Full text of Business Review (Federal Reserve Bank of Philadelphia) : Second Quarter 2014

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

Second Quarter 2014
Volume 97, Issue 2

© Michael Hoffman

Camden County Boathouse, New Jersey

The Paper Trail of Knowledge Transfers 						
Shadow Banking and the Crisis of 2007-08						
Forecast Disagreement in the Survey of Professional Forecasters
Research Rap

INSIDE
Issn: 0007-7011

Second QUARTER 2014

The Business Review is published four
times a year by the Research Department of
the Federal Reserve Bank of Philadelphia.
The views expressed by the authors are not
necessarily those of the Federal Reserve.
We welcome your comments at PHIL.
BRComments@phil.frb.org.

The Paper Trail of Knowledge Transfers

For a free subscription, go to www.
philadelphiafed.org/research-and-data/
publications. Archived articles may be
downloaded at www.philadelphiafed.
org/research-and-data/publications/
business-review. To request permission
to reprint articles in whole or in part,
click on Permission to Reprint at www.
philadelphiafed.org/research-and-data/
publications. Articles may be photocopied
without permission. Microform copies may
be purchased from ProQuest Information
and Learning, 300 N. Zeeb Road, Ann
Arbor, MI 48106.
The Federal Reserve Bank of Philadelphia
helps formulate and implement monetary
policy, supervises banks and bank and
savings and loan holding companies, and
provides financial services to depository
institutions and the federal government. It
is one of 12 regional Reserve Banks that,
together with the U.S. Federal Reserve
Board of Governors, make up the Federal
Reserve System. The Philadelphia Fed
serves eastern Pennsylvania, southern New
Jersey, and Delaware.
Charles I. Plosser
President and Chief Executive Officer
Loretta J. Mester
Executive Vice President and
Director of Research
Colleen Gallagher
Research Publications Manager
Dianne Hallowell
Art Director and Manager

1

Why do firms tend to locate near other firms? Economists suspect that
geographic clustering spurs innovation by letting businesses tap a climate
rich in informal transfers of knowledge. By tracing links between inventors
filing for patents for the same inventions, Jeffrey Lin shares new evidence
supporting the idea that proximity offers businesses tangible benefits.

Shadow Banking and the Crisis of 2007-08

7

In recent decades, institutions that function much like traditional banks
have grown outside regulatory oversight. Yet, as Daniel Sanches explains,
these so-called shadow banks are as vulnerable to runs as regular banks.
Because banking crises can inflict lasting economic harm, economists are
interested in tracing how the panic ensued in the shadow system.

Forecast Disagreement in the Survey
of Professional Forecasters

15

To enact effective policies and spend resources efficiently, firms,
policymakers, and markets need accurate economic forecasts. But even
though economists generally work with similar models and data, their
projections often range widely. To better understand why, Keith Sill
explores what the evidence and theories say about how forecasters form
their views.

Research Rap
Abstracts of the latest working papers produced by the Research
Department of the Federal Reserve Bank of Philadelphia.

25

The Paper Trail of Knowledge Transfers
By Jeffrey Lin

W

hy do firms and inventors tend to locate in dense, costly
areas? One intriguing hypothesis is that such geographic
clustering lets them benefit from local knowledge spillovers.
As Nobel laureate Robert Lucas has noted, the benefits
of one person’s knowledge spilling over to others play a
central role in economic growth and the existence of cities: “What
can people be paying Manhattan or downtown Chicago rents for, if
not for being near other people?” Proximity may improve the sharing
of knowledge, the matching of ideas to firms, or the rate of learning.1
If dense clustering indeed confers these benefits, then that raises the
possibility that individuals and firms may not be fully taking them into
account when deciding where to locate, resulting in underinvestment in
new ideas.

A key counterargument to the importance of knowledge spillovers is that
firms might prefer to keep their work
secret from competitors. For example,
many firms include nondisclosure and
noncompete clauses in employment
contracts for researchers and scientists. Yet, as Alfred Marshall suggested,
knowledge is difficult to keep secret:
“The mysteries of the trade become no
mysteries; but are as it were in the air.”
So are, in fact, knowledge spillovers an important reason why inventors tend to locate near one another?
We know that other factors might
also encourage firms and inventors to
locate near one another. For example,

Gerald A. Carlino’s 2001 Business Review
article discusses such mechanisms.

firms might benefit from better matches with specialized workers.2 They
may benefit from the sharing of local
production inputs such as cheap electrical power or hard-to-find machinery
and parts. Or skilled inventors may be
attracted to superior amenities such as
restaurants, shopping, or safety.
A key challenge, then, is to account for these alternative explanations so that we do not erroneously
infer that knowledge spillovers are
empirically important. To explore
this challenge, I review the empirical literature regarding evidence of
knowledge spillovers contained in
patent citations and nonpatent data.
I then describe the evidence that Ina

1

2

See my 2009 Business Review article.

Jeffrey Lin is an economic advisor and economist at the Federal
Reserve Bank of Philadelphia. The views expressed in this article
are not necessarily those of the Federal Reserve. This article and
other Philadelphia Fed reports and research are available at www.
philadelphiafed.org/research-and-data/publications.

www.philadelphiafed.org

Ganguli, Nick Reynolds, and I found
using a novel measure — cases of
simultaneous invention that result in
rival claims known as patent interferences. First, let us look at what
researchers have found by studying
routine patent applications.
EVIDENCE FROM PATENT
CITATIONS
At its most basic, the challenge of
verifying the existence of knowledge
spillovers was observed by Paul Krugman — namely, that knowledge flows
are invisible: “They leave no paper
trail by which they may be measured or
tracked.” Adam Jaffe, Manuel Trajtenberg, and Rebecca Henderson tackled
this problem by observing that the
flow of knowledge from one inventor
to another could, in fact, be tracked
using patent citations.3 Their paper
and ones that followed have provided
the best evidence to date that local
knowledge spillovers might be one
important mechanism contributing to
the geographic proximity of inventors.4
They exploit the fact that patents
include citations to older patents. If
a new patent cites a previous patent,
this citation is evidence that the older
patent contains knowledge upon which
the new patent relies.
Though a citation to a nearby
inventor is at first glance evidence that
knowledge has passed from the earlier
inventor to the citing inventor, it does

Robert M. Hunt’s 2001 Business Review article
discusses patents as a measure of knowledge
production.

3

4
See also the work by Jaffe (1986) and Keller
(2002), and the surveys by Rosenthal and
Strange (2004) and Audretsch and Feldman
(2004).

Business Review Q2 2014 1

not necessarily indicate that geographic proximity has facilitated this
transfer of knowledge. It might simply
be the case that inventors are located
nearby for some other reasons besides
the opportunity to take advantage of
knowledge spillovers, such as to be
near some common physical input to
the invention process. For example, if a
patent awarded to researchers at a University of Pennsylvania hospital is cited
in a subsequent application for a patent
awarded to researchers at Temple University, that might be because of local
knowledge spillovers. But it could also
be the case that these patentees are
near each other simply because many
hospitals are needed to serve the large
population of the Philadelphia area,
and proximity offers no advantage in
transmitting knowledge.
Expected proximity. To address
this inference problem, Jaffe and his
coauthors develop a clever matching
strategy. They measure the distance
between an earlier “originating” patent
and a subsequent “citing” patent that
references the originating patent as an
important knowledge input. Then they
compare this distance to the distance
between the originating patent and a
matched “control” patent. The control patent is similar to the matched
citing patent in terms of the date of
invention and technological classification, but it does not cite the matched
originating patent. Thus, the control
patent represents the expected proximity of inventors working in the same
research field and time period, not
conditioned on a citation link. If the
inventors of the citation-linked patent
pair are observed in closer proximity
versus this benchmark, then this is
strong evidence that a local knowledge
spillover has occurred, especially since
we have accounted for the underlying geographic distribution of research
activity and hence other reasons why
inventors might be located together. In
fact, Jaffe and his coauthors find that
2 Q2 2014 Business Review

originating patentees are much more
likely to be from the same metropolitan area as citing patentees, compared
with a matched control patentee.
Despite this clever study design,
subsequent researchers have identified several limitations of this analysis.
First, there are the standard drawbacks
to using patent data: Not all inventions are patented, and some patents
do not represent valuable or worthwhile inventions. More recent papers
have tried to correct these problems
by, for example, measuring the quality
of patents based on patent renewals or
subsequent citations. Second, many
patent citations are actually added by
patent examiners, not inventors. Thus,
citations may not actually represent
true knowledge flows for inventors, but
rather noise introduced by the patent
office.5 Third, Peter Thompson and
Melanie Fox-Kean note that Jaffe’s
results are sensitive to the selection
of an appropriate control patent. By
varying how broadly the technology
classifications and dates of application
are specified for the sample of matched
control patents, Thompson and FoxKean found that imperfect matching
explained a significant part of Jaffe’s
original result.
A final issue, which Jaffe and
his coauthors acknowledged in their
original paper, is that many knowledge
inputs are not actually reflected in
citations. This is significant because
we might expect that geographic proximity is especially important for the
transfer of tacit, operational knowledge — that is, knowledge that is not
necessarily codified in a written patent
application. This is the kind of knowledge transmitted in hallways and over
coffee, rather than through literature
searches of previous work. Because
many more knowledge spillovers may

occur than are reflected in citations,
Jaffe and his coauthors suggest that
their findings may actually represent
a lower bound for the occurrence of
knowledge spillovers among inventors.

5
See the papers by Juan Alcácer and Michelle
Gittelman and by Peter Thompson (2006).

6
See the survey article by David Audretsch and
Maryann Feldman.

EVIDENCE FROM RESEARCH
AND NONPATENT DATA
Other papers have sidestepped
patents altogether.6 Bruce Weinberg
found that physicists who moved to
cities where Nobel laureates were
already working were more likely to
begin their own Nobel Prize-winning
work there. Gerald A. Carlino, Jake
K. Carr, Robert M. Hunt, and Tony E.
Smith have shown that research and
development labs are highly geographically concentrated, substantially more
so than the corresponding industry
concentration patterns. In my previous work, I showed that new activities
related to the implementation of new
knowledge are concentrated in metropolitan areas with highly educated
populations. Finally, Petra Moser has
shown localization among prize-winning inventors at World’s Fairs in the
19th century, although these patterns
weakened over time.
Of course, as with inventors, scientists may locate near each other for
reasons besides knowledge spillovers.
Thus, we cannot be certain whether
an increase in their productivity might
stem from knowledge spillovers from
nearby scientists or from some other
reason. Fabian Waldinger investigated
local knowledge spillovers among scientists in Germany. Waldinger relies
on evidence from the expulsion of
Jewish and certain other scientists from
Germany under the Nazis. Some university departments experienced many
expulsions, while other departments
had not employed Jewish scholars and
were therefore unaffected. If knowledge

www.philadelphiafed.org

spillovers are important, one might expect the productivity of the remaining
scientists in the affected departments
to decline following the expulsion of
their Jewish colleagues. Waldinger
finds that the publishing activity of
the scientists whose departments suffered losses did not decline compared
with that of other scientists. Thus, he
concludes that there is no evidence for
local knowledge spillovers among German scientists in this period.
EVIDENCE FROM PATENT
INTERFERENCES
In my work with Ina Ganguli
and Nick Reynolds, I have been using
patent interferences to try to provide
new evidence on the relevance of local knowledge spillovers for invention.
Patent interferences are especially
valuable for measuring local spillovers
of tacit or uncodified knowledge that is
missing in traditional patent studies.
Patent interferences are a unique
historical feature of the U.S. patent system. Until 2011, the United
States had a “first to invent” rule for
assigning priority of invention, versus
the “first to file” rule more common
in the rest of the world.7 When the
U.S. patent office received applications from multiple parties with
identical claims at roughly the same
time, it was obligated to investigate
the competing claims to determine
which party was entitled to patent
protection. This investigation, known
as a patent interference proceeding,
determined who had conceived of the
invention and reduced it to practice
first. Typically, the parties submitted
dated laboratory notebooks, testimony
by associates, and media reports as
evidence of first invention.
There are many famous examples

7
More details about the patent interference
proceedings can be found in Calvert and Sofocleous (1982), Cohen and Ishii (2006), de Simone
et al. (1963), and Kingston (2004).

www.philadelphiafed.org

of patent interferences in U.S. history, including Alexander Graham
Bell’s and Elisha Gray’s simultaneous
invention of the telephone. Because
Bell’s and Gray’s applications arrived at
the patent office on the same day and
contained nearly identical claims, an
interference proceeding was declared.
Eventually, Bell was determined to
have conceived of the idea and reduced it to practice first, and he was
awarded the patent.
Knowledge in common. Importantly for economists, patent interferences create a record of instances
when the same invention is created
by inventors working independently, a
phenomenon that is highly suggestive

interference is evidence of a knowledge
spillover among the inventors.
Several details about the interference process support the argument
that these proceedings are a good
measure of common, independent
knowledge inputs. First, interferences
were declared by a patent examiner
specializing in a particular technological area. Thus, interfering claims were
likely to be detected. (In some cases,
the examiner was alerted to a possible
interference by one of the applicants.
Note that an interference is different
from patent infringement, in which
the holder of an existing patent sues an
infringing party. Private parties cannot
sue for an interference.)

Patent interferences are especially valuable
for measuring local spillovers of tacit or
uncodified knowledge that is missing in
traditional patent studies.
of common knowledge inputs. In other
words, inventors involved in an interference are likely to have command
of similar knowledge. For example,
interfering inventors may have similar
backgrounds in chemistry, or they
may have similar knowledge of market
conditions. This is especially true
if certain inventions require highly
specific knowledge. For example, for
Bell and Gray to have both invented
the telephone contemporaneously,
they must have had similar knowledge about electrical conductivity and
the properties of various conductive
metals, as well as similar expectations
about market demand for a device that
transmitted voices in real time. For
Jon Merz and Michelle Henry, a patent interference is an indication that
“discovery has become ordinary.” That
is, its occurrence suggests that certain
knowledge is shared among several
inventors. In other words, a patent

Second, interferences involved
parties with roughly the same date of
patent application. An inventor who
delayed filing an application in order
to conceal an invention would lose the
priority contest. Thus, interferences
are less likely to reflect secrecy or other
legal strategies of the participants and
are more likely to reflect genuinely
independent, simultaneous inventions
versus infringements or patent “racing”
by inventors who believed that rival
applications were imminent.
Third, during the interference
proceedings, circumstances that suggested stealing, espionage, or collaborative invention typically led to dismissal with prejudice. In other words,
worker poaching and espionage that is
independent of shared knowledge are
unlikely explanations for the bulk of
cases of patent applications interfering
with each other. Note that recruiting other firms’ researchers or spying
Business Review Q2 2014 3

on them does seem to involve shared
knowledge inputs — a shared desire to
solve a common problem, for example.
In addition, patent judges had been
compelled by statute to rule against
applicants found to have stolen a competitor’s idea, deterring would-be spies
from pursuing an interference. In fact,
in our examination of case decisions,
no more than a small handful of judgments mention espionage as a relevant
factor in decisions.
Fourth, competing claims that
are similar but not identical did not
result in a completed interference
case. Fifth, the patent office verified
that interfering parties had independent financial interests (for example,
that they were not different branches
of a multinational conglomerate);
otherwise, the case was dismissed.
Thus, interferences are not the
result of knowledge sharing within
organizations. Finally, the separate
applications were required to have
been made roughly at the same time,
often within a year. Thus, copying subsequent to publication and
disclosure of an older patent are
unlikely to have occurred.
Interferences improve on traditional patent studies in a number
of ways. One, interferences involve
patents that are more valuable than
the average patent. Since an interference requires parties to actively contest
for priority, it is unlikely that inventors
would spend time or money in pursuit
of a worthless patent. Two, we have information on patent interferences over
a long period, from the 19th century to
2011. Three, patent interferences do
not rely on citations to prove common knowledge inputs and are thus
not subject to some of the weaknesses
noted earlier. Specifically, while patent
citations necessarily capture the spillover only of written, publicly available
knowledge, simultaneous invention is
evidence that some kind of spillover,
whether written or not, is likely to
4 Q2 2014 Business Review

have occurred. Thus, interferences
capture spillovers of tacit knowledge.
As I noted earlier, we might expect
that tacit knowledge is especially sensitive to geographic proximity. If so,
then we expect results on the localization of interferences to be stronger
than on the localization of citations.
We have constructed a database
of over 1,000 interference cases from
the early 1980s to 2011 from the U.S.
patent office. This database includes
the names of the inventors involved in
the interference, their patent and application numbers, and the date of the
interference. We match this information to a database of inventor locations
(based on their Zip codes) produced by
the Harvard Business School.
Testing geographic concentration. If local knowledge spillovers are
important, one possible test is to see
whether patent interferences, as measures of shared, possibly tacit knowl-

edge, are more likely to occur between
inventors who are located close to each
other versus those located farther apart.
The black line in Figure 1 shows this
pattern for only the interference cases
involving pairs of U.S.-based inventors.
The horizontal axis measures the distance in miles as the crow flies between
the observed locations of the two parties involved in a patent interference.
(Since a single patent application can
be made on behalf of multiple inventors, we measure the minimum distance
between inventors of the different parties to the interference.) The vertical
axis shows the percent of interfering
inventor pairs in our database that
are separated by at most the distance
indicated by the horizontal axis. Thus,
as we move to the right, we accumulate
our inventor pairs until 100 percent of
our pairs are within 4,258 miles — the
maximum distance observed between
two interfering U.S. inventors.

FIGURE 1
Interfering Inventors More Geographically
Concentrated
Cumulative percent of patent pairs
100
90
80
70
60
50
40
30

Interfering patents

20

Control patents

10
0
0

500

1,000

1,500

2,000

2,500

3,000

3,500

4,000

Distance between inventors (miles)

Sources: U.S. Patent and Trademark Office, Harvard University Patent Network Dataverse.

www.philadelphiafed.org

Despite the large possible range of
distances between inventors, the black
line shows that 20 percent of interfering inventors are within only 100 miles
of each other, and half of interfering
inventor pairs are within 540 miles of
each other. While this is a good starting point for showing that proximity
matters for shared knowledge inputs,
it still might be true that inventors
are located close to each other to take
advantage of some other factor. Similar
to the logic of the patent citation studies, we can compare the localization
of interferences with the localization
of noninterfering patents in the same
technology classification and year. In
that way, we can control for the underlying distribution of research activity
that doesn’t rely on common knowledge inputs, as interferences do.
For each pair of interfering patents, we selected up to 10 control patents. Our goal was to control for the
underlying geographic distribution of
inventive activity by selecting patents
that were similar to the interfering patents but not involved in the interference case. We selected control patents
based on two criteria. First, a control
patent had to share at least one of the
many possible three-digit technological classification codes assigned by the
patent office that the two interfering
patents had shared. Second, the control patent’s application date had to fall

www.philadelphiafed.org

between the application dates of the
two interfering patents. If no eligible
control patent was found, we then
expanded the selection window incrementally by 10 days before the earlier
interfering application and 10 days after the later interfering application until an eligible control patent was found.
Finally, we randomly chose one of the
two interfering patents to match with
the control patent. We then compared
the distance between the interfering
inventors with the distance between
the randomly selected interfering inventor and the control inventor.
The results. The gold line in Figure 1 shows our results for the proximity of interfering inventors to control
patent inventors. It represents the
expected distribution of distances between inventors working in technology
fields and time periods similar to those
of our sample of patent interferences,
but it is not conditioned on an interfering link between inventors.
Comparing the distribution of
distances between interfering inventors
with the control distribution of noninterfering inventors, it is clear that interfering inventors are more geographically concentrated. While 20 percent
of interfering inventors are within 100
miles of each other, less than 1 percent
of noninterfering inventors are within
100 miles of each other. And while
half of interfering inventors are within

540 miles of each other, the same is
true of less than 21 percent of noninterfering inventors.
Interfering inventors are especially
more likely to be geographically concentrated at small distances. For example, more than 3 percent of interfering U.S. inventors are in the same Zip
code, versus none of the noninterfering
inventor-control pairs. Eleven percent
of interfering inventors are within 15
miles of each other, compared with
less than 1 percent of noninterfering
inventors. These results are consistent
with a growing literature documenting that knowledge spillovers attenuate
rapidly with distance.
CONCLUSION
Although local knowledge spillovers are of central interest to economists, the evidence to date on their
existence is mixed. Patterns in our
data on patent interferences suggest
that inventors working independently
but using common knowledge inputs
are substantially more geographically
concentrated than other inventors
working in the same field and time
period who are not linked by common
knowledge inputs. These results suggest
that localized knowledge spillovers may
be especially salient for forms of tacit or
uncodified knowledge, which is difficult to observe using citations but more
likely detectable from interferences. BR

Business Review Q2 2014 5

REFERENCES
Alcácer, Juan, and Michelle Gittelman.
“Patent Citations as a Measure of Knowledge Flows: The Influence of Examiner Citations,” Review of Economics and Statistics,
88:4 (2006), pp. 774-779.

Cohen, Linda R., and Jun Ishii. “Competition, Innovation and Racing for Priority
at the U.S. Patent and Trademark Office,”
University of California, Irvine, Working
Paper (2006).

Audretsch, David B., and Maryann P.
Feldman. “Knowledge Spillovers and the
Geography of Innovation,” in J. Vernon
Henderson and Jacques-François Thisse,
eds., Handbook of Regional and Urban
Economics, 4. Amsterdam: Elsevier, 2004,
pp. 2,713-2,739.

De Simone, Daniel V., James B. Gambrell,
and Charles F. Gareau. “Characteristics of
Interference Practice,” Journal of the Patent
Office Society, 45 (1963), pp. 503-591.

Audretsch, David B., and Maryann P. Feldman. “R&D Spillovers and the Geography
of Innovation and Production,” American
Economic Review 86:3 (1996), pp. 630-640.
Azoulay, Pierre, Joshua Graff Zivin, and
Jialan Wang. “Superstar Extinction,” Quarterly Journal of Economics, 125:2 (2010),
pp. 549-589.
Borjas, George J., and Kirk B. Doran.
“The Collapse of the Soviet Union and
the Productivity of American Mathematicians,” Quarterly Journal of Economics,
Oxford University Press, 127:3 (2012),
pp. 1,143-1,203.
Calvert, Ian A., and Michael Sofocleous.
“Three Years of Interference Statistics,”
Journal of the Patent Office Society, 64:12
(1982), pp. 699-707.
Carlino, Gerald A., Satyajit Chatterjee,
and Robert M. Hunt. “Urban Density and
the Rate of Invention,” Journal of Urban
Economics, 61 (2007), pp. 389-419.
Carlino, Gerald A. “Knowledge Spillovers:
Cities’ Role in the New Economy,” Federal
Reserve Bank of Philadelphia Business Review (Fourth Quarter 2001).
Carlino, Gerald A., Jake K. Carr, Robert
M. Hunt, and Tony E. Smith. “The Agglomeration of R&D Labs,” Federal Reserve Bank of Philadelphia Working Paper
12-22 (2012).
Chapin, Seymour L. “Patent Interferences
and the History of Technology: A HighFlying Example,” Technology and Culture,
12:3 (1971), pp. 414-416.

Duranton, Gilles, and Diego Puga. “MicroFoundations of Urban Agglomeration
Economies,” in Vernon Henderson and
Jacques-Francois Thisse, eds., Handbook of
Regional and Urban Economics, 1. Amsterdam: Elsevier, 2004, pp. 2,063-2,117.
Duranton, Gilles, and Diego Puga. “Nursery Cities: Urban Diversity, Process Innovation, and the Life-Cycle of Products,”
American Economic Review, 91:5 (2001),
pp. 1,454-1,477.
Glaeser, Edward L. “Learning in Cities,”
Journal of Urban Economics, 46 (1999),
pp. 254-277.
Hunt, Robert M. “You Can Patent That?
Are Patents on Computer Programs and
Business Methods Good for the New Economy?” Federal Reserve Bank of Philadelphia Business Review (First Quarter 2001).
Jaffe, Adam B. “Technological Opportunity and Spillovers of R&D: Evidence from
Firms’ Patents, Profits, and Market Value,”
American Economic Review, 76:5 (1986),
pp. 984-1,001.
Jaffe, Adam B., Manuel Trajtenberg,
and Rebecca Henderson. “Geographic
Localization of Knowledge Spillovers as
Evidenced by Patent Citations,” Quarterly Journal of Economics, 108:3 (1993),
pp. 577-598.
Keller, Wolfgang. “Geographic Localization of International Technology Diffusion,” American Economic Review, 92:1
(2002) pp. 120-142.
Kingston, William R. “Light on Simultaneous Invention from U.S. Patent Office
‘Interference’ Records,” World Patent Information, 26:3 (2004), pp. 209-220.
Krugman, Paul. Geography and Trade.
Cambridge: MIT Press, 1991.

6 Q2 2014 Business Review

Lin, Jeffrey. “Technological Adaptation,
Cities, and New Work,” Review of Economics and Statistics, 93:2 (2011), pp. 554-574.
Lin, Jeffrey. “Urban Productivity Advantages from Job Search and Matching,”
Federal Reserve Bank of Philadlpelphia
Business Review (First Quarter 2011).
Lucas, Robert E. “On the Mechanics of
Economic Development,” Journal of Monetary Economics, 22 (1988), pp. 3-42.
Marshall, Alfred. Principles of Economics.
London: Macmillan, 1920.
Merz, Jon F., and Michelle R. Henry. “The
Prevalence of Patent Interferences in Gene
Technology,” Nature Biotechnology, 22:2
(2004), pp. 153-154.
Moser, Petra. “Did Patents Weaken the Localization of Innovations? Evidence from
World’s Fairs,” Journal of Economic History,
71:2 (2011), pp. 363-381.
Romer, Paul M. “Endogenous Technological Change,” Journal of Political Economy,
98:5 (1990), pp. 71-102.
Rosenthal, Stuart S., and William C.
Strange. “Evidence on the Nature and
Sources of Agglomeration Economies,” in
J. Vernon Henderson and Jacques-François
Thisse, eds., Handbook of Regional and
Urban Economics 4. Amsterdam: Elsevier,
2004, pp. 2,119-2,171.
Thompson, Peter. “Patent Citations and
the Geography of Knowledge Spillovers:
Evidence from Inventor- and ExaminerAdded Citations,” Review of Economics and
Statistics, 88:2 (May 2006), pp. 383-388.
Thompson, Peter, and Melanie Fox-Kean.
“Patent Citations and the Geography of
Knowledge Spillovers: A Reassessment,”
American Economic Review, 95:1 (March
2005), pp. 450-460.
Waldinger, Fabian. “Peer Effects in Science: Evidence from the Dismissal of
Scientists in Nazi Germany,” Review of
Economic Studies, 79 (2012), pp. 838-861.
Weinberg, Bruce A. “Geography and Innovation: Evidence from Nobel Laureate
Physicists,” mimeo (2007).

www.philadelphiafed.org

Shadow Banking and the Crisis of 2007-08
By Daniel Sanches

S

ome economists have noted that recessions accompanied
by banking crises tend to be deeper and more difficult
to recover from than other recessions — even those
associated with other types of financial crises. For instance,
the bursting of the dot.com bubble in 2001 was a very
important financial event that was not accompanied by a protracted
recession. The potential of banking crises to do lasting economic harm
led policymakers to adopt safeguards in the 1930s that have essentially
eliminated traditional banking panics in the U.S. Although the Great
Recession of 2007-09 was associated with a protracted financial market
disruption — and the failures of some large banks like Washington
Mutual and IndyMac — we did not observe widespread withdrawals
from commercial banks, as in a traditional banking crisis. However,
economists Gary Gorton and Andrew Metrick show that it can be
viewed as a banking crisis that originated in the shadow banking system.
In the last 30 years, institutions very similar in function to traditional
banks have grown outside regulatory oversight. One lesson of the
financial crisis is that these institutions are as vulnerable to panics as
traditional banks because they are subject to similar risks.

ECONOMIC FALLOUT
OF BANKING CRISES
Banking crises can harm the
economy. Financial crises are usually
associated with bad economic outcomes
— recessions. One particular kind of
financial crisis to which economists
have devoted a lot of attention is the
type that originates in the banking
sector. A banking crisis is a widespread
withdrawal of funds from depository
institutions — that is, a run on the liabilities of a large number of banks.1 Like
other financial crises, banking crises
are usually associated with economic

downturns, and there is evidence that
banking crises often worsen economic
downturns as weaknesses at the banks
spill over into financial problems for
households and firms. When financial
events affect consumption and investment decisions by households and
nonfinancial firms, economists say that
they have real effects.
Many researchers have provided
evidence that banking crises can make
1
A common form of bank liability is the
demand deposit contract — a typical checking
account that most people have at a commercial
bank.

Daniel Sanches is a senior economist at the Federal Reserve Bank
of Philadelphia. The views expressed in this article are not necessarily
those of the Federal Reserve. This article and other Philadelphia
Fed reports and research are available at www.philadelphiafed.org/
research-and-data/publications.

www.philadelphiafed.org

economic contractions more severe
and more protracted, with various
studies emphasizing different channels.
In a highly influential book providing a systematic account of banking crises in the United States from
1867 to 1960, Milton Friedman and
Anna Schwartz identify the reduction
in the wealth of bank shareholders
and the decline in the money supply
that usually follow a business downturn accompanied by a banking crisis
as the main causes of a further drop
in real economic activity. A decline
in the money supply has real effects
because households and firms need
money to pay for their purchases.
Thus, a decline in the money supply
leads to a decline in transactions and
real economic activity.2
Looking at the banking crises of
the 1930s, Ben Bernanke identified
the increased cost of intermediation
services — the costs that banks incur
when assessing the creditworthiness of
borrowers — following the recurrent
banking crises of the early 1930s as
causing a significant reduction in the
flow of funds from lenders to borrowers through the banking system. This
constricted flow of credit impeded the
real economy’s recovery from the Great
Depression.
Michael Bordo, Barry Eichengreen, Daniela Klingebiel, and Maria
Soledad Martinez-Peria provide crosscountry evidence of the real effects
of banking crises over a period of 120
years. They also find that recessions
that are accompanied by banking crises are more severe than those that are

For more on this subject, see my 2012 Business
Review article, “The Optimum Quantity of
Money.”

2

Business Review Q2 2014 7

not. Moreover, they provide evidence
that banking crises plague advanced
and developing economies equally,
confirming the view that a banking
crisis is not just a concern for countries
at low levels of economic development.
Anatomy of a banking crisis.
Although every banking crisis is different, those that occurred up to and including the Great Depression follow a
similar pattern. Let me briefly describe
the typical sequence of events that
leads to a banking crisis.
Bad news arrives. Usually at the
peak of an economic expansion, bad
news about the quality of the assets
held by a group of banks (or a major
bank) leads to larger withdrawals than
usual. For instance, a failed attempt
by the Knickerbocker Trust Company
to corner the copper market and the
subsequent decision of a major bank
to no longer clear checks issued by the
Knickerbocker triggered a run on the
Knickerbocker on October 18, 1907,
sparking the Panic of 1907. The fact
that it was one of the largest depository
institutions in New York contributed
to the public’s perception that other
banks could also be in distress.
Banks sell assets to meet the increase in withdrawals. To meet the
higher demand for cash, a bank initially draws down its cash reserves. But
its reserves may not be enough if the
withdrawal process quickly intensifies.
The bank can also sell some of its assets to cover withdrawals.
Selling assets causes asset prices to
fall. If many banks are trying to sell
assets at the same time, the assets can
be sold only at a large discount. Think
of what would happen if four neighbors
on your block put their houses up for
sale on the same day as you did. All
else equal, you would have to lower
your price to get anyone to buy. Financial asset markets work the same way.
For easily marketable fixed-income
assets such as Treasury securities or
certain corporate bonds traded in large
8 Q2 2014 Business Review

markets, buyers can still be found by
selling at a discount. However, a large
fraction of a bank’s assets consists
of mortgages and commercial and
industrial loans made to households
and firms whose creditworthiness is
unknown to the wider market, which
means the bank would probably find
few if any buyers. And anyone willing
to purchase the loan would demand a
substantial discount to compensate for
the lack of information about the borrower’s creditworthiness. Thus, selling
assets on short notice may be extremely costly for a bank.
Depositors begin a run on healthy
banks. Banks facing large withdrawals
may borrow in the interbank market,
where banks routinely borrow reserves

Banks suspend convertibility. The
final step comes as banks react to
widespread withdrawals. One way to
stop the drain on funds is to temporarily suspend the convertibility of
deposits into cash — banks may simply
lock their doors — in an attempt to
preserve capital until depositors calm
down and things get back to normal.
Strictly speaking, this is a breach of
the demand deposit contract.4
This description of a typical banking crisis clearly reveals why banks
are fragile: They fund illiquid assets
with deposits that can be withdrawn
at will. Economists usually refer to this
practice as maturity transformation. It
is important to mention that this role
played by banks has a value for society.

Although every banking crisis is different,
those that occurred up to and including the
Great Depression follow a similar pattern.
from each other. But if banks want to
borrow more reserves than usual, they
must pay a higher interest rate to the
lending bank. Larger discounts in asset
markets and higher interest rates in
interbank markets are usually signs of
financial strain. If widespread distrust
of banks causes depositors to withdraw
their funds even from healthy banks,
a line is crossed. The number of banks
that want to sell assets increases, resulting in even steeper discounts, and the
number of banks that want to borrow
in the interbank market also increases,
making it harder for each borrower
bank to obtain a loan. As this process
intensifies, we have a full-scale panic.3

3
We can think of these withdrawals as a way
for depositors to monitor their banks. That is,
by withdrawing their money, depositors are
checking whether the bank is healthy enough
to pay. This might explain why people decide to
withdraw their funds even from banks initially
viewed as safe and sound.

People have a preference for holding
highly liquid assets — assets that are
easy to sell without taking a loss —
but the most profitable investments
take a long time to pay off. Banks offer
demand deposit contracts that give
people ready access to their funds and
a higher rate of return than they would
get by holding liquid assets directly.
Banks are able to offer a higher rate of
return to depositors because they pool
resources in such a way that permits
them to invest a significant fraction of their assets in higher-yielding,
long-term projects such as mortgages
and other types of long-term loans.
Normally, funding illiquid assets with
short-term liabilities works fine. But
when depositors begin to worry about
losses, a bank run may ensue.
In the second half of the 19th century, the
decision to suspend convertibility was usually
coordinated by private bank associations.

4

www.philadelphiafed.org

U.S. bank runs essentially disappeared in the 1930s. The introduction of federal deposit insurance in
1933 with the creation of the Federal Deposit Insurance Corporation
(FDIC) ended the banking crises that
had been recurrent events in the U.S.
even before the Great Depression.
The government’s deposit guarantees
largely relieved depositors of the need
to constantly monitor the health of
banks. In turn, the government has
undertaken the monitoring of banks
through regulation and supervision.
But regulations are not costless. FDIC
premiums, capital requirements, and
regulatory restrictions on bank portfolios increase banks’ costs. These costs
are informally referred to as regulatory
taxes. And banks, like any other firm,
have a strong incentive to avoid taxes.
THE RISE OF SHADOW
BANKING
The Great Recession in the U.S.
was associated with a severe financial
crisis, but we did not observe people
rushing to their banks to withdraw
their deposits. However, a closer look
suggests that the crisis was not very
different from a typical banking crisis,
except that it was triggered outside the
traditional banking sector. According
to Gary Gorton and Andrew Metrick,
the financial crisis can be viewed as a
banking crisis that originated in the
shadow banking system.5
The shadow banking system is a
set of institutions that carry out functions very similar to those of traditional banks but that are largely unregulated. They perform the same kind
of maturity transformation traditionally performed by commercial banks.
Thus, the shadow banking system,
despite its somewhat unwholesomesounding name, provides a useful
service to society. This is to say that

shadow banking is not necessarily a
bad thing. The problem is that, under
certain circumstances, these financial
institutions can become fragile — that
is, subject to panics.
An important fact about the shadow banking system is that it has grown
significantly in the last 30 years. For
instance, Gorton and Metrick estimate
that just before the financial crisis of
2007-08, the assets of the shadow banking system were at least as large as the
assets of commercial banks.6 Another
important fact about the shadow banking system is that it has grown outside
the oversight of regulators. Why did
this happen? As banking and finance
in general have expanded in recent decades, part of that growth has occurred
in the shadow system, largely to avoid
the costs associated with regulation.7
As I will now explain, the shadow
banking system works pretty much
like a typical commercial bank even
though the parties involved in the
transactions are not the bankers and
depositors that we typically have in
mind. For the most part, I will follow
Gorton and Metrick and focus on the
market for repurchase agreements (or
repos) as the main cause of the panic
in the shadow banking system and one
of the centers of the financial crisis.
But the shadow banking system also
includes other markets and institutions
such as asset-backed commercial paper
in which the same basic structure
(risky, illiquid assets funded by shortterm liabilities) recurs.
The repo market. The repo market is a market for short-term, mainly
overnight, collateralized loans. To understand why repos work pretty much

According to Gorton and Metrick, this is
probably an underestimate because this comparison involves the assets of only a fraction of
the shadow banking system.
6

For more on the rise of shadow banking,
see the review by Tobias Adrian and Adam
Ashcraft.

7

5

See also chapter 2 in Gary Gorton’s 2010 book.

www.philadelphiafed.org

like banking and to see why the repo
crisis was actually a banking crisis, it is
necessary to look at how repo transactions work.
Let me start by identifying the
“depositors,” the repo lenders. These
are largely institutional investors such
as pension funds and large corporations that need some place to invest
large amounts of money for short periods. They also want to obtain higher
yields than those offered by regulated
commercial banks. Most important,
these institutional investors want their
funds to be safe.8
One alternative is the repo market. A firm can make an overnight
loan to a borrower. To make the loan
safe, the firm receives collateral usually in the form of government bonds,
which are liquid and fluctuate little
in value over short periods. If the borrower is unable to return the funds, the
lending firm will simply seize the collateral. Provided that the value of the
underlying collateral does not change
significantly over short periods, a repo
transaction is safe for the repo lender.
Like a bank depositor, the repo
lender has ready access to its money
and has the opportunity to reallocate
its funds toward some other use on a
daily basis. Thus, a repo transaction
offers the firm both the convenience of
having ready access to its funds and a
level of safety not much different from
that of a federally insured demand
deposit. Until 2011, large commercial
depositors could not receive interest on
their short-term deposits, another motivation for them to seek an alternative
place to park their funds.9 When the

8
It is also important to mention that the
amounts these institutional investors wish to
deposit are typically larger than the maximum
amount insured by the FDIC.
9
As part of the Dodd-Frank Act, the Federal
Reserve Board in July 2011 repealed Regulation
Q, which had prohibited banks from paying
interest on corporate checking accounts.

Business Review Q2 2014 9

repo borrower repurchases the security
from the repo lender, he or she also
pays interest to the lender.
As should be clear by now, the
“banker” in the repo transaction is the
repo borrower, which typically is an
investment bank or the broker-dealer
arm of a large bank holding company.
These institutions use the funds they
borrow in the repo market to finance a
wide range of activities, some of them
quite risky. As long as the repo is collateralized by a Treasury security, it is
not fragile in the same sense as traditional banking because the asset that
collateralizes the repo is highly liquid
and can be easily sold. If the repo borrower can’t repay on time, the repo
lender can simply take the collateral
and sell it for cash.
This is basically how the shadow
banking system works. Depositors
(institutional investors and large corporations) need a place to park liquid
funds that provides them with ready
access to their money, pays an interest rate higher than that offered by
traditional banks, and spares them the
expense and hassle of managing their
own cash.10 Bankers (investment banks
and broker-dealer firms) are willing
to provide such a product in the form
of repo transactions. Finally, safe collateral such as U.S. Treasury bonds are
essential to make this financial transaction work.
The growth of the repo market
increased the demand for collateral.
The growth of the repo market prior
to the financial crisis of 2007-08 was
extraordinary. The volume of repo

As Robert Lucas puts it: “In a monetary economy, it is in everyone’s private interest to try to
get someone else to hold non-interest-bearing
cash and reserves. But someone has to hold it
all, so all of these efforts must simply cancel out.
All of us spend several hours per year in this
effort, and we employ thousands of talented and
highly trained people to help us. These personhours are simply thrown away, wasted on a task
that should not have to be performed at all.”

10

10 Q2 2014 Business Review

transactions reported by primary dealers (those who trade directly with the
Federal Reserve System) had grown
from roughly $2 trillion in 1997 to
$7 trillion in 2008. This estimate, of
course, leaves out unreported transactions. Gorton and Metrick estimate
that the overall size of the repo market
just before the financial crisis was
roughly the same as the size of the
traditional banking sector as measured
by total assets.11
As we have seen, Treasury securities play an important role in the functioning of the shadow banking system.
However, repo markets are not the
only source of demand for Treasury securities. They are also used as collateral in derivative markets and settlement
systems. Furthermore, many foreign
governments, especially the central
banks of developing countries such as
China, demand Treasury securities because they are safe and highly liquid.12
For instance, in 2005 only 48.6 percent
of total U.S. debt was privately held,
according to the Federal Reserve Bank
of San Francisco. About a third of that
privately held debt was held in reserve
by foreign central banks, which means
that only about a third of total U.S.
debt (or $2.6 trillion) was available for
private transactions.
Unlike for other goods and services, higher demand for Treasury
securities doesn’t automatically provide
an incentive to increase supply. The
supply of government bonds is determined by government borrowing, a
direct consequence of fiscal policy. For
instance, the decision to reduce the
fiscal deficit in the U.S. in the 1990s
and early 2000s may have contributed
to a shortage of government bonds

available for repo transactions.
One piece of indirect evidence
that government bonds were in short
supply is the practice in financial markets known as rehypothecation, which
simply means that traders can use the
same collateral to secure more than
one transaction. To the extent that
this practice had become widespread
before the crisis of 2007-08, traders
may have had an incentive to develop
other methods to conduct a growing
number of transactions with a limited
amount of good collateral.13
Mortgage-backed securities
helped satisfy the demand for collateral. The solution to the shortage of
good collateral was found in another
form of financial innovation that had
evolved significantly since the 1980s:
securitization. Commercial banks
make many loans to consumers and
firms. Instead of holding these loans
on its own balance sheet, a bank can
sell them to a shell company the bank
creates and manages for this purpose, called a special purpose vehicle
(SPV). The SPV funds the acquisition
of these assets (mortgages, car loans,
credit card receivables, etc.) by issuing
asset-backed securities (ABS) that, as
the name implies, are backed by the
loans the SPV holds and that become
the SPV’s liabilities when it sells them
to investors in the capital markets.14
Figure 1 shows how commercial banks
fund loans through securitization.
Most important, this organizational form allows financial institutions to
increase the scale of their overall operations without increasing their balance
sheets, which would require them to
increase their regulatory capital. Thus,
setting up an SPV is a way of avoiding

11
They also cite a range of estimates by other
economists of the same order of magnitude.

13

Foreign demand for Treasury securities
increased significantly in the 1990s and 2000s,
the flip side of the large trade surpluses run by
China and some other developing countries.

14
See the chapter by Gary Gorton and Nicholas
S. Souleles.

12

For more on the role of rehypothecation,
see the 2011 Business Review article by Cyril
Monnet.

www.philadelphiafed.org

FIGURE 1
Securitization and Shadow Banking
STEP 1:

STEP 2:

STEP 3:

Bank originates
loans.

Bank creates SPV
to purchase loan portfolio.

SPV issues ABS
to fund portfolio purchase.

CASH

BORROWERS

capital requirements, which increases a
financial institution’s overall degree of
leverage by raising its total assets relative to its capital.15
If carefully chosen, a portfolio of
loans backing the ABS can be safe
and predictable. Thus, by making it
possible to bundle individual loans and
sell claims on the loan portfolio on
the market, this form of financial innovation offers an alternative to using
deposits to fund banks’ illiquid assets.
When carefully executed, securitization is extremely valuable for both
banks and investors.
The housing boom in the U.S. in
the 2000s was financed in this way.
The large increase in the number and
size of mortgage loans created a large
supply of a particular type of ABS
called mortgage-backed securities
(MBS). As the name suggests, MBS
are ABS that bundle mortgages. Given
the growth of the repo market and the
relative scarcity of government bonds,
the use of ABS as collateral in the
repo market seemed to be a reasonable
solution to the shortage of collateral.
Gorton and Metrick have argued that
the use of ABS as collateral in the

In this article, I emphasize avoiding regulatory taxes as a motivation for securitization.
See Ronel Elul’s article for an account of the
efficiency benefits of securitization.

15

www.philadelphiafed.org

SPV

BANK
LOANS

CASH

CASH

LOANS

repo market had increased significantly
prior to the financial crisis. As I discuss below, this is still a controversial
claim. Despite a wealth of anecdotal
evidence, we have no precise estimates
of the share of repo transactions that
used ABS as collateral.
CRISIS IN THE SHADOW
BANKING SYSTEM
ABS as repo collateral created
the conditions for a banking panic.
As long as the repo was collateralized
by Treasury securities, lenders (depositors) didn’t have to worry about the
borrower’s risk of default or about the
value of the underlying collateral. But
this changed when the repo was collateralized by ABS.
In 2007, house prices in the U.S.
started to decline, raising concerns that
homeowners could start defaulting on
their mortgages in large numbers. In
turn, lenders with repo collateralized by
MBS started worrying about potential
losses. What was the reason for their
concern? After all, as I have argued,
when carefully executed, securitization
can generate a safe asset for investors,
and indeed many MBS were built to be
nearly riskless under normal conditions.
Usually, ABS are designed to be
safe. ABS reduce credit risk in two
ways: diversification and overcollateralization. For instance, pooling mort-

INVESTORS
ABS

gages that had been originated in cities
all over the U.S. is one way to create
diversification. Under normal circumstances, large numbers of homeowners
in all regions of the country are very
unlikely to default on their mortgages
at the same time. Overcollateralization
simply involves pooling enough mortgages to guarantee that it can generate
enough cash flow to make the promised payments to investors even if some
of the borrowers default. The amount
of overcollateralization required to
make an MBS safe usually depends on
certain fundamental market indicators,
including the trend in house prices.
Significantly, the statistical models
used to design, price, and provide credit ratings for MBS estimated default
rates based on data collected during
periods of generally rising house prices
and during periods when housing price
declines were localized.16
Bad news arrived. Now consider
a scenario in which investors expect
house prices to rise and, contrary to
their expectations, house prices begin
to fall, and keep falling. That is what
happened in the U.S. in 2007. When
house prices fell for several consecu-

16
Indeed, Christopher Foote, Kristopher
Gerardi, and Paul Willen have documented that
people had overly optimistic beliefs about house
prices.

Business Review Q2 2014 11

tive months, an increasing number
of investors believed that the average rate of default on any given pool
of mortgages was going to rise. Their
fears became more concrete when in
the summer of 2007, two hedge funds
sponsored by Bear Stearns that had
invested heavily in subprime mortgages
filed for bankruptcy and BNP Paribas
suspended withdrawals from three
money market mutual funds that were
exposed to subprime mortgages. An
important indicator of their fears was
that the ABX index, a measure of the
risk of default on subprime MBS, began to rise. This raised concerns that
many SPVs were not holding enough
collateral to generate sufficient cash
flow to make good on the promised
payments to investors.
Another reason to have doubts
about the true value of MBS was that
many investors did not know where
the risks were concentrated. Although
many MBS were wisely built to be
nearly riskless, several classes of MBS
contained a disproportionate fraction
of mortgages that had been extended
to people of dubious creditworthiness.
And the risk of these subprime mortgages was particularly sensitive to the
decline in housing prices.
Repo lenders ran on repo borrowers,
including healthy borrowers. A depositor
with serious doubts about the underlying value of the collateral can do two
things: either ask for more collateral
or simply not renew the repo. Both actions can be interpreted as a decision
to withdraw funds from the shadow
banking system, much like the decision bank depositors make to withdraw
funds from their bank when they believe they might not be able to get all
their money out.
Repo lenders initially asked for
more collateral, but ultimately they
simply refused to renew their loans. In
other words, the repo market froze.17
Because investors could not tell safe
MBS from risky MBS in most cases,
12 Q2 2014 Business Review

they withdrew their funds even from
shadow banks that probably had safe
MBS to secure repos. This problem
was severe enough to turn the initial
panic into a systemic event — a banking crisis.
Thus, the financial crisis was not
very different from the banking crises
of old. Investors in the repo market behaved pretty much like bank depositors
did during U.S. banking crises before
1933. And the outcome was certainly
very similar. The initial banking crisis
spread to other financial markets, and
several financial firms either failed or
had to be rescued by the federal government to prevent further failures.
A caveat. Gorton and Metrick’s
explanation for the events that sparked
the 2007-08 financial crisis depends on
the claim that the fraction of the repo
market that used ABS as collateral was
large enough to generate a systemic
event. But this claim has been a source
of controversy among financial economists. For instance, Arvind Krishnamurthy, Dmitry Orlov, and Stefan
Nagel have argued that a relatively
small share of repo transactions in
which money market mutual funds and
securities lenders were the repo lenders
was collateralized by ABS prior to the
2007-08 crisis. However, these authors
focus on a relatively small segment of
the repo market, the triparty repo market, while Gorton and Metrick study
the larger bilateral repo market, for
which there is as yet no direct evidence
about the collateral used in transactions.18 Furthermore, Krishnamurthy
and coauthors note that while the
share of the transactions collateralized

17
See also Yaron Leitner’s article explaining
why markets freeze. See also Benjamin Lester’s
article for a discussion of regulatory interventions in response to market freezes.

Triparty repo transactions take place between
two counterparties intermediated by a dealer
bank. Bilateral repo transactions occur between
two counterparties without an intermediary.

18

by ABS was modest, such transactions
were more concentrated among a small
number of large banks that experienced significant problems. So focusing on average shares may be misleading. Nonetheless, the details of Gorton
and Metrick’s account of developments in the repo market will remain a
source of controversy until researchers
can collect more complete data. Moreover, some evidence suggests that the
financial crisis was actually triggered
in another part of the shadow banking
system. See the accompanying discussion, Crisis in the ABCP Market.
SHADOW BANKING PANIC
MAY HAVE DEEPENED
RECESSION
It is still too early to fully disentangle the relative importance of the
various factors that led to a particularly deep recession and a particularly
slow recovery. But like many earlier
recessions associated with banking
crises, the crisis in the shadow banking
system may have played a significant
role in the depth of the downturn and
the slow recovery.
The crisis in the shadow banking system has significantly reduced
the ability of commercial banks to
originate and renew loans, creating
ongoing problems for households and
firms that rely on bank loans. Some
economists have even argued that the
effects of the collapse can persist for an
extended period. For instance, Viral
Acharya has argued that traditional
lenders cannot easily fill the role that
shadow banks had played in providing
credit to the economy. This void has
certainly contributed to the delay in
restoring the flow of credit to a volume
consistent with that of a recovery from
a typical recession that had not been
accompanied by a banking crisis.
The shadow banking system has
not fully recovered from the financial
crisis. Even though it has continued to
operate with government support, it is
www.philadelphiafed.org

unclear whether the volume of operations will return to that observed prior
to the crisis anytime soon, or whether
it should. Since the crisis, the government-sponsored enterprises Fannie Mae
and Freddie Mac have carried out nearly all securitizations in housing markets.
At this point it is unclear whether the
private sector will ever play the same

role in the creation of securitized assets
that it had before the crisis.
CONCLUSION
One lesson of the financial crisis is
that institutions quite similar to banks
tend to rise up outside of regulatory
purview. This is an important matter
because this shadow banking system

is fragile and subject to panics. And
banking panics — regardless of where
they occur — have pernicious economic repercussions. This potential for
economic harm had led some economists before the crisis to propose tighter
regulation of the shadow banking
system. In the aftermath, policymakers
were working to write new rules. BR

Crisis in the ABCP Market

S

ome economists have argued that problems in another segment of the shadow banking system can be
identified as the prime cause of the 2007-08 financial crisis. For example, in his discussion of Gorton
and Metrick’s account of the crisis, Andrei Shleifer has provided evidence that the contraction in the
asset-backed commercial paper (ABCP) market happened before the contraction in the repo market.
Thus, he suggests that problems in the ABCP market may have triggered the financial crisis.
Commercial paper is a short-term debt instrument that both financial and nonfinancial firms use
to finance ongoing operations. Financial firms issue commercial paper to fund a wide array of activities, including the
purchase of long-term securities such as MBS. One form of funding through the issuance of commercial paper that
has increased significantly in the last 20 years is ABCP. A financial firm can set up an SPV to purchase a portfolio of
securities by issuing commercial paper on the capital markets. ABCP maturities can vary from one day (as in a typical
repo transaction) to 90 days. The typical maturity of ABCP is 30 days. Again, we have something that looks like a
bank, but it operates outside the regulatory system.
The main investors in ABCP are money market mutual funds. Similar to the investors in the repo market, money
market mutual funds also need a convenient place to invest some of their resources for short periods. These investors also want their investments to be safe and to yield an attractive return. Provided that the assets securing ABCP
are of sufficiently high quality, such an investment vehicle is fairly safe, at least under normal market conditions. The
short-term duration of ABCP gives investors an opportunity to “withdraw” their funds in case they decide to invest
elsewhere or in case they have doubts about the quality of the assets securing ABCP.
The issuers of ABCP are SPVs that are sponsored by large financial institutions, including traditional commercial banks.a The SPVs allow these institutions to fund a wide array of securities at any moment. The short duration of
ABCP means that an SPV has to roll over its debt every time an ABCP matures.
Many SPVs used the proceeds from the sale of ABCP to invest in MBS (i.e., the collateral backing ABCP were
MBS). As we have seen, the perception of MBS as a safe debt instrument can suddenly change once the trend in
house prices becomes clearly downward. Starting in the summer of 2007, many investors stopped refinancing maturing
ABCP because of potential exposure to subprime mortgages via MBS. A full-scale panic ensued as the spread on overnight ABCP over the federal funds interest rate (the rate of interest on unsecured loans in the interbank market in the
U.S.) increased from 10 basis points to 150 basis points. The outstanding amount of ABCP shrank steadily after the
summer of 2007, despite several interventions by the Federal Reserve System in the form of liquidity facilities, offering
short-term credit to banks to refinance maturing ABCP.
The ABCP market also provides another example of financial transactions carried out outside the oversight of
regulators that are very similar to traditional banking. Thus, a closer look at the crisis in the ABCP market has also
demonstrated that it was not very different from previous banking crises.
Perhaps the most balanced view is that while the financial crisis began in the shadow banking system, it had
many epicenters. Furthermore, the structural similarities among many of the institutions in the shadow banking system — illiquid assets funded by short-term liabilities — and the trigger for the crisis — the decline in housing prices
— tell much the same story.

A sponsor financial institution usually provides credit guarantees to the SPV. For a detailed description of the ABCP market, see the paper by
Marcin Kacperczyk and Philipp Schnabl.

a

www.philadelphiafed.org

Business Review Q2 2014 13

REFERENCES

Acharya, Viral. “Understanding Financial Crises: Theory and Evidence from
the Crisis of 2007-2008,” NBER Reporter
(April 2013).
Adrian, Tobias, and Adam B. Ashcraft.
“Shadow Banking: A Review of the Literature,” Federal Reserve Bank of New York
Staff Report 580 (October 2012).
Bernanke, Ben. “Nonmonetary Effects of
the Financial Crisis in the Propagation of
the Great Depression,” American Economic
Review, 73 (1983), pp. 257-276.
Bordo, Michael, Barry Eichengreen,
Daniela Klingebiel, and Maria Soledad
Martinez-Peria. “Is the Crisis Problem
Growing More Severe?” Economic Policy,
16 (2001), pp. 51-82.

Foote, Christopher, Kristopher Gerardi,
and Paul Willen. “Why Did So Many People Make So Many Ex Post Bad Decisions?
The Causes of the Foreclosure Crisis,” Public Policy Discussion Papers 12:2, Federal
Reserve Bank of Boston (2012).
Friedman, Milton, and Anna Schwartz. A
Monetary History of the United States, 18671960. Princeton University Press, 1963.
Gorton, Gary B. Slapped by the Invisible
Hand: The Panic of 2007. New York: Oxford University Press, 2010.
Gorton, Gary B., and Andrew Metrick.
“Securitized Banking and the Run on
Repo,” Journal of Financial Economics, 104
(2012), pp. 425-451.

Elul, Ronel. “The Economics of Asset
Securitization,” Federal Reserve Bank
of Philadelphia Business Review (Third
Quarter 2005).

Gorton, Gary B., and Nicholas S. Souleles.
“Special Purpose Vehicles and Securitization,” in Rene Stulz and Mark Carey, eds.,
The Risks of Financial Institutions. University of Chicago Press, 2006.

Federal Reserve Bank of San Francisco,
“Who Are the Largest Holders of U.S.
Public Debt?” http://www.frbsf.org/education/publications/doctor-econ/2005/july/
public-national-debt.

Kacperczyk, Marcin, and Philipp Schnabl.
“When Safe Proved Risky: Commercial
Paper During the Financial Crisis of 20072009,” Journal of Economic Perspectives, 24
(2010), pp. 29-50.

14 Q2 2014 Business Review

Krishnamurthy, Arvind, Stefan Nagel, and
Dmitry Orlov. “Sizing Up Repo,” NBER
Working Paper 17768 (2012).
Leitner, Yaron. “Why Do Markets Freeze?”
Federal Reserve Bank of Philadelphia
Business Review (Second Quarter 2011).
Lester, Benjamin. “Breaking the Ice:
Government Interventions in Frozen
Markets,” Federal Reserve Bank of
Philadelphia Business Review (Fourth
Quarter 2013).
Lucas, Robert E., Jr. “Inflation and Welfare,” Econometrica, 68 (2000), pp. 247-274.
Monnet, Cyril. “Rehypothecation,” Federal
Reserve Bank of Philadelphia Business
Review (Fourth Quarter 2011).
Sanches, Daniel. “The Optimum Quantity
of Money,” Federal Reserve Bank of Philadelphia Business Review (Fourth Quarter
2012).
Schleifer, Andrei. “Comments on Gorton
and Metrick: Regulating the Shadow
Banking System,” Brookings Papers on
Economic Activity, 2 (2010), pp. 298-303.

www.philadelphiafed.org

Forecast Disagreement
in the Survey of Professional Forecasters
By Keith Sill

M

any people engaged in activities related to business,
financial markets, and policymaking closely follow
economic forecasts. Our interest in forecasts stems from
the fact that, to an important degree, the decisions we
make today are influenced by our expectations about the
economy. Accurate forecasts lead to better decision-making and more
efficient use of economic resources, and so there is a clear benefit to
identifying good forecasts.

An important resource for evaluating the predictions and performance
of professional forecasters is the Survey
of Professional Forecasters, conducted
by the Philadelphia Fed Research
Department’s Real-Time Data Research Center. The SPF is a quarterly
survey that asks a panel of professional
forecasters about their projections for a
range of economic variables, including output growth, unemployment,
inflation, and interest rates. When
examining the SPF data, it becomes
clear that professional forecasters
have wide-ranging views about the
future evolution of the economy. This
is perhaps a bit surprising, since the
statistical methods that underlie good
forecasting models are well known,
and professional forecasters by and
large have access to the same data on
the economy’s past performance.
With forecasters having similar
tools and data to work with, why do

we observe this wide dispersion in
their projections?1 Are expectations
wide-ranging because of differences in
models and methods used to make the
forecasts? Or does the wide disagreement stem from how different forecasters process and analyze information
and then use it as an input into their
forecast-generation process? To design
and implement effective economic policies, it is important to understand how
expectations are formed. One way to
do so is to study forecast disagreement.
In this article we will examine some
features of the forecasts that underlie the SPF and discuss what theories
and evidence tell us about forecaster
behavior and how expectations about
the economy are formed and evolve
over time.
For a discussion on measuring the accuracy of
the survey’s forecasts, which is beyond the scope
of this article, see Stark (2010).

1

Keith Sill is a vice president and director of the Real-Time Data
Research Center at the Federal Reserve Bank of Philadelphia.
The views expressed in this article are not necessarily those of the
Federal Reserve. This article and other Philadelphia Fed reports and
research are available at www.philadelphiafed.org/research-and-data/
publications.

www.philadelphiafed.org

THE SURVEY’S DESIGN
The SPF is the oldest quarterly
survey of macroeconomic forecasts in
the United States, having been initiated in 1968 under the leadership of
Victor Zarnowitz at the American Statistical Association and the National
Bureau of Economic Research. After
conducting what was then known as
the ASA-NBER Quarterly Economic
Outlook Survey for 22 years, the ASANBER turned the survey over to the
Federal Reserve Bank of Philadelphia
in 1990, at which time it was renamed
the Survey of Professional Forecasters. The Philadelphia Fed’s Real-Time
Data Research Center now conducts
the SPF. A panel of professional
forecasters (there are usually around
45 respondents per survey) is asked to
give projections for a range of major
macroeconomic variables over various
time horizons.2
To maintain high quality, the SPF
screens its participants. Most have had
advanced training in economic theory
and statistics and use statistical models
to generate their projections. To keep
the integrity of the survey high, participation is limited to those employed
by firms or paid by clients to generate
forecasts now or in the past. Because of
these criteria and the types of individuals who participate in the SPF, we

2
The survey results are released to the public
free of charge at 10 a.m. on the second or third
Friday of the second month of each quarter.
The release schedule and the results of current
and past surveys, as well as the underlying data,
including anonymized individual forecaster projections, are available at http://philadelphiafed.
org/research-and-data/real-time-center/surveyof-professional-forecasters.

Business Review Q2 2014 15

can surmise that fairly sophisticated
models and statistical methods underlie their projections.
SPF participants use their models
to forecast quarterly values of major
macroeconomic variables for up to
five quarters, including the current
quarter, and annual projections up to
three years ahead. In addition, the
SPF asks for long-term annual averages
for headline and core inflation, real
GDP growth, productivity growth, and
stock and bond returns. A somewhat
unusual feature of the survey is that,
instead of asking participants just for
single forecasts for output growth,
inflation, and the unemployment rate,
it asks them to assign probabilities to
different outcomes and so gives a more
comprehensive picture of the forecasters’ views of the future.3
SPF FORECAST DISAGREEMENT
For the most part, the main, or
“headline,” forecast numbers reported
in the SPF are the median values
across forecasters. Each median,
though, belies the variation that exists
among individual projections for key
variables that describe the macroeconomy. In fact, the range of forecast
values underlying the median can be
substantial, and it changes over time.
At times, the forecasters show more
agreement and at other times more disagreement in their projections.
The Real-Time Data Research
Center website provides data on SPF
forecast disagreement for the variables
that are regularly reported in the SPF.
The measure of disagreement that is
reported is the difference between the
75th percentile and the 25th percentile
of the forecasts, which is called the
interquartile range. In other words,

3
See my 2012 article for more on forecast uncertainty and the forecast probabilities that are
reported in the SPF. That article also presents
some evidence on how forecast disagreement
affects the macroeconomy.

16 Q2 2014 Business Review

suppose there are 100 separate forecasts for annual real GDP growth in
2014. Order the forecasts from highest
value to lowest value, and take the
difference between the 75th slot and
the 25th slot as the measure of disagreement. We measure disagreement in
this way in order to ensure that any
outliers among the forecasts, perhaps
due to mistaken entries in the respondent questionnaires, do not unduly
influence the measure of disagreement.
Figures 1 through 3 show plots of
disagreement measured by the interquartile range from the center’s website
for real GDP growth, GDP price index
inflation, and the unemployment rate.
Each measure of disagreement is for
the four-quarters-ahead forecast as of
the date on the horizontal axis.

Roughly speaking, this suggests that
about 50 percent of the forecasts fall
within a range of about 0.4 percentage point below to about 0.4 percentage point above the median forecast.
The other 50 percent of the forecasts
are even further away from the median. Consequently, the disagreement
among the forecasters seems not too
large but nonetheless represents a
significant difference between the top
and bottom of the distribution. By way
of comparison, the standard deviation
of quarterly real GDP growth from the
first quarter of 1991 to the third quarter of 2013 was about 2.5 percentage
points at an annual rate.
Recall, though, that the measure
of disagreement shown in Figures 1
through 3 is somewhat conservative.

The range of forecast values underlying the
median can be substantial, and it changes
over time.
The charts show that disagreement generally tended to be higher
in the survey’s early years — the late
1970s and early 1980s — compared
with the latter half of the sample.
Broadly speaking, this pattern of declining disagreement tracks the period
known as the Great Moderation from
1984 to 2008, when the overall volatility of the economic data was lower
than in the pre-1984 period.4
How do we interpret the data in
Figures 1 through 3? Take the case
of disagreement for real GDP growth.
Since the early 1990s, the disagreement for forecasts of real GDP growth
four quarters ahead has bounced
around in a range of 0.5 percentage
point to 1.5 percentage points, with
an average of 0.86 percentage point.

For more on the Great Moderation, see my
2004 article, “What Accounts for the Postwar
Decline in Economic Volatility?”

4

To calculate it, we make no use of the
forecasts in the top and bottom 25
percent of the distribution — which,
if included, would widen the disagreement. Indeed, this potentially wide
disagreement is part of the reason that
the SPF generally reports median rather than average forecasts. The median
is the midmost forecast when forecasts
are ranked from high to low. So, unlike
with the average forecast, the effect of
outliers is discounted.
If we use all the forecasts to calculate the standard deviation across
projections of four-quarters-ahead real
GDP growth, we obtain Figure 4. For
the most part, the standard deviation measure tracks the interquartile
range measure fairly closely, though
it is clearly more volatile, especially
early in the sample. This volatility
may partly reflect reporting errors by
members of the forecast panel. Sometimes an SPF respondent will submit
www.philadelphiafed.org

FIGURES 1–3
Significant Dispersion for Key Indicators
Figure 1: Real GDP Growth Rate Dispersion by Quarter
Interquartile range of 4-quarters-ahead forecasts,
Percent
5.0
4.5
4.0
3.5
3.0
2.5
2.0
1.5
1.0
0.5
0
1974

1978

1982

1986

1990

1994

1998

2002

2006

2010

Figure 2: Unemployment Rate Dispersion by Quarter
Interquartile range of 4-quarters-ahead forecasts,
Percent
1.6
1.4
1.2
1.0
0.8
0.6
0.4
0.2
0
1974

1978

1982

1986

1990

1994

1998

2002

2006

2010

Figure 3: GDP Deflator Inflation Rate Dispersion by Quarter
Interquartile range of 4-quarters-ahead forecasts,
Percent
3.0
2.5
2.0
1.5
1.0
0.5

a forecast that seems extreme relative
to those of the other respondents.
If the outlier appears to be an error,
that forecast is removed. In general, though, it is difficult to identify
reporting errors versus actual views.
For example, in the survey for the
first quarter of 2013, the interquartile range for the four-quarters-ahead
real GDP growth projections is 0.6,
while the standard deviation is 1.7.
However, one forecaster had entered a
four-quarters-ahead real GDP growth
forecast of 12.6 percent — which
might have been a reporting error. If
we exclude that forecast, the standard
deviation falls to 0.832, which is much
closer to the interquartile range. The
median and the interquartile range are
less affected by such outliers.
We can see this in Figure 5, which
plots the forecasts for four-quartersahead real GDP growth from the SPF
for the first quarter of 2013. The gold
dots are observations in the upper 25
percent and lower 25 percent tails of
the distribution. We see two outliers,
one calling for 12.6 percent growth
and one calling for a 1.5 percent
contraction. We cannot say for sure
that these were reporting errors, but it
seems possible. Excluding those two,
the remainder fall in a range of about 2
percent to 4.2 percent, while the central tendency ranges from 2.5 percent
to 3.1 percent.
Disagreement still significant. If
we use all the forecasts to calculate the
standard deviation, we can construct
confidence intervals for the forecasters.
A confidence interval indicates the
probability of a forecast falling within
a certain range. Typically, a 95 percent
confidence interval is plus or minus
two standard deviations around the
mean estimate.5 In the case of forecasts for real GDP growth four quarters

0
1974

1978

1982

1986

1990

1994

1998

2002

2006

Source: Survey of Professional Forecasters, Federal Reserve Bank of Philadelphia.

www.philadelphiafed.org

2010
5
This is the case if the observations are drawn
from a normal distribution.

Business Review Q2 2014 17

FIGURE 4
Standard Deviation Generally Tracks GDP
Forecast Range
Interquartile range of 4-quarters-ahead forecasts,
Percent
8
7

Interquartile range
Standard deviation

6
5
4
3
2

Figure 7 shows the forecasts for
GDP deflator inflation from the 2013
first quarter SPF. Again, two look suspicious — both calling for significant
deflation — and would have a large
impact on the standard deviation of
the forecasts. Excluding the deflation
outliers, the forecasts range from about
1.3 percent for inflation four quarters
ahead to almost 4 percent. The central
tendency is narrow at 1.7 percent to 2.3
percent. But we again see that professional forecasters can have strikingly
different views about how the economy
will evolve over the next 12 months.

1
0
1974

1978

1982

1986

1990

1994

1998

2002

2006

2010

Sources: Survey of Professional Forecasters, Federal Reserve Bank of Philadelphia; author’s
calculations.

ahead, this calculation suggests that
we can estimate that 95 percent of the
forecasts will, on average, be in a range
of about 1.7 percentage points above to
1.7 percentage points below the mean
of the forecasts (the average standard
deviation of the four-quarters-ahead
projections from the first quarter of
1992 to the first quarter of 2013 is
0.85). So, if the average forecast was
3 percent, we would estimate that 95
percent of the forecasts for four-quarters-ahead real GDP growth would fall
in a range of about 1.3 percent to 4.7
percent. This range highlights that the
SPF forecasters typically have fairly
divergent views on how real output
growth is likely to evolve in the nottoo-distant future.
When looking at the interquartile
ranges of the forecasts for the unemployment rate and GDP deflator inflation, we also see a tendency toward a
decline in average disagreement in the
post-1990 sample. As it had for real
GDP growth, disagreement among
these forecasts increased following the
18 Q2 2014 Business Review

recession that began in December 2007.
Disagreement for the unemployment
rate forecast is a bit lower than that
for inflation or output growth, possibly
because the high persistence in the unemployment rate makes it a somewhat
easier variable to forecast. Nevertheless,
the range of views on future unemployment rates is significant. Take the
results of the survey taken in the first
quarter of 2013. Figure 6 plots the forecasts for the four-quarters-ahead unemployment rate from high to low. The
gold dots again denote the upper and
lower quartiles of the distribution. We
see that while the interquartile range
was fairly small at around 0.5, the range
of overall forecasts was larger. Some
forecasters thought that the economy
would make little if any progress on
the unemployment rate (the median
current-quarter forecast for the 2013
first quarter unemployment rate was 7.8
percent). But some thought unemployment would fall below 7 percent. Most
thought it would be in a range of 7.3
percent to 7.6 percent.

DIFFERENT MODELS AND
METHODS
The forecasters who make up
the SPF panel use a variety of statistical models to help them make their
projections, and this variety of models
surely plays a role in forecast disagreement. But how large a role might that
be? Their models generally fall into
one of two major categories: reducedform models and structural models.
Reduced-form models impose little, if
any, economic theory to refine their
structure. For example, one of the simplest forecasting models for real GDP
growth is to suppose that current GDP
growth is related to past GDP growth
in a linear fashion. To forecast a greater range of variables, the model can
be expanded by adding more lagged
variables to form a system of equations
that relates current values of variables
such as output growth, inflation, and
interest rates to their lagged values. A
forecaster using such a system of equations chooses which variables to have
in the system as well as the number
of lags of variables to use. Once those
choices are made, the model can be
estimated using historical data and
then used to generate forecasts. One
does not need to bring much economic
theory to bear when specifying such
a reduced-form model, since there are
typically no restrictions on the estiwww.philadelphiafed.org

FIGURES 5–7
Strikingly Different Year-Ahead Projections
Figure 5: Real GDP Growth Rate Forecasts
Q1 2013 4-quarters-ahead forecasts,
Percent
14
12
10
8
6
4
2
0
-2
-4
0

10

20

Top and bottom quartiles
Interquartile range

30

40

50

Individual survey responses

Figure 6: Unemployment Rate Forecasts
Q1 2013 4-quarters-ahead forecasts,
Percent
8.2
8.0
7.8
7.6
7.4
7.2
7.0
6.8
6.6
0

10

20

Top and bottom quartiles
Interquartile range

30

40

50

Individual survey responses

Figure 7: GDP Deflator Inflation Rate Forecasts
Q1 2013 4-quarters-ahead forecasts,
Percent
5
4
3
2
1
0
-1
-2
-3
-4
0

10

20

Top and bottom quartiles
Interquartile range

30

40

Individual survey responses
Source: Survey of Professional Forecasters, Federal Reserve Bank of Philadelphia.

www.philadelphiafed.org

50

mated coefficients and no restrictions
on how the model’s variables interact
with each other.
Another approach to building
a forecast model is to use economic
theory to restrict the way in which
the variables interact. For example,
one might stipulate that the relationship between household consumption
and hours worked in the labor market
is related to the real wage in some
particular way, or that the relationship
between current and future consumption is tied in a specific way to the
interest rate. Models that impose
economic theory on the data’s interrelationships are called structural
models. Like reduced-form models,
structural models relate values of variables to their lagged values. But structural models use economic theory to
impose complex relationships among
those variables. Structural models are
especially useful for honing the story
behind the forecast. For example,
such models may indicate that output
growth will rise because of higher
demand today, or that inflation is expected to fall because firms expect the
marginal costs of production to fall in
the future. These kinds of stories that
are consistent with economic theory
are typically more difficult to tease out
when using reduced-form models.
There is another element that usually is an important part of the forecast
process: judgment. Often forecasters
will examine the historical errors in
the equations of their models — that
is, how much the predicted value from
the equation differed from that actual
value in the data. If the forecast is
persistently missing on the high or low
side, the forecaster may alter the equation away from the estimated values a
bit so that its predictions are more in
line with the most recent data observations. This is a judgment-based adjustment of the forecast whereby forecasters subjectively alter the predictions
generated by the statistical model to
Business Review Q2 2014 19

bring them more into conformity with
the recent behavior of the economy
and their own views on how the future
is likely to unfold. Typically, data,
models, and judgment are combined to
produce the final forecast.
If forecasters are using different models on which to base their
forecasts, then we might expect to
see disagreement in those forecasts.
Some models might be more accurate descriptions of the economy than
others. If one forecasting methodology is consistently better than another
and the data are able to discriminate
among the models, we should see bad
models being driven out over time by
good models. Likewise, we should see
that forecasters using the best models should consistently produce better
forecasts than other forecasters do.
That is, if model heterogeneity is the
most important reason that forecasts
differ and the data are informative
about the models, then we should be
able to identify forecasters and models
that reliably outperform their peers.
Evidence casts doubt. However,
some evidence from the forecast evaluation literature casts some doubt that
model heterogeneity is the key element
behind forecast disagreement. First,
one of the most robust findings from
the forecasting literature is that mean
forecasts systematically outperform individual forecasts. That is to say that
over time, a more accurate forecast can
be had by taking the average of many
different forecasts rather than sticking
with one individual forecaster’s projections. If one forecaster and his or her
model were consistently producing better forecasts, then we wouldn’t expect
to see such a gain from averaging.6
Why does this forecast averaging tend to work so well compared
with any one forecast over time? If
there was one, known, true model of
6
See the 2006 article and references therein by
Allan Timmermann.

20 Q2 2014 Business Review

the economy, then the forecasts from
that model would dominate alternative models. But we don’t have the true
model of the economy. Yet, because
the economy is so complex, different
models may capture different features
of the economy in a successful way. At
certain times, some of those features
may be more important for successful
forecasts than at other times. So by
averaging a wide range of forecasts, we
can incorporate many features of the
economy that are impossible to capture

shocks should generate persistently
better forecasts as long as the shocks
a particular model is good at analyzing are still important drivers of the
dynamics of the economy. We might
then expect that some forecasters will
give more accurate forecasts for several
reporting periods in a row. However,
another finding from the empirical
forecasting literature is that the best
forecaster in any one period is no more
likely to be the best in the next period.7
So, although economic shocks are per-

By averaging a wide range of forecasts, we can
incorporate many features of the economy that
are impossible to capture in any single model
and on average make more accurate forecasts.
in any single model and on average
make more accurate forecasts.
Aside from forecast averaging,
is there other evidence that suggests
model differences may not be the key
factor underlying forecast dispersion?
Models can differ in how they incorporate economic shocks, which are
defined as unpredictable disturbances
to the economy from events such as
the outbreak of war or an unexpected
surge in global commodity prices. For
example, some models might be better
at predicting how the economy will
respond to oil price shocks, while other
models might be good at predicting
how the economy will respond to fiscal
policy shocks. Therefore, depending
on the specific mix of shocks hitting
the economy at any one time, some
models may produce the most accurate
forecasts for one period, only to have
their relative predictive power reversed
when a different set of shocks hits.
However, shocks and their effects tend
to persist. That is, the impact tends to
decline slowly as the shocks fully work
their way through the economy. This
persistence implies that models that
are especially accurate for particular

sistent, forecast performance is not.
The findings that average forecasts tend to outperform individual
forecasts and that top forecasters don’t
stay on top for long suggest that differences in models may not be the most
important element behind forecast
disagreement. But as is often true in
economics, the case is not so clear-cut.
Recent research by Andrew Patton
and Allan Timmermann examines
how forecast disagreement changes
with the forecast horizon. Using survey data from Consensus Economics,
they find greater disagreement among
longer-term forecasts than near-term
forecasts. Because variables such as
real GDP growth and inflation tend to
return to their mean values over time,
the observation that long-horizon forecasts show more disagreement is consistent with the idea that differences in
economic models are an important factor. This is because different models
might be calibrated to yield different
long-run averages for key variables, and
model-based long-run forecasts will
7
See the 2007 article by Michael Bryan and
Linsey Molloy.

www.philadelphiafed.org

FIGURE 8
No Wider Variation in Longer-Term Forecasts
Range of 1-Quarter-Ahead GDP Forecasts
Forecasts for real GDP growth,
Percent
8
6
4
2
0
-2
2010 Q1

2011 Q1

2012 Q1

2013 Q1

Range of 4-Quarters-Ahead Forecasts
Forecasts for real GDP growth,
Percent
8
6
4
2
0
-2
2010 Q1

2011 Q1

2012 Q1

2013 Q1

Source: Survey of Professional Forecasters, Federal Reserve Bank of Philadelphia.
Note: In the upper chart, one outlier forecast from the 2012Q3 survey was eliminated. In the
lower chart, one outlier from each of the following surveys was eliminated: 2010Q1, 2012Q3,
and 2013Q1.

reflect those differences.
This pattern is more difficult to
discern in recent SPF projections for
real GDP growth. However, the longest forecasts we can use from the SPF
are looking only four quarters ahead
from the quarter in which the survey is
conducted, which is more in the realm
of near-term forecasts. With that caveat, Figure 8 plots individual forecasts
of one-quarter-ahead and four-quartersahead real GDP growth from the first
quarter of 2010 to the second quarter
of 2013. Looking at the panels in the
www.philadelphiafed.org

figure, there is no clear tendency for
the longer-horizon forecasts to show
more disagreement than the shorterhorizon forecasts over this period.8 If
we were to extend the data sample back
to 1970, we would see the same basic
pattern: There is no obvious increase in
disagreement when we move from one8
Patton and Timmermann were able to compare forecast horizons as short as one month
with those as long as 24 months. In addition,
they used the Consensus Forecast survey for
1991 to 2008, which typically had about 25
forecasters in the panel.

quarter-ahead to four-quarters-ahead
real GDP forecasts in the SPF.
In considering how near-term
forecasts might differ from long-term
forecasts, it is important to note that
near-term forecast disagreement is more
likely to be influenced by current information that is used to kick off the forecast. As noted earlier, a forecast combines models with data and judgment
to generate a projection. The timeliness of the data that a forecaster has in
hand when making a forecast and the
extent to which the forecaster discerns
the true state of the economy from that
data are critical elements when projecting the economy’s future state. Perhaps
the data and information analysis that
go into forecasts are also key drivers of
forecast disagreement.9
AN IMPERFECT SIGNAL
As new data become available and
are put into the forecasting models,
the projections are modified, sometimes dramatically. So how forecasters
respond to the arrival of new information is an important factor in the
forecast-generation process. When
constructing models of the economy,
economists often assume for simplicity’s sake that people costlessly receive
all the information they need to make
their decisions. In reality, though,
information is costly to process and
often subject to revision over time.
Take the release of quarterly real GDP
data. An initial estimate is released in
the month after the end of the quarter,
a second release two months after the
end of the quarter, and a final release
three months after the end of the
quarter. But even that’s not really the
“final” release, since the estimate will
again be revised in July for the next
several years, and then again every five
years or so with benchmark revisions

9
Note, though, that without some discipline,
models with heterogeneous forecasting rules can
rationalize any disagreement outcome.

Business Review Q2 2014 21

to the national income and product accounts.10 Forecasters looking at today’s
data to put into their models realize
that they have only an imperfect signal
of the true state of the economy at any
point in time. In contrast to the case
of full information, we can say that
there are information “frictions” that
make the true state of the economy
difficult and costly to assess.
There are two prominent theories
of imperfect information in macroeconomics that have different implications
for what we should observe in forecast
surveys such as the SPF. The first,
often referred to as the sticky-information theory, is described in a 2002 paper by Greg Mankiw and Ricardo Reis.
In this theory, economic agents such as
forecasters are assumed to update the
information they use to make decisions and forecasts randomly — with
a certain probability each period that
is independent of economic conditions
or past decisions. It is as if each day
people play an information lottery. If
they win the lottery, they go ahead and
update their view of the world based
on current data. And the assumption
is that they receive full information
about the state of the economy when
they update. If they lose, they don’t
update their information and continue
to make decisions and forecasts based
on stale data.
Clearly, this is an extreme view of
the world and is unlikely to be strictly
true. But if there is a fixed cost to
acquiring and processing new information, then households and firms

10
The most recent benchmark revisions were
conducted in the summer of 2013. According to
the Bureau of Economic Analysis, comprehensive revisions encompass (1) updated definitions
and classifications to more accurately portray
the evolving U.S. economy, (2) changes in
presentations to make the NIPA tables more
informative, and (3) statistical changes that
introduce improved methodologies and newly
available and revised source data. See http://
www.bea.gov/newsreleases/national/gdp/
gdpnewsrelease.htm.

22 Q2 2014 Business Review

will update their information only
infrequently.11
Another prominent theory of
information frictions assumes instead
that agents monitor the flow of data
and update their information continually, but that the information they receive about economic fundamentals is
contaminated by random “noise” that
obscures the signals they are interested
in. For example, take the case of a
monetary policymaker who is concerned about the behavior of inflation
when setting interest rates. At a point
in time, the policymaker has observations on current and past inflation and
tries to discern the trend in inflation
from transitory movements that are
likely to dissipate over time. Inflation
may be higher today because of, say, a
temporary weather shock that affects
food prices. The policymaker is more
likely concerned with the underlying
trend rate of inflation but must make
some inference about that unobserved
trend from the underlying data. So, he
or she doesn’t necessarily possess the
full information. More generally, time
and resources must be spent to best
estimate the desired information, and
because time and resources are costly,
there are tradeoffs in deciding how to
process information.
Information frictions have implications for forecast behavior, including forecast disagreement. How do
forecasters respond to economic shocks
when there are information rigidities?
In the case of sticky information, not
all forecasters are updating their projections in response to the shock at the
same time.12 Since the respondents are
surveyed at the same time, we would
hypothesize that the average forecast

11
That the presence of fixed costs of updating
information can rationalize the sticky-information model is derived in a 2006 article by
Ricardo Reis.

It is plausible, for instance, that an economist
employed by a bank or other firm to generate

12

will be somewhat inertial and slow
to adjust to the shock. Consequently,
forecast errors will persistently be above
or below zero (depending on whether
the shock is positive or negative) for
some time, though eventually everyone
updates his or her information and the
average forecast error returns to zero.13
Similarly, in the case of imperfect
information, the change in the average
forecast in response to the shock is
inertial — in this case only a fraction
of the signal about economic fundamentals is incorporated into the current estimate of the underlying state
of the economy so that adjustment
to the data is only partial. This slow
response is reflected in the persistence of forecasters’ beliefs about the
underlying state of the economy, which
in turn leads to a somewhat inertial
adjustment of the forecasts to shocks
to economic fundamentals.
Effect of shocks on disagreement. So, both imperfect information
theories suggest that forecast errors
might be persistently below or above
zero in response to shocks (but that
the errors converge to zero over time).
But what about forecast disagreement?
Here, the two theories offer different
predictions. The sticky-information
theory predicts that disagreement will
rise in response to shocks. This happens because not all forecasters are updating their information sets and forecasts after a shock, so the forecasts of
those who do update may move further
away from the forecasts of those who
don’t: Disagreement increases. The
noisy-information theory predicts that
disagreement should not respond to
shocks. In this theory, forecasters conforecasts largely for internal use might update
those forecasts less frequently than would a
forecaster who primarily sells forecasts to a wide
clientele.
Key references for the macroeconomic
implications of imperfect information literature
include Lucas (1972), Sims (2003), and Woodford (2003).

13

www.philadelphiafed.org

tinually monitor and react to the flow
REFERENCES
of
data.14 The idea is easier to grasp if
we assume that forecasters are using
the same model. Then, if they are also
monitoring the data continually, the
disagreement among forecasts arises
from idiosyncratic differences in how
the forecasters process the information that ends up being fed into their
models. For example, suppose the data
for the model include the real interest
rate, which is not observed directly but
must instead be inferred by subtracting the expected rate of inflation from
the nominal interest rate. Depending
on how they measure expected inflation, different forecasters can arrive at
different measures of the real interest
rate. As long as the ways that individual forecasters process information do
not themselves respond to shocks, then
the dispersion of the forecasts will not
vary in response to shocks.
To sum up then, both imperfectinformation theories predict that
forecast errors should show some
persistence, but they have different
implications for forecast disagreement.
The sticky-information theory suggests
that disagreement rises in response to
shocks, while the noisy-information
theory suggests that it doesn’t.15
IS THERE EMPIRICAL
EVIDENCE?
Is there empirical evidence on the
role of imperfect information in forecaster behavior? Recent work by Olivier Coibion and Yuriy Gorodnichenko examines whether the implications
This assumes that the dispersion of the idiosyncratic noise shocks that forecasters receive
does not respond to economic fundamentals
such as inflation or output. However, if forecasters who receive the same signal interpret
it differently, then forecast dispersion can be
correlated with economic shocks. See Coibion
and Gorodnichenko (2012) for details.

of imperfect-information theories are
found in the forecast data. They use a
variety of surveys, including the SPF,
to investigate how forecast errors and
disagreement respond to shocks and
whether that response can be rationalized by imperfect-information theories.
Assessing the response of economic variables to shocks can be
tricky because shocks themselves are
often unobserved and so have to be

information problems are resolved.
Indeed, Coibion and Gorodnichenko
generally find serially correlated forecast errors in response to a variety of
shocks, especially TFP and oil price
shocks. For example, after an inflationary shock, they find a predictable
sequence of serially correlated positive
inflation forecast errors. Over time,
though, these errors converge back
to zero — just as the theory predicts

One key prediction of the imperfect information
theories is that forecast errors should show
some persistence in response to shocks.
identified from the data. Coibion and
Gorodnichenko identify several shocks,
including monetary policy shocks,
total factor productivity (TFP) shocks,
oil shocks, and fiscal policy shocks,
and then use regression methods to
see how inflation forecast errors and
forecast disagreement respond to these
shocks.16 For the most part, the forecast horizon under investigation is one
year ahead.
Recall that one key prediction of
the imperfect information theories is
that forecast errors should show some
persistence in response to shocks. If a
shock hits the economy and forecasters
either don’t incorporate it quickly into
their forecasts or else have a hard time
extracting the relevant signals from
the data, then their forecasts are likely
to over- or under-shoot until these

they should. If there were no information problems confronting forecasters,
we would expect that forecast errors
would in turn not show predictable
patterns in the data.
Coibion and Gorodnichenko also
examine how forecast disagreement
responds to shocks. Recall that this response has the potential to distinguish
between the sticky-information and
noisy-information theories. After examining how disagreement changes in
response to many different shocks that
hit the economy over the past 30 years
or so, they conclude that, on balance,
structural shocks do not seem to notably increase disagreement across forecasters. This finding gives an edge to
the noisy-information theory, though
for other dimensions of the data, the
sticky-information theory does better.17

16
Total factor productivity is the residual in
accounting for economic output after the contributions of labor and capital inputs have been
measured. It can be viewed as the contribution
of technological change to output growth. A
monetary policy shock can be thought of as the
surprise component of the monetary policy instrument. It is the difference between a realized
policy outcome — usually a short-term interest
rate — and the rate that had been predicted by
a specific model.

17

14

The two theories have additional predictions
for the data besides those we have discussed
here. See the 2012 paper by Coibion and Gorodnichenko for a fuller exposition.

15

www.philadelphiafed.org

Some of Coibion and Gorodnichenko’s
findings, though, are more consistent with the
sticky-information theory. For example, the
convergence rate of forecast errors is just as
rapid for monetary policy shocks as it is for TFP
shocks. Under the noisy-information theory, if
TFP shocks were more important for determining productivity and economic growth, one
might expect forecasters to pay more attention
to these shocks, which implies that forecast error convergence would differ among shocks.
Business Review Q2 2014 23

It seems that elements of both theories
are present in the data.
This evidence indicates that imperfect-information theories are consistent with the forecast data, though
discriminating between the two theories is difficult. Is imperfect information important or insignificant in the
data? Coibion and Gorodnichenko also
try to answer this question using data
from the SPF on inflation forecasts. In
the context of the sticky-information
model, their estimates imply an average duration of six to seven months
between information updates on the
part of forecasters.18 In the context

of imperfect-information models, the
estimates imply that new information
receives less than half the weight it
would if there was no imperfect information and forecasters did not have to
extract the relevant signals. Thus, the
evidence is consistent with imperfect
information having a significant role in
how forecasters process new information and how often they acquire it.
18
It seems puzzling that professional forecasters
would update their forecasts infrequently, but
other researchers have also found patterns in
the data consistent with this implication. Two
such studies are Clements (2012) and Andrade
and Le Bihan (2010).

CONCLUSION
Forecast disagreement is large and
varies over time. Differences in how
forecasters see the economy evolving in the future can be attributed
to the models they use and the way
they monitor and incorporate into
their forecasts the heavy, continual
flow of information on the state of the
economy. Empirical evidence from
forecast surveys suggests that both
model heterogeneity and imperfect
information about the economy play
roles in the wide dispersion in professional forecasts. BR

REFERENCES
Andrade, Philippe, and Hervé Le Bihan.
“Inattentive Professional Forecasters,”
Banque de France Working Paper 307
(2010).
Bryan, Michael F., and Linsey Molloy.
“Mirror, Mirror, Who’s the Best Forecaster
of Them All?” Federal Reserve Bank of
Cleveland Economic Commentary (2007).
Clements, Michael P. “Do Professional
Forecasters Pay Attention to Data Releases?” International Journal of Forecasting,
28 (April-June 2012), pp. 297-308.
Coibion, Olivier, and Yuriy Gorodnichenko. “What Can Survey Forecasts Tell Us
About Information Rigidities?” Journal of
Political Economy, 120:1 (2012) pp. 116-159.
Coibion, Olivier, and Yuriy Gorodnichenko.
“Information Rigidity and the Expectations
Formation Process: A Simple Framework
and New Facts,” National Bureau of
Economic Research Working Paper 16537
(2010).

24 Q2 2014 Business Review

Lucas, Robert E. “Expectations and the
Neutrality of Money,” Journal of Economic
Theory, 4 (April 1972), pp. 103-124.
Mankiw, N. Gregory, and Ricardo Reis.
“Sticky Information Versus Sticky Prices:
A Proposal to Replace the New Keynesian Phillips Curve.” Quarterly Journal
of Economics, 117 (November 2002),
pp. 1,295-1,328.
Patton, Andrew J., and Allan Timmermann.
“Why Do Forecasters Disagree? Lessons
from the Term Structure of Cross-Sectional
Disagreement,” Journal of Monetary Economics, 57:7 (2010), pp. 803-820.
Reis, Ricardo. “Inattentive Producers,”
Review of Economic Studies, 73 (July 2006),
pp. 793-821.
Timmermann, Allan. “Forecast Combinations,” in G. Elliott, C. Granger, and A.
Timmermann, eds., Handbook of Economic
Forecasting, 1:1 (2006), pp. 135-196.

Sill, Keith. “What Accounts for the Postwar Decline in Economic Volatility?” Federal Reserve Bank of Philadelphia Business
Review (Fourth Quarter 2004).
Sill, Keith. “Measuring Economic Uncertainty Using the Survey of Professional Forecasters,” Federal Reserve Bank
of Philadelphia Business Review (Fourth
Quarter 2012).
Sims, Christopher A. “Implications of
Rational Inattention,” Journal of Monetary
Economics, 50 (April 2003), pp. 665-690.
Stark, Tom. “Realistic Evaluation of RealTime Forecasts in the Survey of Professional Forecasters,” Federal Reserve Bank
of Philadelphia Special Report (May 2010).
Woodford, Michael. “Imperfect Common
Knowledge and the Effects of Monetary
Policy,” in P. Aghion, R. Frydman, J. Stiglitz, and M. Woodford, eds., Knowledge,
Information, and Expectations in Modern
Macroeconomics in Honor of Edmund
Phelps, Princeton, N.J.: Princeton University Press (2003).

www.philadelphiafed.org

Research Rap

Abstracts of
research papers
produced by the
economists at
the Philadelphia
Fed

Economists and visiting scholars at the Philadelphia Fed produce papers of interest to the professional researcher on banking, financial markets, economic forecasting, the housing market, consumer
finance, the regional economy, and more. More abstracts may be found at www.philadelphiafed.org/
research-and-data/publications/research-rap/. You can find their full working papers at
http://www.philadelphiafed.org/research-and-data/publications/working-papers/.

The Continuing Power of the Yield Spread
in Forecasting Recessions
The authors replicate the main results of
Rudebusch and Williams (2009), who show
that the use of the yield spread in a probit
model can predict recessions better than the
Survey of Professional Forecasters. Croushore
and Marsten investigate the robustness of their
results in several ways: extending the sample
to include the 2007-09 recession, changing the
starting date of the sample, changing the ending date of the sample, using rolling windows
of data instead of just an expanding sample,
and using alternative measures of the “actual”
value of real output. The results show that the
Rudebusch-Williams findings are robust in all
dimensions.
Working Paper 14-5. Dean Croushore, University of Richmond and Federal Reserve Bank of
Philadelphia Visiting Scholar; Katherine Marsten,
University of Richmond.
Continuous Markov Equilibria with
Quasi-Geometric Discounting
The authors prove that the standard quasigeometric discounting model used in dynamic
consumer theory and political economics does
not possess continuous Markov perfect equilibria (MPE) if there is a strictly positive lower
bound on wealth. The authors also show that,
at points of discontinuity, the decision maker
strictly prefers lotteries over the next period’s
assets. The authors then extend the standard
model to have lotteries and establish the
existence of an MPE with continuous decision
www.philadelphiafed.org

rules. The models with and without lotteries are
numerically compared, and it is shown that the
model with lotteries behaves more in accord with
economic intuition.
Working Paper 14-6. Satyajit Chatterjee, Federal
Reserve Bank of Philadelphia; Burcu Eyigungor,
Federal Reserve Bank of Philadelphia.
The Economics of Debt Collection:
Enforcement of Consumer Credit Contracts
In the U.S., third-party debt collection
agencies employ more than 140,000 people and
recover more than $50 billion each year, mostly
from consumers. Informational, legal, and other
factors suggest that original creditors should have
an advantage in collecting debts owed to them.
Then, why does the debt collection industry
exist and why is it so large? Explanations based
on economies of scale or specialization cannot
address many of the observed stylized facts. The
authors develop an application of common agency
theory that better explains those facts. The model
explains how reliance on an unconcentrated
industry of third-party debt collection agencies
can implement an equilibrium with more intense
collections activity than creditors would implement by themselves. The authors derive empirical
implications for the nature of the debt collection
market and the structure of the debt collection
industry. A welfare analysis shows that, under certain conditions, an equilibrium in which creditors
rely on third-party debt collectors can generate
more credit supply and aggregate borrower surplus
than an equilibrium where lenders collect debts
owed to them on their own. There are, however,
Business Review Q2 2014 25

situations where the opposite is true. The model also suggests a number of policy instruments that may improve the
functioning of the collections market.
Working Paper 14-7. Viktar Fedaseyeu, Bocconi University
and Federal Reserve Bank of Philadelphia Visiting Scholar; Robert M. Hunt, Federal Reserve Bank of Philadelphia.
Foreclosure Delay and Consumer Credit Performance
The deep housing market recession from 2008 through
2010 was characterized by a steep increase in the number
of foreclosures. Foreclosure timelines — the length of time
between initial mortgage delinquency and completion of
foreclosure — also expanded significantly, averaging up to
three years in some states. Most individuals undergoing foreclosure are experiencing serious financial stress. However,
extended foreclosure timelines enable mortgage defaulters to
live in their homes without making housing payments until
the completion of the foreclosure process, thus providing
a liquidity benefit. This paper tests whether the resulting
liquidity was used to help cure nonmortgage credit delinquency. The authors find a significant relationship between
longer foreclosure timelines and household performance on
nonmortgage consumer credit during and after the foreclosure process. Their results indicate that a longer period of
nonpayment of housing-related expenses results in higher
cure rates on delinquent nonmortgage debts and improved
household balance sheets. Foreclosure delay may have mitigated the impact of the economic downturn on credit card
default. However, credit card performance may deteriorate in
the future as the current foreclosure backlog is cleared and
the affected households once again incur housing expenses.
Working Paper 14-8. Paul Calem, Federal Reserve Bank of
Philadelphia; Julapa Jagtiani, Federal Reserve Bank of Philadelphia; William W. Lang, Federal Reserve Bank of Philadelphia.
Competing for Order Flow in OTC Markets
The authors develop a model of a two-sided asset market
in which trades are intermediated by dealers and are bilateral. Dealers compete to attract order flow by posting the
terms at which they execute trades, which can include prices,
quantities, and execution times, and investors direct their
orders toward dealers that offer the most attractive terms of
trade. Equilibrium outcomes have the following properties.
First, investors face a trade-off between trading costs and
speeds of execution. Second, the asset market is endogenously segmented in the sense that investors with different asset
valuations and different asset holdings will trade at different
speeds and different costs. For example, under a Leontief
technology to match investors and dealers, per unit trading
costs decrease with the size of the trade, in accordance with
the evidence from the market for corporate bonds. Third,
26 Q2 2014 Business Review

dealers’ implicit bargaining powers are endogenous and typically vary across sub-markets. Finally, the authors obtain a
rich set of comparative statics both analytically, by studying
a limiting economy where trading frictions are small, and
numerically. For instance, the authors find that the relationship between trading costs and dealers’ bargaining power can
be hump-shaped.
Working Paper 14-9. Benjamin Lester, Federal Reserve
Bank of Philadelphia; Guillaume Rocheteau, University of California–Irvine; Pierre-Olivier Weill, University of California–Los
Angeles.
Forecasting Credit Card Portfolio Losses in the Great
Recession: A Study in Model Risk
Credit card portfolios represent a significant component
of the balance sheets of the largest US banks. The charge-off
rate in this asset class increased drastically during the Great
Recession. The recent economic downturn offers a unique
opportunity to analyze the performance of credit risk models
applied to credit card portfolios under conditions of economic stress. Specifically, the authors evaluate three potential
sources of model risk: model specification, sample selection,
and stress scenario selection. Their analysis indicates that
model specifications that incorporate interactions between
policy variables and core account characteristics generate the most accurate loss projections across risk segments.
Models estimated over a time frame that includes a significant economic downturn are able to project levels of credit
loss consistent with those experienced during the Great
Recession. Models estimated over a time frame that does not
include a significant economic downturn can severely underpredict credit loss in some cases, and the level of forecast
error can be significantly impacted by model specification
assumptions. Higher credit-score segments of the portfolio
are proportionally more severely impacted by downturn economic conditions and model specification assumptions. The
selection of the stress scenario can have a dramatic impact
on projected loss.
Working Paper 14-10. José J. Canals-Cerdá, Federal Reserve Bank of Philadelphia; Sougata Kerr, Federal Reserve Bank
of Philadelphia.
Misallocation, Informality, and Human Capital:
Understanding the Role of Institutions
The aim of this paper is to quantify the role of formalsector institutions in shaping the demand for human capital
and the level of informality. The authors propose a firm
dynamics model where firms face capital market imperfections and costs of operating in the formal sector. Formal
firms have a larger set of production opportunities and the
ability to employ skilled workers, but informal firms can
www.philadelphiafed.org

avoid the costs of formalization. These firm-level distortions
give rise to endogenous formal and informal sectors and,
more importantly, affect the demand for skilled workers. The
model predicts that countries with a low degree of debt enforcement and high costs of formalization are characterized
by relatively lower stocks of skilled workers, larger informal
sectors, low allocative efficiency, and measured TFP. Moreover, the authors find that the interaction between entry
costs and financial frictions (as opposed to the sum of their
individual effects) is the main driver of these differences.
This complementarity effect derives from the introduction of
skilled workers, which prevents firms from substituting labor
for capital and in turn moves them closer to the financial
constraint.
Working Paper 14-11. Pablo N. D’Erasmo, University of
Maryland and Federal Reserve Bank of Philadelphia; Hernan J.
Moscoso Boedo, University of Virginia; Asli Senkal, University
of Virginia.

banks. Banks accumulate securities like Treasury bills and
undertake short-term borrowing when there are cash flow
shortfalls. A nontrivial size distribution of banks arises out
of endogenous entry and exit, as well as banks’ buffer stocks
of securities. The authors test the model using business cycle
properties and the bank lending channel across banks of different sizes studied by Kashyap and Stein (2000). They find
that a rise in capital requirements from 4% to 6% leads to a
substantial reduction in exit rates of small banks and a more
concentrated industry. Aggregate loan supply falls and interest rates rise by 50 basis points. The lower exit rate causes
the tax/output rate necessary to fund deposit insurance to
drop in half. Higher interest rates, however, induce higher
loan delinquencies as well as a lower level of intermediated
output.
Working Paper 14-13. Dean Corbae, University of Wisconsin–Madison and National Bureau of Economic Research;
Pablo N. D’Erasmo, Federal Reserve Bank of Philadelphia.

Market Exposure and Endogenous Firm Volatility over
the Business Cycle
The authors propose a theory of endogenous firm-level
volatility over the business cycle based on endogenous market exposure. Firms that reach a larger number of markets
diversify market-specific demand risk at a cost. The model is
driven only by total factor productivity shocks and captures
the business cycle properties of firm-level volatility. Using
a panel of U.S. firms (Compustat), the authors empirically
document the countercyclical nature of firm-level volatility. They then match this panel to Compustat’s Segment
data and the U.S. Census’s Longitudinal Business Database
(LBD) to show that, consistent with their model, measures
of market reach are procyclical, and the countercyclicality
of firm-level volatility is driven mostly by those firms that
adjust the number of markets to which they are exposed.
This finding is explained by the negative elasticity between
various measures of market exposure and firm-level idiosyncratic volatility the authors uncover using Compustat, the
LBD, and the Kauffman Firm Survey.
Working Paper 14-12. Ryan Decker, University of Maryland; Pablo N. D’Erasmo, University of Maryland and Federal
Reserve Bank of Philadelphia; Hernan J. Moscoso Boedo,
University of Virginia.

Trade Adjustment Dynamics and the Welfare Gains from
Trade
The authors build a micro-founded two-country dynamic general equilibrium model in which trade responds
more to a cut in tariffs in the long run than in the short run.
The model introduces a time element to the fixed-variable
cost trade-off in a heterogeneous producer trade model.
Thus, the dynamics of aggregate trade adjustment arise from
producer-level decisions to invest in lowering their future
variable export costs. The model is calibrated to match
salient features of new exporter growth and provides a new
estimate of the exporting technology. At the micro level, the
authors find that new exporters commonly incur substantial
losses in the first three years in the export market and that
export profits are back-loaded. At the macro level, the slow
export expansion at the producer level leads to sluggishness
in the aggregate response of exports to a change in tariffs,
with a long-run trade elasticity that is 2.9 times the short-run
trade elasticity. The authors estimate the welfare gains from
trade from a cut in tariffs, taking into account the transition
period. While the intensity of trade expands slowly, consumption overshoots its new steady-state level, so the welfare
gains are almost 15 times larger than the long-run change in
consumption. Models without this dynamic export decision
underestimate the gains to lowering tariffs, particularly when
constrained to also match the gradual expansion of aggregate trade flows.
Working Paper 14-14. George Alessandria, Federal Reserve
Bank of Philadelphia; Horag Choi, Monash University; Kim
Ruhl, New York University Stern School of Business.

Capital Requirements in a Quantitative Model of
Banking Industry Dynamics
The authors develop a model of banking industry
dynamics to study the quantitative impact of capital requirements on bank risk taking, commercial bank failure, and
market structure. They propose a market structure where
big, dominant banks interact with small, competitive fringe
www.philadelphiafed.org

Business Review Q2 2014 27

TM

Now Available:
The Federal Reserve Historical Inventory
and History Web Gateway
The Federal Reserve’s inventory of historical collections offers
students, researchers, and others online access to thousands of
documents and artifacts related to the Fed’s 100-year history
from sources across the Federal Reserve System, universities,
and private collections. To view the inventory, go to http://www.
federalreserve.gov/newsevents/press/other/other20120530a1.
pdf. More content and features will be added over time. Do you
know of materials that should be included? Information may be
submitted at http://www.federalreserve.gov/apps/contactus/
feedback.aspx.
Also available is the new Federal Reserve History Web Gateway.
Designed to encourage deeper reflection on the Fed’s role in the
nation’s economy, the gateway presents biographies, timelines,
photographs, and more that tell the story of the Fed’s purpose,
key economic events affecting U.S. history, and the people who
shaped the Fed. Go to http://www.federalreservehistory.org/.
On December 23, 1913, President Woodrow Wilson signed the
Federal Reserve Act, establishing the Federal Reserve System
as the U.S. central bank. Its mission is to conduct the nation’s
monetary policy; supervise and regulate banks; maintain the
stability of the financial system; and provide financial services to
depository institutions, the U.S. government, and foreign official
institutions. The Federal Reserve Board opened for business on
August 10, 1914, and on November 16, 1914, the 12 regional
Reserve Banks were open for business.
Congress designed the Fed with a decentralized structure. The
Federal Reserve Bank of Philadelphia — serving eastern Pennsylvania, southern New Jersey, and Delaware — is one of 12
regional Reserve Banks that, together with the seven-member
Board of Governors in Washington, D.C., make up the Federal
Reserve System. The Board, appointed by the President of the
United States and confirmed by the Senate, represents the public sector, while the Reserve Banks and the local citizens on their
boards of directors represent the private sector.
The Research Department of the Philadelphia Fed supports the
Fed’s mission through its research; surveys of firms and forecasters; reports on banking, markets, and the regional and U.S.
economies; and publications such as the Business Review.

PRESORTED STANDARD
U.S. POSTAGE
PAID

Federal R eserve Bank of Philadelphia

PHILADELPHIA, PA
PERMIT #583

Ten Independence Mall
Philadelphia, Pennsylvania 19106-1574

ADDRESS SERVICE REQUESTED

www.philadelphiafed.org

Past and current Business Review articles can be downloaded for
free from our website. There you will also find data and other
information on the regional and national economy, consumer
finance issues, resources for teachers, information on our
community development initiatives, the latest research publications
from the Federal Reserve Bank of Philadelphia, and more.

explore

and

learn