path. Interest rates need
not rise and output need not fall significantly because a shift in
expectations brought about by clear communication of the
Federal Reserve’s inflation objective largely brings inflation
under control.
Note that the results of these counterfactual exercises should
be interpreted with caution. Their objective is not to prescribe
an alternative to the policy followed in 2004, but rather to
investigate how a different path for inflation might have been
achieved. In fact, according to Krugman (1998) and Eggertsson
and Woodford (2003), an increase in inflation expectations
might be the best monetary strategy to escape a liquidity trap.
Many have argued that the main objective of the Federal

Policy Analysis Using DSGE Models: An Introduction

This article provides an introduction to dynamic stochastic
general equilibrium models and presents an example of their
use as tools for monetary policy analysis. Given the mainly
educational nature of our presentation, we simplify by using
a small-scale model designed to account for the behavior of
three key macroeconomic variables: GDP growth, core PCE
inflation, and the federal funds rate. Despite its simplicity, our
model is rich enough to reproduce some of the salient features
of the series of interest. It also allows us to highlight the
components common to more articulated and realistic DSGE
specifications.
Our model offers insight into the causes of the abrupt pickup in inflation in the first half of 2004, from levels close to
1 percent at the beginning of 2003 to values steadily above
2 percent through the end of 2008. This exercise highlights the
central role of expectations in the transmission of shocks and
policy impulses in DSGE models. The main lesson that we
derive from the exercise is that the most effective approach
to controlling inflation is through the management of
expectations, rather than through actual movements of the
policy instrument. This lesson seems to be well understood by
the public, given the amount of attention and speculation that
usually surround the pronouncements of central bankers on
their likely future actions. DSGE models have the potential
to broaden this understanding by adding a quantitative
assessment of the link between current policy, expectations,
and economic outcomes—and thus to clarify the effect that
different systematic approaches to policy have on those
outcomes.
16

In its August 2003 statement, the FOMC observed that “on balance, the risk
of inflation becoming undesirably low is likely to be the predominant concern
for the foreseeable future.” Very low or negative levels of inflation are one of
the most likely triggers of a liquidity trap.

Technical Appendix

The table reports information on the prior distribution for
the parameters of the model. Further details on the parameters
and the structure of the model are available from the
corresponding authors.

Parameter

Distribution

Mean

Standard
Deviation








y


Calibrated
Gamma
Gamma
Beta
Beta
Beta
Normal
Normal
Normal
Normal
Normal
Beta
Beta
Beta
Beta
InvGamma
InvGamma
InvGamma
InvGamma
InvGamma

0.99
0.1
1.0
0.6
0.6
0.7
1.5
0.5
2.0
2.0
3.0
0.95
0.5
0.5
0.5
0.5
0.5
0.5
0.2
0.5

—
0.05
0.2
0.2
0.2
0.15
0.25
0.2
1.0
1.0
0.35
0.04
0.2
0.2
0.2
2.0
2.0
2.0
1.0
2.0


 


u


u
 
i

FRBNY Economic Policy Review / October 2010

References

Adolfson, M., M. K. Andersson, J. Lindé, M. Villani, and A. Vredin.
2007. “Modern Forecasting Models in Action: Improving
Macroeconomic Analyses at Central Banks.” International
Journal of Central Banking 3, no. 4 (December): 111-44.
Adolfson, M., S. Laséen, J. Lindé, and M. Villani. 2008. “Evaluating
an Estimated New Keynesian Small Open-Economy Model.”
Journal of Economic Dynamics and Control 32, no. 8
(August): 2690-2721.
An, S., and F. Schorfheide. 2007. “Bayesian Analysis of DSGE Models.”
Econometric Reviews 26, nos. 2-4 (April): 113-72.
Bernanke, B. S., M. Gertler, and S. Gilchrist. 1999. “The Financial
Accelerator in a Quantitative Business Cycle Framework.” In J. B.
Taylor and M. Woodford, eds., Handbook of Macroeconomics,
vol. 1, 1341-93. Elsevier.

Curdia, V., and M. Woodford. 2009. “Conventional and
Unconventional Monetary Policy.” Federal Reserve Bank
of New York Staff Reports, no. 404, November.
Dixit, A. K., and J. E. Stiglitz. 1977. “Monopolistic Competition and
Optimum Product Diversity.” American Economic Review 67,
no. 3 (June): 297-308.
Edge, R. M., M. T. Kiley, and J.-P. Laforte. 2008. “Natural Rate
Measures in an Estimated DSGE Model of the U.S. Economy.”
Journal of Economic Dynamics and Control 32, no. 8
(August): 2512-35.
Eggertsson, G. B., and M. Woodford. 2003. “The Zero Bound on
Interest Rates and Optimal Monetary Policy.” Brookings Papers
on Economic Activity, no. 1: 139-211.

Bils, M., and P. J. Klenow. 2004. “Some Evidence on the Importance
of Sticky Prices.” Journal of Political Economy 112, no. 5
(October): 974-85.

Eichenbaum, M. S., L. Hansen, and K. J. Singleton. 1988. “A Time Series
Analysis of Representative Agent Models of Consumption and
Leisure Choice under Uncertainty.” Quarterly Journal of
Economics 103, no. 1 (February): 51-78.

Calvo, G. A. 1983. “Staggered Prices in a Utility-Maximizing
Framework.” Journal of Monetary Economics 12, no. 3
(September): 383-98.

Erceg, C. J., L. Guerrieri, and C. Gust. 2006. “SIGMA: A New OpenEconomy Model for Policy Analysis.” International Journal
of Central Banking 2, no. 1 (March): 1-50.

Campbell, J. Y. 2003. “Consumption-Based Asset Pricing.” In G. M.
Constantinides, M. Harris, and R. M. Stulz, eds., Handbook of
the Economics of Finance, vol. 1, 803-87. North-Holland.

Galí, J., and M. Gertler. 1999. “Inflation Dynamics: A Structural
Econometric Analysis.” Journal of Monetary Economics 44,
no. 2 (October): 195-222.

Campbell, J. Y., and J. H. Cochrane. 2000. “Explaining the Poor
Performance of Consumption-Based Asset Pricing Models.”
Journal of Finance 55, no. 6 (December): 2863-78.

———. 2007. “Macroeconomic Modeling for Monetary Policy
Evaluation.” Journal of Economic Perspectives 21, no. 4 (fall):
25-45.

Christiano, L., R. Motto, and M. Rostagno. 2007. “Financial Factors
in Business Cycles.” Unpublished paper.

Gerali, A., S. Neri, L. Sessa, and F. Signoretti. 2008. “Credit and
Banking in a DSGE Model.” Unpublished paper, Bank of Italy,
June.

Clarida, R. G., J. Galí, and M. Gertler. 1999. “The Science of Monetary
Policy: A New Keynesian Perspective.” Journal of Economic
Literature 37, no. 4 (December): 1661-1707.
Curdia, V., A. Ferrero, and A. Tambalotti. 2009. “Which Output Gap?
Evaluating Interest Rate Rules in an Estimated DSGE Model.”
Unpublished paper, Federal Reserve Bank of New York.

Policy Analysis Using DSGE Models: An Introduction

Gertler, M., and P. Karadi. 2009. “A Model of Unconventional
Monetary Policy.” Unpublished paper.
Gertler, M., L. Sala, and A. Trigari. 2008. “An Estimated Monetary
DSGE Model with Unemployment and Staggered Nominal Wage
Bargaining.” Journal of Money, Credit, and Banking 40,
no. 8 (December): 1713-64.

References (Continued)

Gilchrist, S., A. Ortiz, and E. Zakrajsek. 2009. “Credit Risk and the
Macroeconomy: Evidence from an Estimated DSGE Model.”
Unpublished paper.
Hansen, L. P., and K. J. Singleton. 1982. “Generalized Instrumental
Variables Estimation of Nonlinear Rational Expectations Models.”
Econometrica 50, no. 5 (September): 1269-86.
———. 1983. “Stochastic Consumption, Risk Aversion, and the
Temporal Behavior of Asset Returns.” Journal of Political
Economy 91, no. 2 (April): 249-65.
Krugman, P. 1998. “It’s Baaack! Japan’s Slump and the Return of the
Liquidity Trap.” Brookings Papers on Economic Activity,
no. 2: 137-87.
Nakamura, E., and J. Steinsson. 2008. “Five Facts about Prices:
A Reevaluation of Menu Cost Models.” Quarterly Journal
of Economics 123, no. 4 (November): 1415-64.
Orphanides, A., and D. W. Wilcox. 2002. “The Opportunistic
Approach to Disinflation.” International Finance 5, no. 1
(spring): 47-71.

Smets, F., and R. Wouters. 2003. “An Estimated Dynamic Stochastic
General Equilibrium Model of the Euro Area.” Journal of the
European Economic Association 1, no. 5 (September):
1123-75.
———. 2007. “Shocks and Frictions in U.S. Business Cycles:
A Bayesian DSGE Approach.” American Economic Review 97,
no. 3 (June): 586-606.
Taylor, J. B. 1993. “Discretion versus Policy Rules in Practice.”
Carnegie-Rochester Conference Series on Public
Policy 39, December: 195-214.
———. 2008. “The Costs and Benefits of Deviating from the
Systematic Component of Monetary Policy.” Remarks delivered
at the Federal Reserve Bank of San Francisco Conference on
Monetary Policy and Asset Markets, San Francisco, California,
February 22.
Woodford, M. 2003. Interest and Prices: Foundations
of a Theory of Monetary Policy. Princeton, N.J.:
Princeton University Press.

Sbordone, A. M. 2002. “Prices and Unit Labor Costs: A New Test of
Price Stickiness.” Journal of Monetary Economics 49, no. 2
(March): 265-92.

The views expressed are those of the authors and do not necessarily reflect the position of the Federal Reserve Bank of New York
or the Federal Reserve System. The Federal Reserve Bank of New York provides no warranty, express or implied, as to the
accuracy, timeliness, completeness, merchantability, or fitness for any particular purpose of any information contained in
documents produced and provided by the Federal Reserve Bank of New York in any form or manner whatsoever.
FRBNY Economic Policy Review / October 2010

Kenneth D. Garbade, Frank M. Keane, Lorie Logan, Amanda Stokes,
and Jennifer Wolgemuth

The Introduction of
the TMPG Fails Charge
for U.S. Treasury Securities
• Prior to May 2009, market convention enabled
a seller of Treasury securities to postpone—
without any explicit penalty and at an
unchanged invoice price—its obligation
to deliver the securities.

• The September 2008 insolvency of Lehman
Brothers exposed a flaw in this convention,
when a decline in short-term interest rates
set the stage for an extraordinary volume
of settlement fails that threatened to erode
the perception of the market as being free
of credit risk.
• In response, the Treasury Market Practices
Group introduced a “dynamic fails charge”
for Treasury securities in May 2009.
• The fails charge incentivizes timely settlement
by providing that a buyer of Treasury securities
can claim monetary compensation from a seller
if the seller fails to deliver on a timely basis.
• The fails charge mitigated a key dysfunctionality
in the market and illustrates the value of public
and private sector cooperation in responding
to altered market conditions.

Kenneth D. Garbade is a senior vice president, Frank M. Keane an assistant
vice president, Lorie Logan a vice president, Amanda Stokes a trader/analyst,
and Jennifer Wolgemuth a counsel and assistant vice president at the Federal
Reserve Bank of New York.
Correspondence: kenneth.garbade@ny.frb.org

1. Introduction

ecurities transactions commonly involve a variety of
market conventions—widely accepted ways of doing
business that persist through time even though not mandated
by law or regulation. Commonplace examples include the
quotation of prices for Treasury bonds in increments of 32nds
(and fractions of a 32nd) of a percent of principal value (rather
than in decimal increments) and the quotation of Treasury bills
in terms of discount rates (rather than prices or yields).
In most cases, market conventions are useful or, at worst,
innocuous. In some cases, however, a new use for an old
instrument can render a convention in need of revision. One
particularly notorious example was the convention—observed
prior to 1982—of ignoring accrued interest on Treasury bonds
sold on repurchase agreements (also known as repos, or RPs).
The convention made sense as long as repos were used
primarily to borrow money from creditworthy lenders that
held the bonds simply to limit their exposure to credit risk. It
made less sense when highly leveraged securities dealers began
to use repos to borrow bonds to deliver on short sales. The 1982
Drysdale episode illuminated the risks involved in ignoring
accrued interest and prompted the Federal Reserve Bank of
New York to orchestrate a change in the convention.1

Garbade (2006).

The authors are grateful to Murray Pozmanter and Cartier Stennis for
assistance in researching this article and to two anonymous referees for helpful
suggestions. They are especially grateful to the members of the Treasury
Market Practices Group, identified in Box 4, for the time and effort they
devoted to making the fails charge a reality, as well as to all of the other
individuals, identified in the appendix, who provided invaluable feedback and
assistance during the implementation of the fails charge. The views expressed
are those of the authors and do not necessarily reflect the position of the
Federal Reserve Bank of New York or the Federal Reserve System.
FRBNY Economic Policy Review / October 2010

A market convention may also require revision following a
change in the economic environment. This article discusses a
recent example: the convention of postponing—without any
explicit penalty and at an unchanged invoice price—a seller’s
obligation to deliver Treasury securities if the seller fails to
deliver the securities on a scheduled settlement date. As
discussed in more detail below, as long as short-term interest
rates were above about 3 percent, the time value of money
usually sufficed to incentivize timely settlement of transactions
in Treasury securities. However, when short-term rates fell to

A market convention may . . . require
revision following a change in the
economic environment. This article
discusses a recent example: the
convention of postponing—without any
explicit penalty and at an unchanged
invoice price—a seller’s obligation to
deliver Treasury securities if the seller fails
to deliver the securities on a scheduled
settlement date.
near zero following the insolvency of Lehman Brothers
Holdings Inc. in September 2008, the time value of money no
longer provided adequate incentive and the Treasury market
experienced an extraordinary volume of settlement fails. Both
the breadth of the fails across a large number of securities and
the persistence of the fails were unprecedented and threatened
to erode the perception of the Treasury market as a market free
of credit risk.2 In response, the Treasury Market Practices
Group (TMPG)—a group of market professionals committed
to supporting the integrity and efficiency of the U.S. Treasury
market—worked over a period of six months to revise the
market convention for settlement fails, developing a “dynamic
fails charge” that, when short-term interest rates are below
3 percent, produces an economic incentive to settle trades
roughly equivalent to the incentive that exists when rates are at
3 percent. Thus, the TMPG fails charge preserves a significant
economic incentive for timely settlement even when interest
rates are close to zero.
2

See, for example, Wrightson, Federal Reserve Data, October 17, 2008 (“The
breakdown in the clearing mechanism for the Treasury market is beginning to
emerge as a top-tier policy concern. The safe-haven status of Treasury
securities is one of the few advantages the government market has left in a year
in which net Treasury borrowing needs . . . are likely to exceed $1 trillion by a
large margin. At some point, though, buyers will think twice about buying a
‘safe-haven’ asset for peace of mind if they have doubts about their
counterparty’s ability to deliver the security.”).

The Introduction of the TMPG Fails Charge

This article describes the introduction of the TMPG fails
charge. The introduction of the fails charge is important for
two reasons. First, it mitigated an important dysfunctionality
in a market of critical significance both to the Federal Reserve
in its execution of monetary policy and to the country as a
whole. Second, it exemplified the value of cooperation between
the public and private sectors in responding to altered market
conditions in a flexible, timely, and innovative fashion.
Our study is divided into three parts. The first part (Sections 2-5)
describes settlement processes and settlement fails in the
Treasury market, explains why sellers usually try to avoid fails,
describes industry and Federal Reserve efforts to mitigate
settlement fails prior to 2008, and briefly reviews three episodes
of chronic fails in the Treasury market. The second part
(Section 6) describes the tsunami of fails that followed
Lehman’s insolvency. The balance of the study (Sections 7-10)
explains the TMPG’s response. Section 11 concludes.

2. Settlements and Settlement Fails
in U.S. Treasury Securities
A transaction in Treasury securities is said to “settle” when the
seller delivers the securities to, and receives payment from, the
buyer. The two most important settlement processes are
bilateral settlement and multilateral net settlement. Before
describing those processes, we explain how market participants
establish and transfer ownership of Treasury securities.

2.1 Establishing and Transferring Ownership
of Treasury Securities
For more than three decades, investors have established
ownership of Treasury securities through Federal Reserve
book-entry securities accounts.3 Book-entry account holders
that own Treasury securities can house their securities directly
in their accounts and can transfer the securities to other bookentry accounts by issuing appropriate instructions to the
Federal Reserve.
Federal Reserve book-entry accounts are generally available
only to depository institutions and certain other organizations,
such as government-sponsored enterprises and foreign central
banks. All other market participants establish ownership of
Treasury securities through commercial book-entry accounts
at depository institutions that act as custodians for their
customers. Depository institutions that offer commercial
3

See “Book-Entry Securities Account Maintenance and Transfer Services,”
Federal Reserve Banks Operating Circular no. 7, August 19, 2005. Garbade
(2004) describes the origins of the Federal Reserve book-entry system.

book-entry accounts hold their customers’ securities in their
Federal Reserve book-entry accounts commingled with their
own securities.
A market participant with a commercial book-entry
account can transfer Treasury securities to another market
participant through their respective custodians. For example,
participant A can transfer a Treasury security to participant B
by instructing its custodian to debit its commercial book-entry
account and to transfer the security to B’s custodian for credit
to B’s commercial book-entry account. Upon receipt of
instructions, A’s custodian will debit A’s account and instruct
the Federal Reserve to 1) debit its Federal Reserve book-entry
account and 2) credit the Federal Reserve book-entry account
of B’s custodian. Following receipt of the security in its Federal
Reserve book-entry account, B’s custodian will complete the
transfer by crediting B’s commercial book-entry account.
(If A and B have a common custodian, the transfer can be
completed on the books of that common custodian without
involving the Federal Reserve.)

2.2 Bilateral Settlement
The simplest type of settlement occurs when a market
participant has sold Treasury securities for bilateral settlement
on a deliver-versus-payment basis. The sale may be a
conventional sale of securities but it may alternatively be

The simplest type of settlement occurs
when a market participant has sold
Treasury securities for bilateral settlement
on a deliver-versus-payment basis.
the starting leg, or the “off” leg, of a repurchase agreement.
(We describe repurchase agreements in more detail below.)
Suppose, for example, an investor sells ten Treasury bonds
at a price of $100 per bond for settlement on June 2. Following
negotiation of the terms of the sale, the seller will instruct its
custodian to send ten bonds to the buyer’s custodian on June 2
against payment of $1,000. The buyer will concurrently
instruct its custodian to receive, on June 2, ten bonds from the
seller’s custodian and to pay $1,000 upon receipt of the bonds.
On June 2, the seller’s custodian will instruct the Federal
Reserve to 1) debit its Federal Reserve book-entry account for
ten bonds, 2) credit the Federal Reserve book-entry account of
the buyer’s custodian for ten bonds and simultaneously debit
the account of the buyer’s custodian for the $1,000 due upon

Exhibit 1

Bilateral Settlement

Buyer

Seller
Delivery
instructions

Receive
instructions
Ten bonds

Seller’s
custodian

Buyer’s
custodian
$1,000

Note: This transfer of bonds and funds is effected through the Federal
Reserve if the seller and buyer have different custodians, and is effected
on the books of the common custodian if they have the same custodian.

delivery, and 3) credit the seller’s custodian’s account for the
$1,000. The resulting transfers of securities and funds are
shown in Exhibit 1.4
Following notification that ten bonds have come into its
Federal Reserve book-entry account and that $1,000 has been
withdrawn, the buyer’s custodian will verify that the bonds and
money are consistent with the buyer’s instructions. In most
cases, they are and the custodian will credit the buyer’s account
for the ten bonds and debit that account for $1,000. In some
cases, however, the buyer will have provided different
instructions—perhaps referencing a different security or a
different invoice price—or no instructions. In any of these
cases, the buyer’s custodian will reverse the settlement,
instructing the Federal Reserve to return the ten bonds and
recover the $1,000 payment. The buyer and seller and their
respective custodians will then have to communicate and come
to a common understanding of the terms of the underlying
transaction, following which the seller will reinitiate the
settlement process.

2.3 Multilateral Net Settlement
Bilateral settlement is a simple process that satisfies the purpose
of settlement: moving securities from sellers to buyers and
moving funds from buyers to sellers. Alternative settlement
structures, however, can sometimes be more efficient.
4

In the event the buyer and seller have a common custodian, settlement can be
completed on the books of the common custodian, with cash and securities
moving between the accounts of the respective customers, without involving
the Federal Reserve.

FRBNY Economic Policy Review / October 2010

Exhibit 2

Bilateral Settlement of Three Transactions

because it will receive more for the bonds sold to C than the
original contract price and C loses for the same reason. These
gains and losses are exactly offset with further agreements to
make small side payments of cash. In particular:

• A agrees to pay $8 to the CCP,
reflecting the $8 gain from marking the price of the eight
bonds bought from C down from $101 per bond to $100
per bond,

Instructions

• B agrees to pay $10 to the CCP,
reflecting the $10 gain from marking the price of the ten
bonds sold to C up from $99 to $100 per bond, and

A’s custodian

Eight bonds

$808

Ten bonds

• the CCP agrees to pay $18 to C,
in compensation for the $8 loss from marking the price
of the eight bonds sold to A down from $101 per bond,
and for the $10 loss from marking the price of the ten
bonds bought from B up from $99 per bond.

$1,000
$990
C’s custodian

B’s custodian
Ten bonds

Instructions

On the night before the settlement date, the CCP nets out
the deliver and receive obligations of A, B, and C and novates5
their respective contracts, becoming the buyer from every net
seller and the seller to every net buyer, all at the common
settlement price. After netting and novation:
• A is obligated to deliver two bonds to the CCP against
payment of $200,

Consider, for example, the case where:

• B has no deliver or receive obligations, and

• participant A sells ten bonds to participant B at a price
of $100 per bond for settlement on the following
business day,

• the CCP is obligated to deliver two bonds to C against
payment of $200.

• B sells ten of the same bonds to participant C at a price
of $99 per bond, also for settlement on the following
business day, and
• C sells eight of the same bonds to A at a price of $101 per
bond, again for settlement on the following business day.
As shown in Exhibit 2, bilateral settlement of the three
transactions requires the delivery of twenty-eight bonds against
payments of $2,798.
As an alternative, the participants might agree to settle
through a central counterparty (CCP). The CCP first marks all
of the deliver and receive obligations to a common price—say,
$100 per bond. After marking to the common price,
• A is obligated to deliver ten bonds to B against payment
of $1,000,
• B is obligated to deliver ten bonds to C against payment
of $1,000, and
• C is obligated to deliver eight bonds to A against
payment of $800.
Marking to a common price results in gains for some
participants and losses for others. In the example, B gains

The Introduction of the TMPG Fails Charge

On settlement day, the obligations of A to deliver two bonds
to the CCP and the CCP to deliver two bonds to C are settled
with bilateral deliver-versus-payment settlements. In addition,
A, B, and the CCP make the agreed-upon side payments of
cash. Exhibit 3 shows that multilateral net settlement requires
the delivery of four bonds and payments of $436—about
15 percent of the deliveries and payments shown in Exhibit 2.

2.4 Some Concrete Identities
The foregoing description of settlement processes referred to
abstract entities like “participant A” and an unnamed “central
counterparty.” Before we begin to discuss settlement fails, it
may be helpful to identify some of the key participants in the
Treasury market.
At the center of the market is a group of dealers that provide
liquidity to customers, quoting bid prices at which they are
willing to buy and offer prices at which they are prepared to
sell. A subset of dealers, called “primary dealers,” make markets
5

To novate is to substitute one legal obligation for another.

Box 1

Exhibit 3

Multilateral Net Settlement of Three Transactions

Instructions

Primary Dealers in Mid-2008a
Banc of America Securities LLC
Barclays Capital Inc.
Bear, Stearns & Co., Inc.b
BNP Paribas Securities Corp.
Cantor, Fitzgerald & Co.

A’s custodian
$8

Two bonds

$200
Instructions

Central
Instructions
Central
counterparty’s
counterparty
custodian

Two bonds

$10

Citigroup Global Markets, Inc.
Credit Suisse Securities (USA) LLC
Daiwa Securities America Inc.
Deutsche Bank Securities Inc.
Dresdner Kleinwort Securities LLC

B’s custodian

$200

Goldman, Sachs & Co.
Greenwich Capital Markets, Inc.
HSBC Securities (USA) Inc.
J. P. Morgan Securities Inc.
Lehman Brothers Inc.c

$18
C

Instructions

C’s custodian

Note: Security settlements are shown with solid blue and black lines. Side
payments, represented by dashed black lines, take place independently
of security settlements.

to the Federal Reserve Bank of New York when the Bank is
conducting open market operations on behalf of the Federal
Reserve System.6 Box 1 identifies the primary dealers as of mid2008.
Dealers sometimes trade directly with each other, but more
commonly through specialized interdealer brokers. A dealer
that sells securities to another dealer through an interdealer
broker agrees to deliver securities (against payment) to the
broker. The broker, in turn, agrees to deliver the same
securities (also against payment) to the ultimate buyer. This
arrangement allows the dealers to trade on a “blind,” or
undisclosed, basis.
All of the dealers, and all of the interdealer brokers,
maintain commercial book-entry accounts at one of two banks:
JPMorgan Chase Bank, N.A., and The Bank of New York
Mellon. These two “clearing” banks offer custodial services
refined over many years to meet the needs of brokers and
dealers that deliver and receive large volumes of securities on a
daily basis.
The Fixed Income Clearing Corporation (FICC), a
subsidiary of the Depository Trust & Clearing Corporation, is
the central counterparty in the Treasury market. All of the
6

See, generally, Federal Reserve Bank of New York, “Administration of
Relationships with Primary Dealers,” January 11, 2010, available at
http://www.newyorkfed.org/markets/pridealers_ policies.html.

Merrill Lynch Government Securities Inc.d
Mizuho Securities USA Inc.
Morgan Stanley & Co. Incorporated
UBS Securities LLC
a

Federal Reserve Bank of New York, “Primary Dealers List,” July 15, 2008,
available at http://www.newyorkfed.org/newsevents/news/markets/2008/
an080715.html.
b

Removed October 1, 2008, following its acquisition by J. P. Morgan
Securities Inc.

Removed September 22, 2008.

Removed February 11, 2009, following its acquisition by Bank of America
Corporation.

primary dealers and all of the interdealer brokers, as well as a
number of other market participants, are netting members of
FICC. FICC maintains commercial book-entry accounts at
both JPMorgan Chase and The Bank of New York Mellon and
is prepared to receive securities from, and deliver securities to,
any of its netting members in a timely and efficient fashion.
Beyond the dealers, the interdealer brokers, and FICC,
the Treasury market consists of a large number of other participants, including “real-money” investors such as mutual funds,
pension funds, and corporate treasurers, and “leveraged
accounts” such as hedge funds. Some of these participants trade
directly with dealers, others trade anonymously in electronic
markets. All use custodians that offer more or less complex
(and more or less costly) services tailored to their needs.

FRBNY Economic Policy Review / October 2010

2.5 Settlement Fails
A settlement fail occurs when the obligation of a seller to
deliver securities to a buyer remains outstanding following the
close of business on the scheduled settlement date of a
transaction. This can occur either because the seller’s custodian
failed to tender any securities to the buyer’s custodian, or
because the buyer’s custodian rejected whatever securities were
tendered by the seller’s custodian. In the event of a settlement
fail in Treasury securities, the market convention is to
postpone settlement to the following business day without any

A settlement fail occurs when the
obligation of a seller to deliver securities
to a buyer remains outstanding following
the close of business on the scheduled
settlement date of a transaction.

change in the funds due upon delivery and (prior to May 2009)
without any explicit penalty or charge.7 The process of failing
(to settle) and deferring settlement to the next business day can
take place repeatedly, day after day, until settlement occurs or
the trade is canceled.
Settlement fails can occur for any of several reasons. First, a
fail can result from miscommunication. A buyer and seller may
not have a common understanding of the terms of a trade, or
one or the other may have failed to communicate settlement
instructions to its custodian, or may have communicated
incorrect instructions, or one of the custodians may have
misunderstood the instructions that it received. In any of these
cases, the buyer’s custodian will reject whatever securities are
tendered by the seller’s custodian. After becoming aware of the
failed attempt to settle (or of the absence of any attempt to
settle), the buyer and seller and their respective custodians
communicate to resolve the problem. This usually results in
successful settlement within a day or two.
A fail may also stem from operational problems. One
well-known instance occurred on Thursday, November 21,
1985, when a computer outage at The Bank of New York
7

This convention was memorialized in Chapter 8, Section C, of the
Government Securities Manual of the Public Securities Association: “If
securities are not delivered on the agreed upon settlement date, there is a fail.
Regardless of the date the securities were actually delivered, the buyer of the
securities pays the seller the original settlement date figures.” The Public
Securities Association was the forerunner of the Bond Market Association,
which joined with the Securities Industry Association in 2006 to form the
Securities Industry and Financial Markets Association.

The Introduction of the TMPG Fails Charge

(a predecessor of The Bank of New York Mellon) prevented
that bank from effecting deliveries of Treasury securities. The
bank was unable to resolve the problem until the following day,
and had to finance overnight (at its own expense) the customer
securities that it was unable to deliver. It borrowed in excess of
$20 billion from the Federal Reserve Bank of New York and
incurred interest expenses of $5 million.8
A settlement fail can also occur because the seller does not
have the requisite securities in its commercial book-entry
account. This is the most common reason for failing when fails
are chronic, but it is usually avoided at other times by
borrowing securities and delivering the borrowed securities.

3. Repurchase Agreements and
Borrowing Securities to Avoid
or Cure Settlement Fails
A repurchase agreement is a sale of securities coupled with an
agreement to repurchase the same securities at a specified price
on a later date.9 Market participants use repos to borrow
money when they buy securities but do not have sufficient cash
on hand to pay for them, that is, to finance long positions, as
well as to borrow securities when they sell securities they do not
already own, that is, to finance short positions.
A repo is analogous to a loan, where the proceeds of the
initial sale correspond to the principal amount of the loan and
the excess of the repurchase price over the original sale price
corresponds to the interest paid on the loan. A market
participant might, for example, sell securities for $10 million
and simultaneously agree to repurchase the securities ten
days later for $10,008,333. This is analogous to borrowing
$10 million for ten days at an interest rate of 3 percent per
annum.10 Market participants commonly think of repos as
loans, rather than as purchases and sales, and quote repos in
terms of interest rates rather than in terms of sale and
repurchase prices.11
8

A Computer Snafu Snarls the Handling of Treasury Issues,” Wall Street
Journal, November 25, 1985, p. 58; Committee on Banking, Finance, and
Urban Affairs (1985); “Fed is Queried on Failure of Bank Computer System,”
Wall Street Journal, December 13, 1985, p. 49; “Fed Weighs a Penalty,”
New York Times, December 13, 1985, p. D2; Sender (1986).
9
Repurchase agreements are complex financial instruments whose contracting
conventions have evolved over the past four decades. See Garbade (2006) and
Fleming and Garbade (2003, 2004).
10
$8,333 = (repo term of 10 days / 360 days in a year) × 3 percent per annum
× $10 million, where the calculation uses the money market convention of a
360-day year.
11
The quotation convention does not change the nature of a repo—
a transaction in which one party sells securities subject to an agreement
to repurchase them at a later date.

Repos are most commonly arranged on an overnight basis
but can run for days or weeks. They can also be arranged on an
“open,” or continuing, basis (with a daily adjustment of the
interest rate) at the mutual consent of the parties. Industry
standard documentation for a repo provides that if the original
seller fails to repurchase the securities on the agreed-upon
repurchase date, the original buyer has the contractual right to,
among other things, sell the securities to a third party and use
the proceeds to satisfy the original seller’s repurchase
obligation. Conversely, if the original buyer does not deliver
the securities back to the original seller on the repurchase date,
the original seller has the contractual right to, among other
things, use the funds that it otherwise would have used to
repurchase the securities to “buy in,” or replace, the securities.

Exhibit 4

Lending Treasury Bond B (against Borrowing Money
at 2 Percent) on a Special Collateral Repurchase
Agreement and Relending the Money on a General
Collateral Repurchase Agreement at 3 Percent

Investor lending bond B (against
borrowing money at 2 percent)
and lending money at 3 percent
(against general collateral)

Bond B

General
collateral
Funds at
2 percent

3.1 Types of Repurchase Agreements

Participant borrowing
bond B against lending
money at 2 percent

Repos come in two flavors: general collateral repos (used to
borrow money) and special collateral repos (used to borrow
securities).

Starting leg of
special collateral
repurchase agreement

General collateral repos: A general collateral repo is a repo in
which the lender of funds is willing to accept any member of a
stated class of securities as collateral. Any of a variety of
securities is acceptable because the lender is concerned

Repos come in two flavors: general
collateral repos (used to borrow money)
and special collateral repos (used to
borrow securities).
primarily with earning interest on its money and having
possession of liquid assets that can be sold quickly in the event
of a default by the borrower.
Interest rates on overnight general collateral repos are
usually quite close to rates on overnight loans in the federal
funds market. This reflects the essential character of a general
collateral repo as a device for borrowing and lending money.
Repo rates for the most liquid and creditworthy collateral,
Treasury securities, are lowest. Repo rates for other classes of
collateral, such as fixed-income securities issued by a federal
agency or mortgage-backed securities issued by a government
sponsored enterprise, are somewhat higher.

Special collateral repos: A special collateral repo is a repo in
which the lender of funds designates a particular security as the
only acceptable collateral.12 Treasury market participants

Funds at
3 percent

Participant borrowing
money at 3 percent
against general collateral

Starting leg of
general collateral
repurchase agreement

Note: For simplicity, the separate roles of custodians are not shown explicitly.

commonly lend money on special collateral repos in order to
borrow specific securities that they need.
The interest rate on a special collateral repo is called a
“specials rate.” The owner of a Treasury security that other
market participants want to borrow may be incentivized to
lend the security if that owner is offered an opportunity to
borrow money at a specials rate less than the Treasury general
collateral repo rate. For example, if the rate on a special
collateral repo involving bond B is 2 percent and the general
collateral repo rate is 3 percent, an investor who owns bond B
can earn a 100 basis point spread by lending the bond and
borrowing money on a special collateral repo and then
relending the money on a general collateral repo (Exhibit 4).
The difference between the general collateral repo rate for
Treasury securities and the special collateral repo rate for a
particular Treasury security is a measure of the “specialness” of
the security and is commonly called the security’s “specialness
spread.” We show below that a security’s specialness spread is
exactly the opportunity cost of borrowing the security to avoid
or cure a settlement fail.
12

See Duffie (1996), Keane (1996), Jordan and Jordan (1997), Fisher (2002),
and Fleming and Garbade (2002).

FRBNY Economic Policy Review / October 2010

3.2 Incentives, prior to May 2009,
to Borrow Securities to Avoid
or Cure a Settlement Fail
Prior to May 2009, sellers of Treasury securities, including
short sellers, borrowed securities to avoid or cure settlement
fails primarily because they did not get paid until they delivered
the securities that they had sold. Prior to May 2009, market
participants usually quantified the cost to a seller of a
settlement fail in Treasury securities as the overnight Treasury
general collateral repo rate—the rate the seller could have

Prior to May 2009, a seller had an
incentive to avoid failing to deliver a
security (by borrowing the security on a
special collateral repo and delivering the
borrowed security) as long as the cost of
borrowing the security was less than the
cost of failing.
earned on a riskless overnight investment of the sale proceeds
that it did not receive. (It should be noted, however, that even
prior to May 2009, the cost of a settlement fail was not limited
to foregone interest earnings. Settlement fails also expose
market participants to the risk of counterparty insolvency and
can lead to increased capital charges for some participants.
These other costs are discussed in Box 2.)
A seller who does not have the securities needed to settle a sale
can avoid failing by borrowing (on a special collateral repo) the
securities that it needs. However, borrowing securities is not
costless because the borrower has to lend money (on the special
collateral repo) at a rate lower than the general collateral repo
rate that it could have earned on the money. The cost of
borrowing securities to avoid a fail in Treasury securities may be
quantified as the difference between the overnight Treasury
general collateral repo rate (the rate the borrower could have
earned on its money) and the overnight special collateral repo
rate on the borrowed securities (the rate the borrower actually
earns on its money)—that is, the securities’ specialness spread.
Prior to May 2009, a seller had an incentive to avoid failing
to deliver a security (by borrowing the security on a special
collateral repo and delivering the borrowed security) as long as
the cost of borrowing the security was less than the cost of
failing. This was certainly the case if the specialness spread for
the security was less than the general collateral repo rate or,
equivalently, if the special collateral repo rate for the security

The Introduction of the TMPG Fails Charge

Box 2

Other Costs of Settlement Fails
Settlement fails impose risks on both buyers and sellers and can
lead to increased capital charges for some market participants.
A buyer who fails to receive securities faces the risk that the
seller might become insolvent before the transaction is settled and
that, to replace the securities, it will have to pay more than the price
negotiated with the insolvent seller. Conversely, a seller who fails
to deliver securities faces the risk that the buyer might become
insolvent before the transaction is settled and that the seller will
then have to sell its securities to someone else at a price lower than
the price negotiated with the insolvent buyer. These risks may be
small for fails that are no more than a day or two old (because
securities prices usually do not change much from day to day), but
they can become significant when fails persist for weeks or months.
In view of the greater risks faced by both buyers and sellers
during the life of a long-outstanding, or “aged,” fail, some market
participants are required to absorb incremental capital charges
when a settlement fail has been outstanding for more than a few
weeks.a Such capital charges can drain capital from more
rewarding activities and limit balance sheet capacity, thereby
imposing opportunity costs on market participants—and if
sufficiently widespread and chronic can threaten overall market
functioning.b
a

See, for example, the net capital rule of the Securities and Exchange
Commission, Code of Federal Regulations, Chapter 17, Section 15c3-1.

See, for example, “Minutes of the Meeting of the Treasury Borrowing
Advisory Committee of the Bond Market Association, November 4, 2003,”
November 5, 2003, available at http://www.ustreas.gov/press/releases/
js933.htm (“While the situation is much improved since this past summer,
members commented that fails were still at an elevated level which does
hurt general market liquidity because dealers are forced to reduce their
market making activities as the fails take up space on their balance sheets.”).
One Treasury official suggested that opportunity costs resulting from
higher capital charges might not be all bad. See “Remarks by Jeff Huther,
Director of the Office of Debt Management, to the Bond Market
Association’s Annual Meeting,” April 22, 2004, available at http://
www.treas.gov/press/releases/js1455.htm (noting that “capital charges
resulting from chronic—widespread and persistent—fails soak up dealer
capital that might otherwise be used to support profit-making activities,
thereby focusing management attention on the underlying fails problem
and incentivizing managers to remedy the situation.”).

was greater than zero.13 As long as a seller could earn more than
a de minimis amount of interest on a special collateral repo, it
made economic sense to lend the money, earn the interest,
and avoid the fail.

Using economic terminology, let Rgc denote the general collateral
repo rate and Rsp denote the special collateral repo rate for the security.
The specialness spread Rgc – Rsp will be less then Rgc if and only if Rsp
is greater than zero.

3.3 Equilibrium in the Market
for Special Collateral Repos
The market for borrowing and lending a particular Treasury
security comes into equilibrium as a result of fluctuations in
the special collateral repo rate for the security relative to the
Treasury general collateral repo rate. If the demand to borrow
the security is modest relative to the supply available for
lending, a market participant seeking to borrow the security
will usually be able to lend its money at a specials rate no lower
than about 15 to 25 basis points below the general collateral
repo rate. However, if the demand to borrow the security
expands, some borrowers (in order to avoid failing) will be
willing to accept less interest on the money they lend.
Downward pressure on the specials rate for the security relative
to the general collateral repo rate makes lending the security
more remunerative, thereby attracting additional lenders. It
may also ration some borrowers out of the market, particularly
short sellers who decide to liquidate their short positions rather
than continue to finance those positions on special collateral
repos earning lower rates of interest. The collateral market will
return to equilibrium, that is, to a state where the quantity of
the security sought to be borrowed at the prevailing specials
rate equals the quantity of the security available for lending at
that rate, when the lower specials rate has attracted enough new
lenders and/or rationed enough borrowers out of the market.

4. Federal Reserve and Industry
Efforts to Mitigate Settlement Fails
Treasury market participants have an interest in mitigating
settlement fails in order to limit their net interest expenses as
well as their exposure to the risk of counterparty insolvency—
a risk explained in Box 2. The Federal Reserve has a separate
interest in mitigating settlement fails to maintain the liquidity
and efficiency of the market in which it conducts open market
operations. (A high volume of fails can lead market participants to reduce, or even withdraw from, their normal activities.
Such activities include dealers making markets for customers,
investors lending securities to dealers to facilitate settlement of
dealer sales, and arbitrageurs seeking to exploit, and thereby
eliminate, price relationships that present abnormal profit
opportunities.)
Since 1969, the Federal Reserve has sought to mitigate
settlement fails by lending Treasury securities to primary
dealers to facilitate settlement of dealer sales.14 (However, the
Federal Reserve lends against collateral, rather than cash, to
insulate the supply of reserves available to the banking system

from securities lending operations.) Pursuant to the terms and
conditions of the lending program in effect in mid-2008,15 each
business day at noon the Federal Reserve Bank of New York
offered to lend on an overnight basis up to 90 percent of the
amount of each Treasury security beneficially owned in the
Federal Reserve’s System Open Market Account (SOMA),
subject to an upper limit of the amount of an issue actually in

Treasury market participants have an
interest in mitigating settlement fails in
order to limit their net interest expenses
as well as their exposure to the risk of
counterparty insolvency.
the account.16 Primary dealers bid for a security by specifying
the quantity desired (in increments of $1 million) and the
overnight loan fee they were willing to pay, expressed as a rate
per annum, subject to a minimum fee of 50 basis points.17
Bidding closed at 12:15 p.m. and loans were awarded to the
highest bidders at their bid rates18 until all of the securities
available for lending were allocated or all of the bidders
had been satisfied.19 During the first six months of 2008,
the Federal Reserve Bank of New York lent an average of
$12.2 billion of securities per day, distributed over an average
of twenty-three different issues, at an average loan fee of
61 basis points per annum.
14

In authorizing the loan of Treasury securities from the System Open Market
Account in 1969, the Federal Open Market Committee stated that the action
“was taken after the Manager [of the System Open Market Account] had
advised that the problem of delivery failures in the Government securities
market had worsened significantly over the past year, partly because private
facilities for lending such securities had become inadequate; that delivery
failures were markedly impairing the performance of the market; and that the
functioning of the market would be improved if securities held in the System
Open Market Account could be lent, for the express purpose of avoiding delivery
failures, to Government securities dealers doing business with the Federal
Reserve Bank of New York . . . .” (Federal Reserve Bulletin, January 1970, p. 32).
15
“SOMA Securities Lending Program Terms and Conditions (Revised),”
Federal Reserve Bank of New York, August 22, 2008, available at http://
www.newyorkfed.org/newsevetns/news/ markets/2008/an082208.html.
Details of the history of the Fed’s lending program appear in Fleming and
Garbade (2007).
16
In order to avoid failing itself, the Fed does not agree to lend securities unless
it has actual possession of the securities at the time of an auction. Thus, it will
not agree to lend securities that it lent the preceding business day and that have
not yet been returned.
17
A loan fee for a security is approximately equivalent to the security’s
specialness spread. See Fleming and Garbade (2007). The minimum fee avoids
crowding out private lenders when security loan markets are functioning
normally, but it has been reduced to nearly zero when those markets are not
functioning well.
18
The auction for each security is a discriminating, or multiple-price, auction.

FRBNY Economic Policy Review / October 2010

Treasury market participants have also acted cooperatively
to mitigate settlement fails and otherwise reduce the cost of
settling transactions. Between 1986 and 1988, dealers in
Treasury securities organized the Government Securities
Clearing Corporation (GSCC) to serve as a central counterparty in interdealer transactions in Treasury and related
securities. As explained earlier, multilateral net settlement
through a central counterparty economizes on the quantity of
securities that have to be delivered to settle a given volume of
transactions. GSCC also implemented a trade confirmation
protocol that essentially eliminated interdealer fails due to
miscommunication between dealers, as well as a procedure for
marking failed trades to current market prices that materially
reduced the consequences of counterparty insolvency. The
GSCC extended its net settlement system to include Treasury
auction takedowns in 1994 and repurchase agreements in
1995.20 In 2002, GSCC became a wholly owned subsidiary of
Depository Trust & Clearing Corporation and was renamed
the Fixed Income Clearing Corporation.

5. Chronic Settlement Fails
Demand to borrow a security (relative to the supply available for
lending) can sometimes be large enough to drive the specials rate
for the security down to near zero. Prior to May 2009, sellers
would then become largely indifferent between a) failing and
b) borrowing the security to avoid failing. In this extreme case,
any unsatisfied demand to borrow would spill over into fails.
Fails could expand further if security lenders, observing a
growing incidence of settlement fails, declined to continue
lending out of concern that their securities may not be returned
on a timely basis.21 Fails could expand still further if, as explained
in Box 3, market participants concluded that they could acquire
a cheap option on a future increase in a specials rate by
contracting to sell a security in a special collateral repo and then
strategically failing to deliver the security. More generally,
settlement fails could become chronic when the specials rate for
a security was driven down to near zero.
Three episodes of chronic fails have been described in the
literature: in May and June of 1986,22 following the terrorist
attack on the World Trade Center on September 11, 2001,23
and during the summer of 2003.24

Box 3

Strategic Fails
When the specials rate for a security is close to zero in a market
without a fails charge convention, a market participant with no
position in the security may sometimes agree to lend the security
on a term repurchase agreement and then fail, intentionally, on
the starting leg of the repo.
Suppose, for example, the three-week specials rate for a fiveyear note is 10 basis points and that XYZ Co. believes the specials
rate will be 50 basis points in one week. If XYZ contracts (in the
specials market) to borrow $50 million for three weeks against
lending the note, it will owe interest of $2,917 at the end of three
weeks.a It will owe this amount even if it fails to deliver the note any
time during the three-week interval.
XYZ Co. has effectively purchased (for $2,917, payable in three
weeks) an option on an exchange of $50 million for the five-year
note at any time during the next three weeks for the balance of the
three-week interval.
XYZ could choose to let its option expire unexercised and simply
pay the $2,917 premium at the end of three weeks. However, if XYZ
Co.’s expectations prove correct, it can exercise the option after one
week by borrowing the five-year note for two weeks against lending
$50 million at 50 basis points (earning interest of $9,722b) and
delivering the note in (delayed) settlement of its earlier negotiated
three-week repurchase agreement. The $50 million received from
delivering the note funds the loan that allows XYZ Co. to borrow the
note, and XYZ Co. has net interest earnings of $6,805 ($6,805 =
$9,722 interest income, less $2,917 interest expense).
More generally, a very low specials rate presents an opportunity
to speculate on an increase in the rate—by lending on, and then
failing on, a special collateral repurchase agreement—with limited
downside exposure.c In the limit, a repo rate of zero may be viewed
by some participants as a “risk-free” opportunity to intentionally
fail and either profit or break even. Such practices can lead to an
increase in aggregate settlement fails and the associated indirect
costs discussed in Box 2.
a

$2,917 = (21/360)  0.10 percent per annum  $50 million.

$9,722 = (14/360)  0.50 percent per annum  $50 million.

Strategic fails are noted in Fleming and Garbade (2002, p. 47), Fleming
and Garbade (2004, pp. 3-4), and in “Remarks by Jeff Huther, Director of
the Office of Debt Management to the Bond Market Association’s Annual
Meeting,” April 22, 2004, available at http://www.treas.gov/press/releases/
js1455.htm.

Awards were subject to three limitations: 1) no dealer could have
outstanding borrowings of more than 25 percent of the amount of an issue
beneficially owned in the System Open Market Account, 2) no dealer could
have outstanding borrowings of more than $750 million of any single issue, and
3) no dealer could have outstanding borrowings of more than $3 billion of
securities in total.
20
See Garbade and Ingber (2005) and Ingber (2006).
21
This can quickly lead to a self-reinforcing and destabilizing cycle, with
lenders withdrawing collateral out of a concern that borrowers may fail to
return the securities, thereby increasing the incidence of settlement fails
and triggering further collateral withdrawals.
54

The Introduction of the TMPG Fails Charge

Cornell and Shapiro (1989). Cornell and Shapiro do not discuss fails directly,
but do document a near-zero specials rate for the 9 1/4 percent Treasury bond
of February 2016 and discuss the reasons for that low rate. The existence of
widespread settlement fails in the issue was common knowledge among market
participants at the time.
23
Fleming and Garbade (2002).
24
Fleming and Garbade (2004). See also Fleming and Garbade (2005).

5.1 The 1986 Episode
An “on-the-run” Treasury security is the most recently
auctioned security in a given series, such as the most recently
auctioned six-month bill or ten-year note. In late April 1986,
dealers began to sell the on-the-run thirty-year Treasury bond
(the 9 1/4 percent bond of February 2016) short in anticipation
of bidding for the new thirty-year bond that would be
announced on April 30 and auctioned on May 8 for settlement
on May 15.25 Such short selling in advance of an auction
announcement was normal and customary dealer behavior.26
However, dealers soon found themselves unable to borrow
enough 9 1/4 percent bonds to finance their short positions, in
part because a significant quantity of the bonds was owned by
investors who declined to lend.27 Strong dealer demand and
limited supply combined to drive the special collateral repo
rate for the bonds down to about 5 basis points, and dealers
began to fail on their settlement obligations. Failing, however,
was expensive because the Treasury general collateral repo rate
was about 6.75 percent, so dealers with short positions had an
economic incentive to cure their fails another way: by buying
(rather than borrowing) the 9 1/4 percent Treasury bonds and
delivering the purchased bonds—thereby closing out their
short positions. They bid up the price of the 9 1/4 percent
bonds, relative to the prices of other Treasury bonds with
similar maturities, until the higher price induced holders to sell
the 9 1/4 percent bonds and replace them with higher yielding
substitutes, thereby allowing dealers to cover their short
positions.28

Department of the Treasury, Securities and Exchange Commission, and
Board of Governors of the Federal Reserve System (1992, p. B-1, footnote 1)
(“participants sold the outstanding 9 1/4 percent bond . . . to prepare for the
roll into the WI [“when-issued”] thirty-year bond.”). Following the announcement of the forthcoming auction, dealers planned to buy the on-the-run thirtyyear bond (thereby covering their previous short sales of that bond) against
selling the WI thirty-year bond short. The transactions would leave them with
short positions in the WI bond that they could cover in the auction.
26
Department of the Treasury, Securities and Exchange Commission, and
Board of Governors of the Federal Reserve System (1992, p. 10, footnote 11)
(“dealers . . . sold [the 9 1/4 percent bond] short as part of a trading strategy
that had worked in the past as they prepared to bid for a new thirty-year
bond.”). The sequence of shorting the on-the-run thirty-year, rolling the short
into the WI thirty-year, and then bidding to buy the WI thirty-year in the
auction was part of the process whereby dealers distributed new bonds to
market participants.
27
Department of the Treasury, Securities and Exchange Commission, and
Board of Governors of the Federal Reserve System (1992, p. 10, footnote 11;
p. B-1, footnote 1) (“some institutional investors did not make the
[9 1/4 percent bonds] available to the repo market” and “securities needed
to [finance] short positions were not readily available to the repo market.”).
28
Cornell and Shapiro (1989, pp. 303-4) suggest that the 9 1/4 percent bond
of February 2016 was overvalued by as much as 7 percent of principal value
compared with one close substitute (the 9 7/8 percent bond of November
2015).

5.2 The 2001 Episode
The 2001 fails episode was attributable, in the first instance, to
the terrorist attack on the World Trade Center on September 11,
2001. The attack destroyed the offices of several interdealer
brokers and impaired telecommunication services throughout
lower Manhattan. GSCC recorded $266 billion in interdealer
settlement fails on September 11 and $440 billion in interdealer
fails on September 12.29 Sellers tried to borrow the securities

Settlement fails began to shrink to more
normal levels after the Treasury reopened
the on-the-run ten-year note in an
extraordinary unscheduled auction
offering on Thursday, October 4, and after
officials indicated that they might reopen
the on-the-run five-year note as well.
needed to cure their fails but holders realized that, in view of
the severe operational problems, their securities might not be
returned on a timely basis and they consequently declined to
lend.30 The contraction in the supply of collateral pushed
specials rates to near zero and settlement fails remained
elevated. Daily average fails in Treasury securities reported by
primary dealers to the Federal Reserve31 reached $200 billion
per day during the week of September 13-19 and continued
high through early October. Settlement fails were particularly
high for the on-the-run five-year note (the 4 5/8 percent note
29

Fleming and Garbade (2002, p. 46).
See “After Attack, Settlement Woes Still Clogging Repo Market,” Dow Jones
Newswires, September 26, 2001, 9:05 (noting “a general reluctance among large
portfolios to lend their securities” and “in a chain reaction, the fear of failing
trades is ‘causing portfolio managers, securities lending desks and foreign
central banks to hold even tighter on to their collateral,’ which is exacerbating
the situation . . . .”); “Treasury Market is Faced with Incomplete Trades,” Wall
Street Journal, October 3, 2001, p. C10 (“Because of the rate of fails . . . dealers
are reluctant to use their securities as collateral. They are worried that they
might not have securities delivered to them . . . . ”); “U.S. Sells $6 Billion in 10Year Notes to Help Overcome Shortage,” Bloomberg News, October 4, 2001,
13:16 (quoting Peter Fisher, Under Secretary of the Treasury for Domestic
Finance, as saying that “the cause of the fails is [in part] the result of . . .
reluctance by institutional investors to lend into a market that is suffering from
extraordinarily high fails levels.”); and “U.S. Acts on Shortage of Treasuries,”
New York Times, October 5, 2001, p. C1 (“With the prospect that securities
might not be returned, both dealers and large investors have become unwilling
to lend them in the repo market.”).
31
Fleming and Garbade (2005) describe the settlement fails data reported by
primary dealers to the Federal Reserve. Unless otherwise noted, this article
measures settlement fails as the daily average over weekly intervals of the
average of cumulative primary dealer fails to receive Treasury securities during
a week and primary dealer fails to deliver Treasury securities over the same
week. The Federal Reserve does not publish data on settlement fails on a dayby-day basis.
30

FRBNY Economic Policy Review / October 2010

of May 2006) and the on-the-run ten-year note (the 5 percent
note of August 2011).
Settlement fails began to shrink to more normal levels after
the Treasury reopened the on-the-run ten-year note in an
extraordinary unscheduled auction offering on Thursday,
October 4, and after officials indicated that they might reopen
the on-the-run five-year note as well.32 Peter Fisher, the Under
Secretary of the Treasury for Domestic Finance, stated that the
Treasury reopened the ten-year note “to reduce the risk that . . .
settlement problems turn into a much bigger problem for the
Treasury market . . . .”33 Fisher went on to observe that “we
have something that is self-compounding. There is some point
at which your fails pile up, and that is the point at which you
damage the price-discovery process and the smooth operating
of the Treasury market.”
The actions of Treasury officials convinced market
participants that the Treasury would take unprecedented steps
to facilitate settlements and maintain market liquidity. Holders
of the on-the-run five- and ten-year notes began to make the
notes available, and the level of fails subsided.34

$25 billion per day during the week ending June 18 to
$103 billion per day during the week ending July 2, and topped
out at $232 billion per day during the week ending August 20.
Settlement fails persisted for months36 and were not fully
resolved until the end of the year, following an offering of a new
series of ten-year notes in November.37

5.4 Proposals to Mitigate Chronic
Settlement Fails
The 2003 episode had a strong impact on the thinking of
market participants. Unlike the 1986 episode, which was shortlived and quickly forgotten, and unlike the 2001 episode, which
clearly stemmed from unusual circumstances, the 2003 episode
was lengthy, large-scale, and stemmed from a marketplace

The 2003 [fails] episode raised the
question of whether something should
be done, by government officials or by
private sector market participants, to
mitigate chronic fails.

5.3 The 2003 Episode
The 2003 fails episode was attributable, in the first instance, to
a heavy volume of short sales of the on-the-run ten-year note
(the 3 5/8 percent note of May 2013) in late June 2003 by
market participants seeking to hedge their interest rate risk on
long positions in other fixed-income securities.35 The short
sales created an unusually large demand to borrow the note
that drove the specials rate for the note down to zero, after
which the residual, unsatisfied demand spilled over into fails.
The fails became chronic when investors began to withdraw
from lending the note. Daily average fails in Treasury securities
reported by primary dealers to the Federal Reserve went from
32

“U.S. Sells $6 Billion in 10-Year Notes to Help Overcome Shortage,”
Bloomberg News, October 4, 2001, 13:16, and “U.S. Acts on Shortage of
Treasuries,” New York Times, October 5, 2001, p. C1 (both quoting the Under
Secretary of the Treasury for Domestic Finance as saying that the Treasury
might reopen the five-year note in the next week).
33
“U.S. Acts on Shortage of Treasuries,” New York Times, October 5, 2001,
p. C1.
34
See, for example, “Remedial Reopenings and the Treasury Supply Outlook,”
Money Market Observer, October 8, 2001 (“Dealers reported a dramatic
reduction in the volume of fails on [October 5] after the settlement of the
additional $6 billion of ten-year notes . . . .”), and “Another Emergency
Treasurys Sale Looks Unlikely as Shortages that Hamstrung ‘Repo Market’
Ease,” Wall Street Journal, October 11, 2001, p. C17.
35
“Supply Dries Up Following Fall in Prices,” Financial Times, August 23,
2003, p. 27 (reporting that “Demand for Treasuries from some quarters has
also risen as prices have fallen because many institutions want to borrow the
securities and ‘short’ them in the expectation that prices will continue to drop.
Traders say hedged positions for the [on-the-run ten-year note] now exceed
the amount of Treasury securities available.”).

The Introduction of the TMPG Fails Charge

activity—hedging—that was a very ordinary occurrence. The
2003 episode raised the question of whether something should
be done, by government officials or by private sector market
participants, to mitigate chronic fails.
The key difference between the 1986 and 2003 episodes
was the level of the Treasury general collateral repo rate.
In May 1986, the overnight general collateral repo rate was
about 6.75 percent. That made it costly to continue to fail
even after the special collateral repo rate on the 9 1/4 percent
bonds of February 2016 had been driven down to near zero
and the economic incentive to avoid failing by borrowing
the bonds had been eliminated. The high cost of failing
incentivized short sellers to cover their short positions with
outright purchases, and they bid up the price of the 9 1/4 percent bonds to a level that gave holders an economic incentive
to swap out of the issue and into higher yielding substitutes.
In mid-2003, however, the overnight Treasury general
collateral repo rate was about 1 percent, so the cost of failing
was modest. Short sellers had little incentive to cover their
36

See, for example, “Treasury Issue Remains a Headache,” Wall Street Journal,
November 17, 2003, p. C13.
37
“California Standoff Dims Prospects,” Wall Street Journal, December 9,
2003, p. C17 (reporting “progress for the . . . May ten-year note, which traders
said appeared to be emerging from six months of gridlock, thanks to supply
that entered the market last week. The note was trading in positive territory in
the repurchase-agreement market. For months, it had been stuck at 0% . . . .”).

short positions with outright purchases after demand to
borrow the May 2013 ten-year note had driven the specials rate
on the note down to near zero and eliminated the incentive to
borrow the note to avoid failing.
Market participants and government officials learned from
the 2003 episode that settlement fails were liable to become
chronic quickly when short-term interest rates are low,38 and
they began to contemplate institutional innovations to avoid,
or at least mitigate, chronic settlement fails. Most discussions
centered around three possibilities:

unpredictable reopening program would raise borrowing costs
over the long run.42

A Treasury lending facility: Like reopenings, a Treasury
lending facility would involve additional issuance from the
Treasury. Unlike reopenings, a Treasury lending facility would
increase supply on only a temporary basis. Such a facility was
put forth as a “straw man” in a Treasury white paper published

Market participants and government
officials learned from the 2003 episode
that settlement fails were liable to become
chronic quickly when short-term interest
rates are low.

• a regular program to reopen an issue when settlement
fails in the issue become chronic,
• a securities lending facility run by the Treasury
Department, and
• a fee to be paid by failing sellers to their counterparties
to incentivize the sellers to resolve their fails.

Reopenings: Reopening an issue to alleviate chronic fails was
exactly what the Treasury did when it reopened the on-the-run
ten-year note on October 4, 2001.39 However, Treasury
officials were reluctant to institutionalize reopenings as a
device to mitigate chronic fails. Three months after the 2001
reopening, Under Secretary Fisher told market participants
that while “it would be imprudent of me to say that the
Treasury will never again hold such an auction . . . you should
not count on it, you should not expect it . . . .”40 The problem
was that reopenings in response to chronic fails ran counter
to “regular and predictable” issuance, a cornerstone of
Treasury debt management since the 1970s.41 Treasury officials
were concerned that the uncertainties engendered by an
38

See, for example, “Remarks by Jeff Huther, Director of the Office of
Debt Management, to the Bond Market Association’s Annual Meeting,”
Department of the Treasury, April 22, 2004, available at http://www.treas.gov/
press/releases/js1455.htm (“The heart of the pricing problem last year was,
unquestionably, the low federal funds rate and the consequently low ceiling on
the cost of financing a short position.”); Department of the Treasury (2006,
p. 26, p. 174, footnote 2) (“The potential for chronic fails episodes thus
increases in a very low interest rate environment such as that prevailing during
the summer of 2003.”); and “Statement of Under Secretary for Domestic
Finance Randal K. Quarles to Bond Market Association Annual Meeting,”
Department of the Treasury, May 19, 2006, available at http://www.treas.gov/
press/releases/js4274.htm (“When the central bank wishes to establish very low
short-term rates, the maximum degree of specialness will be quite small.
During these periods, we might expect to see greater incidence of fails episodes
because the cost of failing is low.”).
39
That reopening was not the first time the Treasury increased the supply of a
security in response to unusual market conditions. On November 3, 1992, the
Treasury announced that it would reopen the 6 3/8 percent note of August
2002, at that time the on-the-run ten-year note, “in order to alleviate an acute,
protracted shortage” of the note. See “Treasury November Quarterly
Financing,” Office of Financing, Department of the Treasury, November 3,
1992.
40
“Remarks by Peter R. Fisher, Under Secretary of the Treasury for Domestic
Finance, Before the Bond Market Association Legal and Compliance
Conference,” January 8, 2002, available at http://www.treas.gov/press/releases/
po.906.htm.

in May 2006.43 The white paper was written to stimulate public
discussion of mechanisms to make available “an additional,
temporary supply of Treasury securities on rare occasions
when market shortages threaten to impair the functioning of
the market for Treasury securities and broader financial
markets . . .”
However, Treasury officials questioned whether the
Secretary of the Treasury has statutory authority to issue
securities on a temporary basis to alleviate chronic settlement
fails. Federal law provides that “the Secretary of the Treasury
may borrow on the credit of the United States Government
amounts necessary for expenditures authorized by law and may
issue bonds of the Government for the amounts borrowed.”44
Similar provisions authorize the issuance of notes and bills.45
The 2006 Treasury white paper suggested that “the Treasury
would likely need to pursue new authority to issue securities
for the purpose of securities lending. . . .”46
41

Garbade (2007) describes the emergence of “regular and predictable” as a
Treasury debt management strategy.
42
See “Remarks of Undersecretary of the Treasury Peter R. Fisher to the
Futures Industry Association, Boca Raton, Florida,” March 14, 2002, available
at http://www.treas.gov/press/releases/po1098.htm.
43
Department of the Treasury (2006). The public responses to the white paper
are available at http://www.ustreas.gov/offices/domestic-finance/debtmanagement/slf-comments.pdf. See also Garbade and Kambhu (2005).
A Treasury lending facility was also recommended by the Treasury Borrowing
Advisory Committee following the chronic fails of late September and early
October 2001; see “Report to the Secretary of the Treasury from the Treasury
Advisory Committee of the Bond Market Association,” October 30, 2001,
available at http://www.ustreas.gov/offices/domestic-finance/debtmanagement/adv-com/reports/rpt-2001-q4.pdf (“members overwhelmingly
felt that Treasury should expand their ability to enhance liquidity in the
Treasury market. To accomplish this, they could set up a repo facility to help
alleviate protracted shortages, in particular, large and persistent fails . . . .”).
44
31 U.S.C. § 3102 (2010).
45
See 31 U.S.C. § 3103-3104 (2010).
46
Department of the Treasury (2006, pp. 26, 178).

FRBNY Economic Policy Review / October 2010

A fails charge: In 2002, two economists at the Federal Reserve
Bank of New York suggested that “chronic fails can also be
alleviated by increasing the cost of failing with a penalty fee.”47
The economists noted that such a fee would give sellers an
economic incentive to borrow securities to avoid failing even
when the special collateral repo rate for the securities was close to
zero. They further noted that a fails charge might lead market
participants to borrow securities against lending money at
negative specials rates (in order to avoid the fails charge) and that
such negative specials rates could attract additional securities
lenders (because they would receive, rather than pay, interest on
the money they borrowed against lending securities).
The economists suggested that a fails charge might be set at
some threshold rate minus the general collateral repo rate, with
a minimum of zero.48 The fails charge would be above zero
only if the general collateral repo rate was below the threshold
rate and would not be higher than what was necessary to bring
the total cost of failing to the threshold rate. (For example, if
the threshold rate is 5 percent and the general collateral repo
rate is 3 percent, the fails charge would be 2 percent and the
total cost of failing would be 5 percent.) They further suggested
that the fails charge could be instituted through a “goodpractice” recommendation of the Bond Market Association.49
The economists noted that a fails charge could be implemented
implicitly by reducing the invoice price on a transaction each
day the seller fails—a material departure from the existing
convention of deferring settlement at an unchanged invoice
price—but observed that “the operational burden of changing
an invoice price following a delay in settlement would
undoubtedly be substantial.”50

5.5 Inaction prior to the Insolvency
of Lehman
Following the 2003 episode of chronic settlement fails, both
government officials and private sector market participants
understood that chronic fails are prone to blossom in an
environment of low interest rates. Several parties had identified
ways to address the problem, but each of the suggestions had a
material deficiency. Treasury officials asked private sector
participants to address the problem, but nothing substantive
came of their requests.51 No significant progress was made with
47

Fleming and Garbade (2002, p. 52).
Fleming and Garbade (2002, p. 53).
49
The Bond Market Association joined with the Securities Industry
Association in 2006 to form the Securities Industry and Financial Markets
Association.
50
Fleming and Garbade (2002, p. 52).
48

The Introduction of the TMPG Fails Charge

respect to addressing the problem of chronic fails before the
insolvency of Lehman in the fall of 2008.

6. Chronic Settlement Fails in the
Wake of the Insolvency of Lehman
The announcement, early in the morning of Monday,
September 15, 2008, that Lehman was insolvent triggered a
“flight to safety” that, by the close of trading that day, pushed
the yield on four-week Treasury bills down to 36 basis points,

The first response to the rising tide of
settlement fails was the decision of the
Federal Reserve to relax the terms of
its securities lending program.
100 basis points lower than the yield on the preceding Friday.
Yields on longer term bills also moved sharply lower. By the
close of trading on Wednesday, September 17, yields on fourweek bills were down to 7 basis points. Over the balance of the
month, four-week-bill yields fluctuated between about 10 basis
points and 100 basis points—well below the 1.50 to 1.85 percent
range that had prevailed since the beginning of August (Chart 1).
Greater demand for high-quality, short-term debt also
drove down repo rates on Treasury collateral. The overnight
Treasury general collateral repo rate averaged 90 basis points
between September 15 and September 30, well below the 2 percent level that had prevailed during the preceding six weeks
(Chart 2).
In the wake of Lehman’s insolvency and in the midst of the
ensuing flight to safety, investors became increasingly reluctant
to lend Treasury securities.52 Unable to replace their maturing
borrowings, dealers began to fail on their delivery obligations.
51

“Minutes of the Meeting of the Treasury Borrowing Advisory Committee,”
November 4, 2008, available at http://www.treas.gov/press/releases/
hp1239.htm (stating that “Since November 2003, Treasury has repeatedly
asked the private sector to address [the fails] issue proactively. On several
occasions, market participants have emphatically stated that they would
resolve the situation without government intervention, but such steps have not
been implemented.”). See also Wrightson, Federal Reserve Data, October 17,
2008 (stating that “the repo market has managed to fend off regulatory reform
in past cycles.”), and “The Treasury Market Reaches Breaking Point,”
Euromoney, December 1, 2008 (quoting a former Treasury employee as saying
that “It was politically difficult to convince the market to put a stop to fails to
deliver in treasuries. There were some forceful voices insisting that if the
Treasury got involved, they would take the incentives out of the specials market
altogether. Those making their living as specialist dealers, as well as those
making a living shorting securities outright, were worried about potential
supply changes which would eliminate trading opportunities for them.”).

Chart 1

Chart 2

Yields on Four-Week Treasury Bills

Target Federal Funds Rates and Rates on
Overnight General Collateral Repurchase
Agreements on Treasury Securities

August to October 2008
Percent

2.0

August to October 2008

2.0

2.5
Percent
1.5

2.5

1.5

Repo rates on dates
prior to September 15

Dates prior to September 15

2.0

2.0
1.0

1.5

1.0
1.5
Repo rates on dates
after September 14

0.5
0.5

Target federal funds rate

Dates after September 14

0.5

0.0
0.5

0
August

September

October

Source: Federal Reserve Statistical Release H.15.

The fails persisted because the low general collateral repo rate
left sellers with little incentive to cure the fails. Primary dealer
settlement fails in Treasury securities mushroomed to an
average of $253 billion per day during the week of Thursday,
September 18, to Wednesday, September 24—far in excess of
the level that had prevailed in August and the first half of
September (Chart 3). And unlike earlier episodes, fails in the
wake of Lehman’s insolvency were not concentrated in one or
two issues; rather, they involved securities across the entire
yield curve.
The first response to the rising tide of settlement fails was
the decision of the Federal Reserve to relax the terms of its
securities lending program. As shown in Table 1, on Tuesday,
September 23, the Fed raised the limit on total borrowings by a
single dealer from $3 billion to $4 billion. Loans to primary
dealers from the SOMA portfolio reached new heights but
primary dealer settlement fails continued to rise, averaging
$342 billion per day over the interval from September 25 to
October 8 (Chart 3).
On Wednesday, October 8, both the Federal Reserve and the
Treasury acted in response to the continuing crisis. The Federal
Reserve further eased the terms of its securities lending
program by reducing the minimum loan fee from 50 basis
52

1.0

“Demand for Short-Term Treasury Debt Puts a Crimp in World-Wide
Supply,” Wall Street Journal, September 25, 2008, p. C1 (reporting that “some
foreign central-bank officials … are reluctant to lend out their safest
collateral—U.S. Treasurys.”); “U.S. Treasury Steps Up Debt Sales to Reduce
Shortages (Update 2),” Bloomberg.com, October 8, 2008, 12:43 EDT (quoting
the head of interest rate strategy at Credit Suisse Securities as saying that
“people are so nervous about the financial crisis that they’re holding on to their
collateral and not lending it out.”); and “More Treasury Bonds on Way to Ease
Crisis,” Wall Street Journal, October 9, 2008, p. A6 (reporting that “investors
have been unwilling to lend Treasury securities to other market participants.”).

0.0
0
August

September

October
November

Sources: Federal Reserve Statistical Release H.15; Board of Governors
of the Federal Reserve System.

points to 10 basis points and by expanding the limit on total
borrowings by a single dealer to $5 billion (Table 1). Treasury
officials took the unprecedented step of reopening four off-therun Treasury notes, announcing at 10:40 a.m. that they would
“reopen multiple securities which have created severe
dislocations in the market causing acute, protracted
shortages.”53 Two of the reopened notes were auctioned later
the same day (at 11:30 a.m. and 1:00 p.m., respectively), and
the other two notes were auctioned the following day (Table 2).
The decision to reopen a substantial amount ($10 billion each)
of so many different notes made clear the scale of the fails
problem; the decision to auction one note with less than an
hour of notice and a second note with less than three hours of
notice emphasized the urgency of the situation.
Although the reopenings helped to mitigate settlement fails
in the issues that were reopened, aggregate primary dealer
Treasury fails continued to rise, reaching a daily average level of
$379 billion per day over the interval from October 9 to
October 22 (Chart 3). Comments to the effect that “Treasury
market functioning remains impaired” and “the repo market is
not functioning” became commonplace. On October 17, a
widely read market letter remarked that “the breakdown in the
clearing mechanism for the Treasury market is beginning to
emerge as a top-tier policy concern.”54

U.S. Department of the Treasury, “Statement on Treasury Market
Conditions and Debt Management Actions,” October 8, 2008, available
at http://www.treas.gov/press/releases//hp1186.htm.
54
Wrightson, Federal Reserve Data, October 17, 2008.

FRBNY Economic Policy Review / October 2010

Chart 3

Daily Average (Over Weekly Intervals) Primary Dealer Settlement Fails in Treasury Securities
August to October 2008
Billions of dollars
400
350
Denotes business days

300

Denotes weekends and holidays

250
200
150
100
50
0
August

September

October

Source: Federal Reserve Bank of New York.
Notes: The first square marks the Lehman insolvency on Monday, September 15. The second square marks the effective date of a revision in the
terms and conditions of the Federal Reserve System Open Market Account (SOMA) securities lending program on Tuesday, September 23.
The third square marks the announcement of the surprise reopening of four Treasury notes on Wednesday, October 8, and the effective date
of a further revision in the terms and conditions of the SOMA securities lending program on the same day.

Table 1

Terms and Conditions of Federal Reserve Security Loan Auctions
Limits on Outstanding Borrowings
by a Single Dealer
Theoretical Amount
of a Single Issue Offered
(Percentage of SOMA Holdings)a

Minimum Loan Fee
(Basis Points)

September 23, 2008

Per Issue

Total

Lesser of $750 million and 25 percent
of amount beneficially owned in
SOMA portfolio

$3 billion

Lesser of $750 million and 25 percent
of amount beneficially owned in
SOMA portfolio

$4 billion

October 8, 2008

Lesser of $750 million and 25 percent
of amount beneficially owned in
SOMA portfolio

$5 billion

December 18, 2008b

Lesser of $750 million and 25 percent
of amount beneficially owned in
SOMA portfolio

$5 billion

Effective Date
Terms and conditions prior
to Lehman insolvency
November 26, 2007

Post-Lehman terms and conditions

Source: Federal Reserve Bank of New York.
Notes: SOMA is the Federal Reserve System Open Market Account. Entries in bold indicate a change in terms.
a

Amount actually offered is the lesser of the theoretical amount offered and the amount of the issue actually in the SOMA account at the time of an auction.
Last revision prior to the end of 2008.

The Introduction of the TMPG Fails Charge

Table 2

Treasury Notes Reopened in October 2008
4 1/8 Percent Note Maturing
May 15, 2015

4 1/4 Percent Note Maturing
August 15, 2015

4 Percent Note Maturing
February 15, 2015

3 1/2 Percent Note Maturing
February 15, 2018

$10 billion

October 8, 2008, 11:30 a.m.

October 8, 2008, 1:00 p.m.

October 9, 2008, 11:30 a.m.

October 15, 2008

Amount bid competitively

$12.1 billion

$21.1 billion

$23.7 billion

$23.1 billion

Closing market yield on
October 7, 2008 (percent)

2.87

2.98

2.79

3.57

Auction yield (percent)

3.31

3.44

3.23

3.79

Closing market yield on
October 9, 2008 (percent)

3.35

3.57

3.22

3.92

Amount offered
Auction date and time
Issue date

Sources: U.S. Treasury Department; Wall Street Journal.
Note: Over the interval from October 7 to October 9, the closing market yield on the on-the-run five-year note (the 3 1/8 percent note of September 30,
2013) rose from 2.47 percent to 2.79 percent and the closing market yield on the on-the-run ten-year note (the 4 percent note of August 2018) rose
from 3.50 percent to 3.81 percent.

7. The TMPG Steps Up
By mid-October 2008, Treasury and Federal Reserve officials
and private sector market participants understood that the
volume and persistence of settlement fails in Treasury
securities was a major problem, but what could or should be
done was far from obvious. The four reopenings had reduced
fails in the reopened notes, but speculation over whether the
Treasury would reopen other chronically failing issues was
contributing to unwanted volatility in the prices of other
Treasury securities. Additionally, there was some indication
that the reopenings had not been well received. The first
auction, of $10 billion of the 4 1/8 percent notes of May 2015,
attracted only $12.1 billion of tenders, and the notes were sold
at a price almost 3 points below where outstanding notes of the
same series traded prior to the auction.
An alternative approach was to revise the market
convention of postponing—without any explicit penalty and at
an unchanged invoice price—a seller’s obligation to deliver
Treasury securities if the seller failed to deliver the securities on
a scheduled settlement date. However, precisely because the
treatment of settlement fails was a matter of market
convention, rather than law or regulation, it could not be
changed except through widespread adoption of an alternative
convention.
Fortuitously, in early 2007 the Federal Reserve Bank of
New York had sponsored the organization of a new forum—
the Treasury Market Practices Group—for discussing Treasury

market practices and for advocating the adoption of practices
deemed to be in the best interests of the market.55
The TMPG is a group of private sector market professionals
committed to supporting the integrity and efficiency of the
market for U.S. Treasury securities. Membership includes
senior business managers and legal and compliance
professionals from broker-dealer firms, banks, buy-side firms,

The TMPG is a group of private sector
market professionals committed to
supporting the integrity and efficiency of
the market for U.S. Treasury securities.
and other organizations involved in Treasury market
infrastructure. (Box 4 identifies the membership in October
2008.) The TMPG routinely meets about eight to ten times a
year to discuss trading issues and best-practice recommendations for the Treasury market and publishes “Treasury Market
Best Practices,” 56 a “living document” that aims to support
55

Federal Reserve Bank of New York, “Statement on Formation of PrivateSector Treasury Market Best Practices Group,” February 9, 2007, available at
http://www.newyorkfed.org/newsevents/news/markets/2007/an070209.html.
See also Federal Reserve Bank of New York, “Statement Regarding New York
Fed Meeting with Primary Dealers,” November 6, 2006, available at http://
www.newyorkfed.org/newsevetns/news-archive/markets/2006/an061105.html.
56
Available at http://www.newyorkfed.org/tmpg/TMPG-bestpractices_033109.pdf.

FRBNY Economic Policy Review / October 2010

Box 4

at a crossroad. . . . At this critical juncture it is incumbent
that TMPG take the leadership position on this issue
and work as a group to provide practical, real time
solutions. . . . Our goal as members of this committee
is to support the integrity and efficiency of the U.S.
Government Treasury Market. . . .

Membership of the Treasury Market Practices Group
in October 2008
Thomas Wipf, Chair

Morgan Stanley

Fran Bermanzohn

Goldman Sachs (last meeting
in October 2008)

Arthur Certosimo

The Bank of New York Mellon

Daniel Dufresne

Citadel Investment Group, LLC

Peter Economou

State Street

John Fath

BTG

Michael Haddad

Caxton Associates (last meeting
in January 2009)

Curt Hollingsworth

Fidelity Investments

James Hraska

Barclays Capital

Murray Pozmanter

Depository Trust & Clearing
Corporation

Gerald Pucci

BlackRock

John Roberts

Barclays Capital

Bill Santangelo

Countrywide Securities Corp.
(last meeting in October 2008)

Peter Stebbing

Reserve Bank of Australia

Nancy Sullivan

The Bank of New York Mellon

Matthew Zames

JPMorgan Chase

Subsequent Additions prior to May 1, 2009
Glenn Hadden

Goldman Sachs

Stuart Wexler

Merrill Lynch

Treasury market integrity and efficiency. Best-practice
recommendations include guidelines for promoting market
liquidity, for integrating compliance and trading functions in a
meaningful fashion, and for managing large positions in ways
that avoid adverse consequences for market liquidity. TMPG
practice guidance has also addressed the efficient clearing and
settlement of trades. Thus, the TMPG was well positioned in
October 2008 to provide the leadership required to revise the
market convention for settlement fails.57
The first meeting of the TMPG after the reopening auctions
of October 8 and 9 was on Thursday, October 23. The
chairman, Tom Wipf of Morgan Stanley, opened the meeting
by reminding members of the urgency of the situation:
To overstate the obvious, the work of today’s meeting
around settlement fails in Treasuries finds our committee
57

The Association of Primary Dealers provided similar leadership in revising
the market convention for the treatment of accrued interest in repurchase
agreements after the 1982 failure of Drysdale Government Securities
(Garbade 2006).

The Introduction of the TMPG Fails Charge

William Dudley, Executive Vice President of the Federal
Reserve Bank of New York and Manager of the Fed’s System
Open Market Account, echoed Wipf’s call for leadership: “This
[meeting] is happening at a critical time in the market place
where leadership is important to creating confidence and
stability—we believe this group can, should and will provide
that leadership.”

7.1 The November 12 Recommendations
During the October 23 meeting, and in a series of subsequent
telephone conference calls, TMPG members discussed changes
in market practices that might reduce chronic fails and limit
the likelihood of a recurrence. The group unveiled its
recommendations on Wednesday, November 12, 2008.58
The principal recommendation suggested that “market
participants agree that the invoice price . . . on any cash or
financing transaction that fails to settle on the originally
scheduled date be reduced at a fails [charge] rate equal to the
greater of a) 3 percent per annum minus the fed funds target
rate … and b) zero.” As shown in Chart 4, this would penalize
fails at a rate that starts at zero when the target federal funds
rate is at or above 3 percent and rises to 3 percent as the target
funds rate declines toward zero. It follows that the economic
cost of failing would never fall below about 3 percent per
annum.59 The TMPG concluded that the “out-of-pocket cost
to the party failing to deliver securities will provide a
compelling incentive to resolve fails promptly.”

Treasury Market Practices Group, “Treasury Market Practices Group
Endorses Several Measures to Address Widespread Settlement Fails,”
November 12, 2008, available at http://www.newyorkfed.org/tmpg/
PR081112.pdf. The recommendations were reported in “Treasury’s Warning
Adds to Plunge in Failed Trades,” Bloomberg.com, November 12, 2008,
15:59 EST, and “Repo Experts Propose Plans to Counteract Rise in ‘Fails,’”
FT.com, November 12, 2008, 20:00.
59
This follows because the sum of the target federal funds rate, which is usually
at or slightly above the Treasury general collateral repo rate, and the fails charge
rate would never be less than 3 percent. The TMPG could have referenced the
overnight Treasury general collateral repo rate in lieu of the target federal funds
rate, but the target funds rate is more familiar and more readily observable to
market participants. There is no definitive and widely disseminated measure of
overnight Treasury general collateral repo rates. The Federal Reserve, by
comparison, publicly announces the target funds rate.

The TMPG recognized that the “the introduction of [the
recommended convention] raises operational, legal and other
implementation issues that may vary across Treasury market
participants” and promised to engage in “further analysis of
these issues,” with a goal of announcing by January 5, 2009,
its recommendations for implementation.61

Chart 4

TMPG Fails Charge Rate as a Function
of Target Federal Funds Rate
TMPG fails charge rate (percent per annum)
3.0
2.5
2.0

8. The Crisis Recedes but Support
for Revising the Market
Convention Persists

1.5
1.0
0.5
0
0

1
2
3
4
Target federal funds rate (percent per annum)

The TMPG explicitly based its recommendation on the
dysfunctionality of the existing market convention for
settlement fails:
Past experience—for example, during the summer
of 2003—shows that settlement fails in a particular
[security] may become widespread and persistent when
the special collateral repo rate for that [security] nears
zero. Special collateral repo rates cannot exceed the
Treasury general collateral repo rate. As a result,
settlement fails across a wide variety of [securities]
can . . . become widespread and persistent when the
Treasury general collateral repo rate is near zero—
as is currently the case.
The underlying problem is the Treasury market
contracting convention that a seller can deliver securities
after the originally scheduled settlement date at an
unchanged invoice price [and] without incurring any
penalty. Introduction of a dynamic fails [charge] with a
finite cap rate would remedy this problem. In particular,
a dynamic fails [charge] would provide an incentive for
sellers to resolve fails promptly, and could lead to repo
contracting conventions [that is, negative repo rates] that
would give beneficial owners of Treasury securities an
opportunity to earn as much as the [3 percent] cap rate in
securities loan fee income regardless of the level of
nominal interest rates.60
60

Treasury Market Practices Group, “Treasury Market Practices Group
Endorses Several Measures to Address Widespread Settlement Fails,”
November 12, 2008.

By the time the TMPG made its November 12 recommendation,
settlement fails in the Treasury market were receding rapidly.
As shown in Chart 5, primary dealer fails declined from a daily
average of $379 billion during the week of October 16-22 to a
daily average of $70 billion during the week of November 13-19
and averaged less than $50 billion a day in December.
Support for a revised market convention for settlement
fails persisted in spite of the receding volume of fails, largely
because the crisis of late September and early October had
given added currency to the view that the existing convention
was dysfunctional. The discussion of settlement fails during
the November 4, 2008, meeting of the Treasury Borrowing
Advisory Committee, as well as the views expressed in a
prominent market newsletter in early January 2009, illustrates
the growing consensus.62

The TMPG made three additional recommendations on November 12:
1) that market participants undertake a study of the most efficient way to
margin fails in Treasury securities (in order to reduce counterparty credit risk
exposure), 2) that market participants examine whether the Fixed Income
Clearing Corporation, the two clearing banks, or other interested parties might
develop “new or enhanced … multilateral netting arrangements” that might
reduce settlement fails, and 3) that market participants pursue consensual cash
settlement of transactions in Treasury securities that have been failing for more
than five days. The TMPG also expressed its support for “discussion of a
standing facility by the U.S. Department of the Treasury to provide temporary
new supply of specific securities at a penalty rate when settlement fails persist,”
but noted that the creation of such a facility was a long-term goal and that
progress on a fails charge should not be contingent on the development of
a Treasury security lending facility.
62
William Dudley, Executive Vice President, Federal Reserve Bank of New
York, and Manager of the System Open Market Account, stated during a public
conference call on January 14, 2009, on the TMPG fails initiative that:
“Although settlement fails have declined recently from record levels amid
reduced trading volumes, the extremely low level of interest rates suggests
that fails could again rise significantly when trading activity picks up. The
fundamental incentive to deliver securities under current market conditions
is simply not sufficient at very low nominal interest rates to reduce the
probability of large chronic fails to acceptable levels.”

FRBNY Economic Policy Review / October 2010

Chart 5

Daily Average (Over Weekly Intervals) Primary Dealer Settlement Fails in Treasury Securities
September 2008 to April 2009
Billions of dollars
400
350
300
250
200
150
100
50

2008

Apr. 16 - Apr. 22

Apr. 2 - Apr. 8

Mar. 19 - Mar. 25

Mar. 5 - Mar. 11

Feb. 19 - Feb. 25

Feb. 5 - Feb. 11

Jan. 22 - Jan. 28

Jan. 8 - Jan. 14

Dec. 25 - Dec. 31

Dec. 11 - Dec. 17

Nov. 27 - Dec. 3

Nov. 13 - Nov. 19

Oct. 30 - Nov. 5

Oct. 16 - Oct. 22

Oct. 2 - Oct. 8

Sep. 18 - Sep. 24

Sep. 4 - Sep. 10

2009

Source: Federal Reserve Bank of New York.

8.1 The November Meeting of the Treasury
Borrowing Advisory Committee
The Treasury Borrowing Advisory Committee (TBAC) is a
committee of market professionals selected to advise the
Secretary of the Treasury on matters relating to Treasury debt
management. At its November 4 meeting, the committee
discussed the upcoming midquarter refunding and, inter alia,
the fails situation.
The TBAC’s discussion of settlement fails focused initially
on better ways for the Treasury to reopen outstanding issues
than the “snap” reopenings of October 8 and 9, but then turned
to the market convention for settlement fails. Several
committee members observed that investors had “little
economic incentive to lend securities when general collateral
[repo] rates stood at 20 basis points,” and one member
suggested that “a negative [repo] rate of 200 or 300 basis points
. . . would create the correct economic incentives to cause
holders of securities in low interest rate environments to lend
securities again.”63

The Introduction of the TMPG Fails Charge

In its ensuing report to the Secretary of the Treasury, the
TBAC expressed the view that the low level of short-term
interest rates “has made the cost of failing negligible, [leaving]
little desire for short-sellers to close out their positions” and
noted the suggestion of one committee member “that there
should be a cost in the form of a penalty rate associated with
fails in a low-rate environment.” The report further noted that
such a cost would encourage negative-rate repo trading,
“which would allow the free market to determine the effective
cost of the fail, and change the economics of securities
lending.”64

63
U.S. Department of the Treasury, “Minutes of the Meeting of the Treasury
Borrowing Advisory Committee of the Securities Industry and Financial
Markets Association, November 4, 2008,” November 5, 2008, available at
http://www.ustreas.gov/press/releases/hp1239.htm.
64
U.S. Department of the Treasury, “Report to the Secretary of the Treasury
from the Treasury Borrowing Advisory Committee of the Securities Industry
and Financial Markets Association, November 4, 2008,” November 5, 2008,
available at http://www.ustreas.gov/press/releases/hp1238.htm.

8.2 Comments in Wrightson’s
Money Market Observer
The January 5, 2009, edition of Wrightson’s Money Market
Observer devoted substantial space to the TMPG proposal to
revise the market convention for settlement fails. The
newsletter noted that creating an explicit charge for settlement
fails was “the most straight-forward way to remedy the obvious
structural flaws that lead to delivery logjams in today’s
market,” and pointed out the anticipated benefits of restoring
competitive market forces to the special collateral repo
markets: “The TMPG believes (correctly, in our view) that the
repo market will be more elastic—and the Treasury clearing
process more efficient—if the floor on special repo rates is set
low enough [that is, below zero] to preserve a spread relative to
general collateral rates even in the current rate environment.”

In lieu of adjusting invoice prices, several TMPG members
suggested that an economically equivalent result could be
obtained if a seller who makes a late delivery agrees to make a
side payment to the buyer in an amount equal to what became
known as the “TMPG fails charge.” The charge for a fail on a
given business day would be computed as:
n
C = ---------  .01  max  3 – R trgt  0   P ,
360

(1)
where:

C = charge, in dollars,
n = number of calendar days to the next following

business day,
R trgt = target federal funds rate at the close of business

on the business day preceding the fail, in percent
per annum, and
P = total proceeds due from the buyer, in dollars.

9. Getting It Right
Although a consensus had emerged in support of revising the
market convention for settlement fails, the TMPG
recommendation for reducing invoice prices itself needed
some revision.
The TMPG recommendation would have required a seller
and a buyer to reduce the invoice price on a failing transaction
by matching amounts on a daily basis. If one party reduced the

Although a consensus had emerged in
support of revising the market convention
for settlement fails, the TMPG
recommendation for reducing invoice
prices itself needed some revision.
invoice price and the other did not, or if the two parties
reduced the invoice price by different amounts, any attempt by
the seller to deliver securities against payment would be
rejected by the buyer (because the buyer would be looking to
pay a different amount than what the seller was looking to
receive). TMPG members who understood the complex
architecture of broker-dealer and custodian settlement systems
pointed out that requiring matching daily price reductions
would impose a major operational burden on market
participants and could lead to an explosion in rejected
deliveries (and thus in settlement fails).

For example, if P = $10 million, R trgt = 1 percent, and n =
three days, then C = $1,666.66.65 This procedure had the
advantage of not requiring any change in existing settlement
systems.
The idea of replacing invoice price adjustments with side
payments illustrates an important aspect of the TMPG
initiative: by working collaboratively, the TMPG was able to
achieve its objectives while accommodating an existing
institutional structure: back-office settlement systems. The
difference between a price adjustment and a side payment may
seem trivial, but the success of the TMPG initiative hinged on
recognizing the difference.

9.1 The January 5 Announcement
On January 5, 2009, the TMPG announced that it was
recommending a fails charge in the form of a side payment
on transactions that failed to settle on a timely basis and
that it was making several additional refinements to its
November 12 recommendation.66 The three key refinements
$1,666.66 = (3/360)  .01  max[3 1, 0]  $10,000,000.
Treasury Market Practices Group, “Timeline for New Market Practices to
Address Widespread Settlement Fails in U.S. Treasury Securities,” January 5,
2009, available at http://www.newyorkfed.org/tmpg/PR090105c.pdf, and
Treasury Market Practices Group, “Claiming a Fails Charge for a Settlement
Fail in U.S. Treasury Securities,” January 5, 2009, available at http://
www.newyorkfed.org/tmpg/PR090105a.pdf. The January 5 announcement
was reported in “Repo Arena Gets a Plan on Penalties,” Wall Street Journal,
January 6, 2009, p. C3, and “Penalty for Failed Trades Set to Spark New Era for
US Repo,” Financial Times, January 7, 2009, p. 22.

65
66

FRBNY Economic Policy Review / October 2010

were 1) a statement of the process for claiming a fails charge, 2)
a timeline suggesting that market participants begin claiming
for settlement fails on transactions agreed to on or after May 1,
2009, and 3) replacement of the target federal funds rate (in the
formula for the fails charge, equation 1 above) with a “TMPG
reference rate.” The latter rate was defined as the target federal
funds rate if the Federal Open Market Committee specified a
target rate or the lower limit of the target band for the federal
funds rate if the FOMC specified a target band.67 In the event
the FOMC specified neither a target rate nor a target band, the

On January 5, 2009, the TMPG announced
that it was recommending a fails charge in
the form of a side payment on transactions
that failed to settle on a timely basis and
that it was making several additional
refinements to its November 12

The January 5 announcement suggested that the best way to
initiate the fails charge would be for buyers to tender claims
directly to sellers.70 A seller could either pay what was claimed
or dispute the claim and negotiate with its counterparty over
the amount due.
The TMPG further suggested that if an investor employed a
professional asset manager and that manager contracted to sell
securities that were not delivered on a timely basis, the claim
for the fails charge should be directed to the asset manager
(rather than to the investor or to the investor’s custodian). This
suggestion was based on the pragmatic notion that since the
sale had been negotiated by the asset manager, the asset
manager would be in the best position to recognize the sale and
identify who was responsible for the settlement fail, be it the
asset manager, the investor’s custodian, or some other party,71
or whether the claim should be left for the account of the
investor.72

9.2 Trading Practice and Market Practice
Recommendations

recommendation.

TMPG committed to recommending some other similar,
readily observable, short-term interest rate as a reference rate
for the fails charge formula.68
The decision to recommend a side payment (in lieu of an
invoice price adjustment) required the TMPG to specify a way
for buyers to collect from sellers who failed to deliver securities
on a timely basis. In the case of buyers and sellers who settled
through FICC, a collection process could be added to other
similar processes previously implemented by FICC (such as
the collections and disbursements that result from marking
transactions to current market prices).69 However, the
collection process was not as simple for transactions that
settled bilaterally, as was the case for most transactions between
dealers and their nondealer customers.
67

This charge was necessitated by the December 16, 2008, decision of the
Federal Open Market Committee to establish a target range for the federal
funds rate of 0 to 1/4 percent.
68
In late March 2009, the TMPG announced a slightly different form for the
fails charge computation:

1
C = ---------  .01  max 3 – RTMPG  0   P .
360
In this form, the charge is computed for each calendar day that a seller’s
delivery obligation is failing. R TMPG is the TMPG reference rate on the
business day preceding the day for which the charge is computed. See Treasury
Market Practices Group, “Treasury Market Practices Group Announces
Updates to Fails Charge Recommendation,” March 30, 2009, available at
http://www.newyorkfed.org/tmpg/tmpg_033009.pdf.

The Introduction of the TMPG Fails Charge

Following the January 5 announcement, TMPG members
and other market participants collaborated to publish two
documents providing guidance on how to implement the
TMPG fails charge. The documents were important for
clarifying how fails charges should be calculated and claimed
and generally for enhancing the transparency of the new
market convention.
69

The addition required a change in FICC rules that had to be approved by the
Securities and Exchange Commission. FICC filed the proposed rule change on
February 25, 2009 (Securities and Exchange Commission Release no. 3459569, March 12, 2009), and the Commission granted approval two months
later (Securities and Exchange Commission Release no. 34-59802, April 20,
2009).
70
The January 5 announcement noted the possibility of setting up a central
industry utility to receive and process claims, but observed that the design of
such a facility raised novel questions regarding the identification of buyers and
sellers and would require further consultation with market participants.
71
Some investors retain an agent to lend securities from the investor’s
portfolio. Such agents are commonly called “agent sec lenders.” In most cases,
an agent sec lender is obliged to reclaim securities out on loan if the investor’s
asset manager decides to sell the securities. If an agent sec lender fails to reclaim
securities on a timely basis and thereby causes a settlement fail, the fail may be
the responsibility of the agent sec lender, rather than the asset manager or the
custodian.
72
A sale of securities negotiated by an asset manager may fail to settle on a
timely basis because the investor’s custodian failed to receive the same
securities on an unrelated purchase. Such fails cannot be attributed to faulty
behavior by the asset manager or the investor’s custodian, so the resulting fails
charge would be left for the account of the investor. The investor can, of course,
direct its asset manager to file a claim on the seller who failed to deliver
securities to the investor.

Trading practice recommendations: On January 15, 2009, the
TMPG and the Securities Industry and Financial Markets
Association (SIFMA) published a “U.S. Treasury Securities
Fails Charge Trading Practice”73 to give market participants

TMPG members and other market
participants collaborated to publish two
documents providing guidance on how to
implement the TMPG fails charge. The
documents were important for clarifying
how fails charges should be calculated
and claimed and generally for enhancing
the transparency of the new market
convention.
guidance on exactly the types of transactions that were covered
by, and excluded from, the TMPG fails charge. The “Trading
Practice” also recommended the form of a letter that a market
participant could send to counterparties, advising them of the
participant’s adoption of the new policy for settlement fails,
and suggested a statement that could be added to trade
confirmations indicating that a transaction was subject to the
fails charge.

Market practice recommendations: On April 23, 2009, SIFMA
published a “Treasury Market Practices Group Fails Charge
Market Practice”74 that recommended procedures for buy-side
firms to use in connection with the new fails charge. The
recommended procedures included processes for researching
and tracking fails, calculating fails charges, determining
responsibility for a claim, sending and receiving claims, and
accounting for claims. The procedures also included
suggestions made earlier by SIFMA and adopted by the
TMPG75 that claims be submitted at the beginning of a month
for fails settled during the preceding month (to accommodate
custodians and asset managers who structured their control
73

Treasury Market Practices Group and Securities Industry and Financial
Markets Association, “U.S. Treasury Securities Fails Charge Trading Practice,”
January 15, 2009, available at http://www.sifma.org/capital_markets/docs/
Fails-Charge-Trading-Practice.pdf.
74
Securities Industry and Financial Markets Association, “Treasury Market
Practices Group Fails Charge Market Practice,” April 23, 2009, available at
http://www.theasset manager.com/docs/AM_Custodian_IndustryProcedures_
TMPG_FailsCharge.pdf.
75
See Treasury Market Practices Group, “Treasury Market Practices Group
Announces Updates to Fails Charge Recommendation,” March 30, 2009,
available at http://www.newyorkfed.org/tmpg/tmpg_033009.pdf.

systems around settled transactions) and be in excess of $500
per issue per settlement (to limit costly research and billing
efforts to nontrivial claims).

10. Implementation
The TMPG fails charge went into effect on May 1, 2009,
replacing the former market convention of postponing—
without any explicit penalty and at an unchanged invoice
price—a seller’s obligation to deliver Treasury securities when
the seller fails to deliver the securities on a scheduled settlement
date. Henceforth, the cost of failing to settle a sale of Treasury
securities in a timely fashion would not be less than 3 percent
per annum.
It would be premature to argue that the TMPG fails charge
has eliminated the possibility of yet another episode of chronic
settlement fails in Treasury securities; past episodes were rare
to begin with and some future event may demonstrate the
existence of an unsuspected flaw in the new system. It may also
be the case that the 3 percent benchmark rate is too low and
that chronic fails would be better mitigated with a 3 1/2 or
4 percent rate. However, there is no evidence to date that the

The new convention is not yet out of its
infancy, but there is reason to anticipate
that the TMPG fails charge will . . . dampen
future eruptions [of settlement fails].
new market convention, and the 3 percent benchmark rate, are
not working. Chart 6 shows daily average settlement fails over
weekly intervals from the beginning of 2009 to July 2010. Fails
averaged a bit over $14.4 billion per day during the first four
months of 2009, but only $4.2 billion per day since
implementation of the fails charge.76 More important, the
relatively modest eruptions of settlement fails that appeared
during the first week of July 2009 and the first week of January
2010 quickly subsided. The new convention is not yet out of its
infancy, but there is reason to anticipate that the TMPG fails
charge will similarly dampen future eruptions.

The fails charge was never intended to eliminate all settlement fails. (Fails
attributable to miscommunication or operational problems are unlikely to be
eliminated by the fails charge—although they may be resolved more quickly.)
Rather, the fails charge was aimed at mitigating episodes of chronic fails that
can threaten market liquidity and efficiency.

FRBNY Economic Policy Review / October 2010

Chart 6

Daily Average (Over Weekly Intervals) Primary Dealer Settlement Fails in Treasury Securities
January 2009 to July 2010
Billions of dollars
35
30

After May 1, 2009

Prior to May 1, 2009

25
20
15
10
5

2009

Jul. 15 - Jul. 21

Jun. 17 - Jun. 23

May 20 - May 26

Apr. 22 - Apr. 28

Mar. 25 - Mar. 31

Feb. 25 - Mar. 3

Jan. 28 - Feb. 3

Jan. 1 - Jan. 6

Dec. 3 - Dec. 9

Nov. 5 - Nov. 11

Oct. 8 - Oct. 14

Sep. 10 - Sep. 16

Aug. 13 - Aug. 19

Jul. 16 - Jul. 22

Jun. 18 - Jun. 24

May 21 - May 27

Apr. 23 - Apr. 29

Mar. 26 - Apr. 1

Feb. 26 - Mar. 4

Jan. 29 - Feb. 4

Jan. 1 - Jan. 7

2010

Source: Federal Reserve Bank of New York.

11. Conclusion
The TMPG fails charge initiative is important both for what it
accomplished and for how it was accomplished. Substantively,
the initiative revised an outmoded convention and mitigated
an important dysfunctionality in a market of critical national
significance. Procedurally, the initiative demonstrated how
cooperation between the public and private sectors can speed
innovative and efficient responses to changing circumstances.
At the time of the May 1, 2009, implementation of the fails
charge, the Federal Reserve Bank of New York welcomed the
new convention:
We applaud the dedicated efforts of the TMPG in
spearheading the development and implementation of
this targeted solution to the settlement fails problem. This
significant milestone in the evolution of Treasury market
practice demonstrates that groups, such as the TMPG, are
effective in addressing deficiencies in market functioning
and facilitating market best practices.77

The Introduction of the TMPG Fails Charge

In a subsequent letter to the TMPG membership expressing his
personal thanks for the Group’s dedication and commitment
to making the fails charge a reality, William Dudley, now
president of the Federal Reserve Bank of New York, reflected
on the significance of the new market convention:
The implementation of the fails charge marks a rare and
significant evolution in Treasury market architecture. In
my view, one would need to look back to 1982 to find a
development of similar magnitude, when the collapse of
Drysdale Securities led to the adoption of a new market
practice to include accrued interest in repo contracts.
The fails charge stands among relatively few revisions
to contracting conventions in the Treasury market since
the development of a liquid national market following
World War I.
77

Federal Reserve Bank of New York, “New York Fed Applauds
Implementation of the TMPG’s Fails Charge Recommendation,” May 1, 2009,
available at http://www.newyorkfed.org/newsevents/news/markets/2009/
ma090501.html.

Appendix: Additional Individuals Who Provided Feedback and Assistance
in the Implementation Phase of the TMPG Fails Charge Initiative

The following list attempts to include all non-TMPG (Treasury Market Practices Group)
individuals who either served on formal subgroups that contributed to the timely implementation
of the TMPG fails charge or participated in less formal conference calls and meetings. The authors
apologize for any inadvertent omissions.
David Aman
Marc Baum
Brandon Becker
Brent Blake
Gary Buki
Kevin Caffrey
Maria Carina
Michael Cetta
Brayton Cherry
Ed Corral
David Cosgrove
Brian Crowe
Craig Delany
Laura Dietel
Craig Dudsak
Steve Dunn
Joe Finan
Marcellus Fisher
Barbara Friedman
Robert Good
Simon Griffiths
Olivier Grimonpont
Chris Haas
Janice Hamilton
Robert Hennessy
Eugene Ing
Dyann Kiessling
Bradley Koehler
Marty Kruse
Michael Landolfi
Lourdes Leon
Christine Lin
Fred Lipinski
Colin Lloyd
Frank Lupica
Diane Madera
Claudia Maia
Jennifer Manor
Frank Martino
Christopher Marzullo
Jason McCann
Shirley McCoy
Katherine McGaugh
Kevin Meagher
Omar Medina

Cleary Gottlieb Steen & Hamilton LLP
Ramius LLC
WilmerHale
State Street
The Bank of New York Mellon
The Bank of New York Mellon
Euroclear
AllianceBernstein Holdings
Brown Brothers Harriman & Co.
JPMorgan Chase
ICAP
Fidelity Investments
JPMorgan Chase
State Street
Citigroup
JPMorgan Chase
Morgan Stanley & Co.
PIMCO
New York Life Investment Management LLC
Goldman, Sachs & Co.
JPMorgan Chase
Euroclear
Merrill Lynch
The Northern Trust Company
The Bank of New York Mellon
Depository Trust & Clearing Corporation
Fidelity Investments
Euroclear
BNY Mellon Asset Servicing
State Street
Morgan Stanley & Co.
Citigroup
Depository Trust & Clearing Corporation
Cleary Gottlieb Steen & Hamilton LLP
Lord, Abbett & Co. LLC
Morgan Stanley & Co.
Euroclear
Fidelity Investments
New York Life Investment Management LLC
Lord, Abbett & Co. LLC
Reserve Bank of Australia
JPMorgan Chase
State Street
Fidelity Investments
Goldman Sachs Asset Management

FRBNY Economic Policy Review / October 2010

Appendix: Additional Individuals Who Provided Feedback and Assistance
in the Implementation Phase of the TMPG Fails Charge Initiative (Continued)
Fatima Mehladi
Eric Miller
Tamara Molinary
Carolyn Monroe-Koatz
Penny Morgan
Louis Nazarro
Edward Neeck
Peter Novello
Elisa Nuottajarvi
Paul Parseghian
Judith Polzer
Joseph Pomo
Thomas Ponti
Nancy Prior
Pawan Puneet
Christopher Ramsay
Bill Rose
Theodore Rothschild
Timothy Ryan
Joseph Sack
Randy Snook
Guido Stroemer
Brian Swann
Rob Toomey
Diane Trupia
Raymond Tyrrell
Jason Ward
Andrew Waskow
Mark Willis
Patricia Yak
Lawrence Young
Anthony Zook

Euroclear
Citadel Solutions, LLC
AllianceBernstein Holdings
JPMorgan Chase
Western Asset Management
JPMorgan Chase
JPMorgan Chase
Lord, Abbett & Co. LLC
SIFMA
Prudential Investment Management Inc.
JPMorgan Chase
Goldman Sachs Asset Management
State Street
Fidelity Investments
State Street IMS
Citadel Investment Group, LLC
BTG
JPMorgan Chase
SIFMA
SIFMA
SIFMA
UBS
Goldman Sachs
SIFMA
SIFMA
Brown Brothers Harriman & Co.
Fidelity Investments
Goldman Sachs
Merrill Lynch
Credit Suisse First Boston
Credit Suisse First Boston
JPMorgan Chase

The Introduction of the TMPG Fails Charge

References

Committee on Banking, Finance, and Urban Affairs. 1985. “The Federal
Reserve Bank of New York Discount Window Loan of $22.6 Billion
Extended to the Bank of New York.” Hearing before the
Subcommittee on Domestic Monetary Policy of the Committee
on Banking, Finance, and Urban Affairs. U.S. House of
Representatives, 99th Cong., 1st Sess., December 12.
Cornell, B., and A. Shapiro. 1989. “The Mispricing of U.S. Treasury
Bonds: A Case Study.” Review of Financial Studies 2, no. 3
(fall): 297-310.
Department of the Treasury. 2006. “Consideration of a Proposed
Treasury Securities Lending Facility.” Federal Register 71,
no. 85 (May 3): 26,174-8.
Department of the Treasury, Securities and Exchange Commission, and
Board of Governors of the Federal Reserve System. 1992. “Joint
Report on the Government Securities Market.” January.
Duffie, D. 1996. “Special Repo Rates.” Journal of Finance 51, no. 2
(June): 493-526.
Fisher, M. 2002. “Special Repo Rates: An Introduction.” Federal
Reserve Bank of Atlanta Economic Review 87, no. 2 (second
quarter): 27-43.
Fleming, M., and K. Garbade. 2002. “When the Back Office Moved
to the Front Burner: Settlement Fails in the Treasury Market after
9/11.” Federal Reserve Bank of New York Economic Policy
Review 8, no. 2 (November): 35-57.
———. 2003. “The Repurchase Agreement Refined: GCF Repo.”
Federal Reserve Bank of New York Current Issues in Economics
and Finance 9, no. 6 (June).
———. 2004. “Repurchase Agreements with Negative Interest Rates.”
Federal Reserve Bank of New York Current Issues in Economics
and Finance 10, no. 5 (April).

———. 2007. “Dealer Behavior in the Specials Market for U.S.
Treasury Securities.” Journal of Financial Intermediation 16,
no. 2 (April): 204-28.
Garbade, K. 2004. “Origins of the Federal Reserve Book-Entry
System.” Federal Reserve Bank of New York Economic Policy
Review 10, no. 3 (December): 33-50.
———. 2006. “The Evolution of Repo Contracting Conventions in the
1980s.” Federal Reserve Bank of New York Economic Policy
Review 12, no. 1 (February): 25-40.
———. 2007. “The Emergence of ‘Regular and Predictable’ as a
Treasury Debt Management Strategy.” Federal Reserve Bank of
New York Economic Policy Review 13, no. 1 (March): 53-71.
Garbade, K., and J. Ingber. 2005. “The Treasury Auction Process:
Objectives, Structure, and Recent Adaptations.” Federal Reserve
Bank of New York Current Issues in Economics and
Finance 11, no. 2 (February).
Garbade, K., and J. Kambhu. 2005. “Why Is the U.S. Treasury
Contemplating Becoming a Lender of Last Resort for Treasury
Securities?” Federal Reserve Bank of New York Staff Reports,
no. 223, October.
Ingber, J. 2006. “A Decade of Repo Netting.” Futures & Derivatives
Law Report 26, no. 2 (February): 3-12.
Jordan, B., and S. Jordan. 1997. “Special Repo Rates: An Empirical
Analysis.” Journal of Finance 52, no. 5 (December): 2051-72.
Keane, F. 1996. “Repo Rate Patterns for New Treasury Notes.” Federal
Reserve Bank of New York Current Issues in Economics and
Finance 2, no. 10 (September).
Sender, H. 1986. “The Day the Computers Went Down.”
Institutional Investor 11, no. 3 (March): 203-7.

———. 2005. “Explaining Settlement Fails.” Federal Reserve Bank of
New York Current Issues in Economics and Finance 11, no. 9
(September).

Full text of Economic Policy Review (Federal Reserve Bank of New York) : Volume 16, Number 2

FRASER