The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
Conducting Monetary
Policy Without
Government Debt:
The Fed’s Early Years
David C. Wheelock
he Federal Reserve System is the largest single
owner of U.S. Treasury securities in the world,
excluding federal government accounts and
trust funds such as the Social Security and Medicare
trust funds. As of January 2, 2002, the Federal
Reserve System Open Market Account held $554.8
billion of U.S. government securities, or about 16
percent of the stock of Treasury securities held
outside of government accounts and trust funds.1
The Fed has acquired this portfolio through its
monetary policy operations. Currently, the Fed
implements its monetary policy by setting a target
for the federal funds rate and using open market
operations in U.S. government and agency securities to achieve that target. Purchases of securities
supply reserves to the banking system, and thus
tend to put downward pressure on the funds rate,
whereas sales of securities remove reserves and
put upward pressure on the funds rate.2
Federal Reserve holdings of government securities are the principal source of the nation’s currency
and depository institution reserve balances, and
hence the U.S. monetary base. In principle, the Fed
could add to the stock of bank reserves and currency
by purchasing any asset, but U.S. Treasury securities
offer at least two advantages over alternative assets
for conducting monetary policy. First, because the
Treasury market is extremely large and highly liquid,
the Fed is able to conduct large transactions that
give it substantial control over depository institution
reserve balances and, hence, the federal funds
rate, without having a disruptive impact on market
prices.3 Second, by using open market transactions
in Treasury debt to implement monetary policy,
the Fed avoids directly affecting the allocation of
private capital, which Federal Reserve Chairman
T
Alan Greenspan (2001) and other System officials
have cited as an important consideration in the
conduct of monetary policy.4
Because of the close relationship between
Federal Reserve holdings of U.S. Treasury securities
and the monetary base, as well as the advantages
of using open market operations in Treasury securities to implement monetary policy, a substantial
decline in the outstanding stock of Treasury debt
would pose a major challenge to policymakers.
Indeed, the stock of Treasury securities available
to private holders, including the Fed, declined from
1997 to 2001, prompting Federal Reserve officials
to consider how the Fed might conduct monetary
policy in a world without a deep, liquid Treasury
securities market.
The substantial size and liquidity of the U.S.
Treasury securities market emerged during World
War II. The stock of outstanding Treasury debt ballooned during the war and remained large during
the ensuing 50 years because of nearly continuous
annual federal budget deficits. Thus, for evidence on
how monetary policy might be conducted without
substantial reliance on Treasury debt, it is necessary
to look either to the experiences of other countries
or to the Fed’s own history before World War II.
Accordingly, this article describes the implementation of Federal Reserve monetary policy before World
War II and highlights that era’s experiences that
offer lessons for the conduct of policy in a possible
future world without Treasury debt.
The record of Federal Reserve policy before
World War II is not good, and some scholars contend
that the poor performance of monetary policy was
caused by the System’s desire also to affect the
1
Board of Governors of the Federal Reserve System Release H.4.1
( January 2, 2002).
2
Fed transactions take the form of outright purchases and sales, as well
as repurchase and matched sale-purchase transactions, of Treasury
and agency securities. The Fed enters the market nearly every business
day to offset influences on reserve markets beyond the Fed’s immediate
control, such as changes in the amount of currency in circulation or
in the size of U.S. Treasury balances at Federal Reserve Banks, that
otherwise would cause the federal funds rate to deviate from the Fed’s
target. See Board of Governors of the Federal Reserve System (1994) for
additional information about the implementation of monetary policy.
3
In 2001, the average daily volume of outright transactions in U.S.
government securities, as reported by primary dealers, was $298
billion. By contrast, in 2001, the Federal Reserve purchased an average
of $5.7 billion of Treasury securities per month. As noted, however,
the Federal Reserve also engages in repurchase and matched purchasesale agreements. See Dupont and Sack (1999) for an overview of the
U.S. Treasury securities market.
4
See also Broaddus and Goodfriend (2001).
David C. Wheelock is an assistant vice president and economist at the
Federal Reserve Bank of St. Louis. Heidi L. Beyer provided research
assistance.
© 2002, The Federal Reserve Bank of St. Louis.
M AY / J U N E 2 0 0 2
1
REVIEW
Wheelock
Figure 1
U.S. Government Debt / GDP
Annual data, 1917-2011*
1.2
financing speculative activity. Those conflicts help
explain the Fed’s failure to respond aggressively to
the Great Depression and illustrate how a policy
focused on the usage of Federal Reserve credit can
interfere with the implementation of an effective
stabilization policy.
1.0
THE RISE AND (POSSIBLE) FALL OF
THE STOCK OF TREASURY DEBT
0.8
0.6
0.4
0.2
0.0
1915 1925 1935 1945 1955 1965 1975 1985 1995 2005 2015
*CBO forecasts for 2001-11.
SOURCE: Board of Governors of the Federal Reserve System,
Bureau of Economic Analysis, and Robert Gordon,
Macroeconomics (2000).
private allocation of credit. This article describes the
Fed’s desire that the credit it supplied to markets not
encourage financial speculation or other forms of
“unproductive” activity and how that desire caused
policymakers to tighten credit overzealously in
response to a perceived misallocation in the late
1920s and remain too tight as the economy collapsed
into the Great Depression. The views that led to this
outcome do not influence Federal Reserve policy
today. However, as Goodfriend (1994) and Schwartz
(1992) contend, the Fed is at times pressured to
conduct a targeted credit policy. Moreover, if the
Fed were forced to conduct monetary policy using
private debt instruments, which could occur if the
stock of U.S. Treasury debt were to decline substantially, such pressures might increase. The Fed’s
prewar experience described here provides one
example of how the conduct of an effective monetary policy can be compromised by pressures to
affect the allocation of private-sector credit.5
The next section discusses recent changes in
the size of the U.S. Treasury debt and possible implications of a substantial decline in the outstanding
stock of debt for the implementation of monetary
policy. Following sections describe how the Fed’s
founders intended the System to conduct policy, the
development of Federal Reserve monetary policy
during the 1920s, and the conflicts created by the
Fed’s desire to prevent Federal Reserve credit from
2
M AY / J U N E 2 0 0 2
As of December 31, 2001, the outstanding debt
of the U.S. federal government totaled $5943.4
billion, of which $2549.0 billion was held by U.S.
government agencies and trust funds such as the
Social Security and Medicare trust funds. The remaining $3394.4 billion of outstanding debt was held
by the “public,” consisting of private individuals,
financial institutions and other firms, state and
local governments, foreign concerns, and the Federal
Reserve.6 The stock of government debt held by the
public reached a year-end peak in 1997 at $3846.7
billion. Since then, the surplus revenues of government trust funds invested in Treasury securities
have exceeded the amount by which total Treasury
debt has increased.7
Often, the stock of government debt is measured
relative to national output and, presumably, the
implications of a given amount of debt for monetary policy depend on the size of the economy. As
Figure 1 illustrates, the ratio of U.S. government
debt held by the public to gross domestic product
(GDP) soared during World War II; it then declined
steadily to the mid-1970s before rising again to a
peak in 1993. Through September 2001 (the end
of the fiscal year), the stock of Treasury debt then
grew at a slower rate than did U.S. GDP. Projections
by the Congressional Budget Office and other forecasters indicate that the debt-to-GDP ratio will continue to decline for at least the next decade. Past
experience indicates that the size of the federal debt
5
Whereas the implementation of monetary policy using private credit
instruments can lead to potential conflicts between the conduct of
stabilization policy and the allocation of credit, analogous conflicts
can arise if the central bank conducts open market transactions in any
asset. For example, in the nineteenth and early twentieth centuries,
U.S. silver mining interests often pressed the federal government to
purchase and coin silver. Conceivably, such efforts to raise the relative
price of silver could conflict with monetary policy objectives. See
Friedman and Schwartz (1963, pp. 483-91) for a discussion of a silver
purchase program implemented in the 1930s.
6
See <www.publicdebt.treas.gov/opd/opdpdodt.htm>.
7
On September 30, 2001, the close of fiscal year 2001, the stock of
debt held by the public totaled $3339.3 billion.
FEDERAL RESERVE BANK OF ST. LOUIS
is difficult to forecast at horizons of more than a
year or two (see Kliesen and Thornton, 2001). Nevertheless, if true, recent projections indicate that by
mid-decade the debt-to-GDP ratio could fall to a level
not observed since the 1920s, and by about 2010 fall
to a level not observed since before World War I.8
A substantial decline in the volume of outstanding Treasury securities would have repercussions
for both the Fed and the financial system. Treasury
securities, especially Treasury bills, serve as liquid,
risk-free investments and collateral for banks and
other financial market participants. For the Fed,
either a substantial increase in discount window
borrowing, which typically is secured by loans and
other private claims, or greater use of securities
other than those issued by the U.S. Treasury in the
conduct of open market operations would expose
the System to more credit risk than it faces today.
Aside from possibly affecting monetary policy, such
exposure could complicate other duties the Fed
performs. For example, if the Fed were to become a
major creditor of the banks that it supervises, then
any aggressive actions it might take as a bank supervisor to deal with problem banks could increase the
probability of losses by the Fed as a bank creditor.
The Fed has relied mainly on open market operations in Treasury debt to implement monetary
policy since World War II. Recently, Federal Reserve
Chairman Alan Greenspan (2001) summarized why
Treasury securities are a convenient asset for the
Fed to conduct its operations: “First, the liquidity of
the market allows the Federal Reserve to make substantial changes in reserves in a short period of time,
if necessary. Second, the size of the market has
meant that the effects of the Federal Reserve’s purchases on the prices of Treasury securities have
been minimal. Third, Treasury securities are free
of credit risk…[and] we believe that the effects of
Federal Reserve operations on the allocation of
private capital are likely to be minimized when
Federal Reserve intermediation involves primarily
the substitution in the public’s portfolio of one type
of instrument that is free of credit risk—currency—
for another—Treasury securities.”
Greenspan went on to identify how the Fed
might respond to a substantial decline in the stock
of Treasury debt: “One possibility is to expand the
use of the discount window by auctioning such credit
to financially sound depository institutions…
Another possibility is to add new assets to those
the Fed is currently allowed by law to buy for its
portfolio.” Greenspan cited Ginnie Mae securities
Wheelock
and certain types of municipal or foreign government obligations as examples of securities that the
Fed might use for open market operations. Either
increased reliance on the discount window (in which
depository institutions borrow reserves mainly
against collateral other than U.S. Treasury securities)
or an expansion of the financial assets the Fed
purchases in the open market would cause the
implementation of monetary policy to resemble
more closely the methods used by the Fed before
World War II.
WHAT THE FED’S FOUNDERS
INTENDED
In discussing the advantages of conducting open
market operations in U.S. Treasury debt, Chairman
Greenspan (2001) argued that “it is important that
government holdings of assets not distort the private
allocation of capital” and that “if the Treasury debt
is paid down…then the Federal Reserve will have
to find alternative assets that still provide substantial
liquidity and minimize distortions to the private
allocation of capital.” The notion that actions to
implement stabilization policy should not distort
the private allocation of capital is not controversial
today. When the Federal Reserve System was established in 1914, however, its founders very much did
intend the System to favor certain uses of private
sector capital over others. Moreover, the founders
had no expectation that the Fed would conduct
stabilization policy as we know it today. In short,
the Fed’s founders envisioned that the Fed would
conduct credit policy, but not monetary policy, in
that Federal Reserve operations were expected to
influence the private allocation of credit but not
regulate the growth rate of the monetary base or
the level of interest rates to achieve macroeconomic
stability.9
The Discount Window
The Fed’s founders expected that Federal Reserve
Banks, like the central banks of Europe, would serve
mainly as lending institutions for their member
commercial banks. Before the Fed’s establishment,
the U.S. commercial banking system suffered numer8
Projected annual ratios of debt to GDP for 2001-11 plotted in Figure 1
are from a Congressional Budget Office (2001) forecast made in August
2001, which was the latest available as of year-end 2001.
9
This distinction is discussed in Goodfriend (1994) and Broaddus and
Goodfriend (2001).
M AY / J U N E 2 0 0 2
3
REVIEW
Wheelock
ous banking panics and, at times, high failure rates
(see Dwyer and Gilbert, 1989). Proponents of the
Federal Reserve System argued that the Fed would
make the banking system more stable by providing
commercial banks with a ready source of reserves
to meet fluctuations in the demands for credit and
currency. To perform that role, Federal Reserve
Banks were given a lending facility—their discount
windows—through which they would rediscount
eligible financial assets for member commercial
banks in exchange for currency or reserve deposit
balances.10 Member banks were required to maintain minimum reserve balances with their Reserve
Bank, and the Reserve System provided currency
(Federal Reserve notes) and payments services for
member institutions.11
The Federal Reserve Act provided that “any
Reserve Bank may discount notes, drafts, and bills
of exchange arising out of actual commercial transactions; that is, notes, drafts, and bills of exchange
issued or drawn for agricultural, industrial, or commercial purposes.” In addition, to be eligible for
rediscount, agricultural paper could have a maturity
of no more than six months, whereas nonagricultural
paper had to mature in 90 days or less.
The limits that the Fed’s founders placed on
the type of securities eligible for rediscount with
Reserve Banks reflected conventional banking principles of the time, the so-called Real Bills Doctrine.
By limiting discount loans to short-term commercial
and agricultural loans, the Fed’s founders expected
that the System would supply a sufficient volume
of credit to accommodate growth and fluctuations
in real economic activity without causing inflation
or speculation. The Federal Reserve Act explicitly
prohibited the rediscount of “notes, drafts, or bills
covering merely investments or issued or drawn
for the purpose of carrying or trading stocks, bonds,
or other investment securities, except bonds and
notes of the Government of the United States.” By
ruling such securities ineligible, the Fed’s founders
sought to prevent Federal Reserve credit from being
used to finance transactions or investments that
had no obvious, direct connection to the production, distribution, or sale of specific products or
commodities.12
Apparently, the authors of the Federal Reserve
Act believed it more important to specify precisely
the type of securities that would be eligible for rediscount than to specify criteria for setting the discount
rate. The Federal Reserve Act stated only that Reserve
Bank discount rates should be determined “with a
view of accommodating commerce and business.”
4
M AY / J U N E 2 0 0 2
Reserve Banks were required to maintain a gold
reserve against their liabilities, however, which
imposed implicit bounds on their lending rates.
Open Market Operations in Bankers
Acceptances
The coalition of interests supporting the founding of the Federal Reserve System included both
small community banks and large money center
banks.13 The large money center banks were particularly interested in promoting the dollar as an
international currency and thereby increasing the
share of international transactions financed by U.S.
banks (Broz, 1997). Accordingly, the Federal Reserve
Act permitted U.S. commercial banks to issue bankers
acceptances to finance foreign trade. The Act further
sought to encourage the development of a U.S.
acceptance market by permitting Federal Reserve
Banks to acquire acceptances by rediscount or open
market purchase. The Reserve Banks established
interest rates at which they purchased all eligible
acceptances offered to them, and such purchases
were a major source of Federal Reserve credit during
the System’s first two decades.
Open Market Operations in U.S.
Government Securities
In addition to open market purchases of bankers
acceptances, the Federal Reserve Act authorized
Reserve Banks “to buy and sell, at home or abroad,
bonds and notes of the United States, and bills, notes,
revenue bonds, and warrants…issued…by any State
[sic], county, district, political subdivision, or municipality in the continental United States.” This provision was intended to provide the Reserve Banks with
a source of revenue in the event that income from
10
At the time, most commercial and agricultural loans were made on a
discount basis. Hence, when such loans were offered to Reserve Banks
in exchange for currency or reserve balances, the Reserve Banks
rediscounted the paper at the current discount rate.
11
The Fedwire system, which the Fed established in 1918 to effect interbank payments electronically, was among the innovations enhancing
the liquidity of the payments system. See Gilbert (1998) for an analysis of how the founding of the Federal Reserve affected the efficiency
of the U.S. payments system.
12
Burgess (1936) and West (1977) discuss the objectives of the Fed’s
founders in detail.
13
This is not to say that all banks favored the creation of the Fed. Numerous small, state-chartered banks elected not to join the System, and
many opposed the Fed’s check collection practices (see Gilbert, 1998).
For further analysis of the Fed’s “membership problem,” see White
(1983).
FEDERAL RESERVE BANK OF ST. LOUIS
Wheelock
rediscounts and the provision of services was insufficient to meet Bank expenses (Chandler, 1958, p. 76).
The Fed’s founders did not contemplate that open
market operations would be used to influence the
level of market interest rates or the growth rate of the
money stock in an effort to stabilize the price level
or economic activity. It was not long, however, before
Federal Reserve policymakers discovered that open
market operations could affect the level of interest
rates and, potentially, influence economic activity.
Figure 2
THE SOURCES OF FEDERAL RESERVE
CREDIT
1500
Federal Reserve credit constitutes the Fed’s contribution to the stock of bank reserves and currency
in circulation, the sum of which is referred to as
the monetary base or high-powered money. Other
sources of the monetary base include Treasury
currency outstanding (e.g., coins) and, historically,
the monetary gold stock. The principal sources of
Federal Reserve credit are discount window loans
and Fed purchases of U.S. government securities.
Before World War II, Fed purchases of bankers
acceptances also contributed meaningfully to Fed
credit.14
Figure 2 illustrates the relative volumes of discount window loans, Federal Reserve holdings of
acceptances, and Fed holdings of U.S. government
securities during 1914-41. Total Federal Reserve
credit grew sharply during World War I, when the
Fed committed itself to helping finance the war
effort. The Federal Reserve Act was amended in
1916 to permit member banks to borrow directly
from the Fed using U.S. government securities and
other eligible assets as collateral. Discount loan
volume soared when the Fed established preferential
discount rates for advances secured by U.S. government securities that guaranteed banks a profit on
their holdings of such securities. In June 1917, discount window loans outstanding (rediscounts and
advances) totaled $197 million, of which 13 percent
were advances against member bank holdings of
U.S. government securities. In June 1919, by contrast,
discount loans totaled $1818 million; of these, 87 percent represented rediscounts of, or advances against,
U.S. government securities (Board of Governors,
1943, p. 340).
The Federal Reserve also purchased considerable U.S. government securities in the open market
during the war. Fed holdings of Treasury securities
increased from $66 million in June 1917 to $236
Principal Sources of Federal Reserve Credit
Annual data, 1914-41
$ millions
3500
3000
2500
2000
1000
500
0
1915
1920
1925
1930
1935
1940
Discount loans and advances
Fed holdings of bankers acceptances
Fed holdings of U.S. government securities
SOURCE: Board of Governors of the Federal Reserve System.
million in June 1919. The percentage of total Fed
credit outstanding accounted for by Fed purchases
of U.S. government securities remained at approximately 10 percent over the two years. Meanwhile,
the Fed also acquired bankers acceptances in the
open market. Throughout the war, more Federal
Reserve credit was extended by the purchase of
acceptances than by open market purchases of U.S.
government securities.
The Fed retained preferential discount rates on
loans secured by U.S. government securities after
World War I. By late 1919, however, declining reserve
ratios at several Reserve Banks prompted the Banks
to increase their discount rates and to discontinue
preferential rates on loans secured by U.S. government securities.15 Discount window loan volume
dropped sharply in 1921-22. From a peak of $2808
million in October 1920, discount loans outstanding
fell to less than $400 million in August 1922 (Board
of Governors, 1943, p. 374).
The Reserve Banks sought to offset the loss of
revenue associated with the decline in discount
loans by purchasing government securities in the
14
Other sources of Fed credit include check float and Federal Reserve
purchases of foreign currency.
15
Reserve Banks were required by law to maintain reserves of gold and
eligible securities against their outstanding liabilities.
M AY / J U N E 2 0 0 2
5
REVIEW
Wheelock
open market (Chandler, 1958, p. 209). Open market
operations in government securities remained an
important source of Federal Reserve credit throughout the 1920s and even more so during the early
1930s, when discount window loans and Federal
Reserve purchases of bankers acceptances dwindled. By 1932, open market purchases of U.S. government securities had become the dominant source
of Federal Reserve credit, and for the remainder of
the decade the Fed made almost no discount loans
and purchased almost no bankers acceptances.
MONETARY VERSUS CREDIT POLICY
The changes in the composition of Federal
Reserve credit during the 1920s reflect well the
evolution from the self-regulating credit policy
envisioned by the Fed’s founders toward a modern
monetary policy. By the mid-1920s, the Fed had
begun to use open market operations in U.S. government securities to influence money market conditions, with the twin goals of promoting domestic
economic stability and the international gold standard. Credit policy came back to the fore in 1928-29,
however, when the Fed sought to check stock market
speculation without unduly restricting the flow of
credit to “legitimate” borrowers. The consequent
stance of monetary policy proved extremely restrictive, the stock market crashed, and the U.S. economy
collapsed.16 The Fed did not ease monetary policy
aggressively in response to the collapse, however,
in part because some Fed officials feared that loose
monetary policy would reignite financial speculation.
This section discusses the origins of Federal Reserve
monetary policy during the 1920s and the conflict
between monetary and credit policy that emerged
during the late 1920s. The following section explores
how this conflict influenced the setting of monetary
policy during the Great Depression.
The Birth of Monetary Policy
Although the provisions of the Federal Reserve
Act permitting the Reserve Banks to acquire government securities were little more than an afterthought, Reserve Bank officials, especially at the
Federal Reserve Bank of New York, observed that
their purchases affected market interest rates and
credit conditions. After World War I, the Reserve
Banks formed a committee of Reserve Bank governors, headed by Federal Reserve Bank of New York
Governor Benjamin Strong, to establish policies
for the conduct of open market operations and to
6
M AY / J U N E 2 0 0 2
coordinate open market purchases for all Reserve
Banks.17 Strong, according to Chandler (1958), was
among the first Reserve System officials to comprehend the impact of open market operations on
money market and credit conditions, as well as to
favor the use of open market operations to achieve
general macroeconomic goals.
Under Strong’s leadership, the Fed began to
use open market operations in U.S. government
securities during the 1920s to implement an active
monetary policy. This policy clashed with the credit
policy objectives of members of the Federal Reserve
Board and some Reserve Banks. This conflict came
to a head over how to control stock market speculation in 1928-29 and how to respond to the economic
depression that followed the stock market crash in
1929.
Strong directed two major monetary policy
operations during the 1920s, involving substantial
open market purchases in 1924 and 1927. The
motivation for these operations has been debated.
Chandler (1958), Friedman and Schwartz (1963),
Meltzer (1997), and Wicker (1966) all argue that
Strong was motivated by a desire to ease money
market conditions, but they disagree about Strong’s
ultimate objective. Friedman and Schwartz (1963)
contend that in both years Strong was motivated
primarily by a desire to promote domestic recovery
from a recession. Wicker (1966), however, argues
that Strong’s primary motivation was to redirect the
international flow of gold away from the United
States toward the United Kingdom, in an effort to
help Britain first restore, then preserve, gold convertibility of the pound. Chandler (1958) and Meltzer
(1997) contend that both objectives were important,
and Wheelock (1991) reports econometric evidence
consistent with their conclusion.
Whatever Strong’s motivation, his use of open
market operations caused considerable controversy
within the Federal Reserve System. Strong’s initiative irritated members of the Federal Reserve Board,
who believed that the committee of governors had
overstepped its authority. Several members of the
Board also opposed open market purchases, espe16
Schwartz (1981) and Hamilton (1987) argue that tight monetary policy
in 1928-29 was an important cause of the Great Depression.
17
From 1914 to 1935, the chief executive officer of each Federal Reserve
Bank held the title “governor,” as did the chair of the Federal Reserve
Board. The Banking Act of 1935 changed the title of Reserve Bank
chief executives to “president” and assigned the title “governor” to
each member of the Federal Reserve Board, which was renamed the
Board of Governors of the Federal Reserve System.
FEDERAL RESERVE BANK OF ST. LOUIS
cially in 1927, on economic grounds. Most members
of the Board, and officials of some Reserve Banks,
believed that Federal Reserve credit should be
extended only at the initiative of member commercial banks through the rediscounting of commercial
and agricultural loans. Otherwise, those officials
argued, the Fed risked contributing to speculative
activities that could prove harmful to the economy.
Strong, on the other hand, contended that the
Fed could help lift the economy during a period of
weakness by using open market purchases to ease
monetary conditions. At a meeting of Reserve Bank
governors in 1926, Strong argued: “Should we go
into a business recession while the member banks
were continuing to borrow [from Reserve Bank discount windows]…we should consider taking steps
to relieve some of the pressure which this borrowing
induces by purchasing Government securities and
thus enabling member banks to reduce their indebtedness” (quoted in Chandler, 1958, pp. 239-40). In
Strong’s view, by enabling banks to repay their discount window borrowings, open market purchases
would ease money market conditions and promote
economic recovery.
The Stock Market and the Return of
Credit Policy
Although the disagreements about open market
purchases in 1924 and 1927 were sharp, the most
heated debates within the System during the 1920s
focused on how the Fed should respond, if at all, to
the rising stock market. The rapid increase in stock
prices and growth of loans from banks and brokers
to finance stock purchases in 1928 and 1929 concerned System officials who sought to ensure that
reserves supplied by the Fed were not being used
to finance speculation. Most members of the Federal
Reserve Board favored a “direct action” policy in
which member banks with outstanding loans to
finance stock purchases would be prohibited from
borrowing at the discount window. Board members
thought that by enforcing this restriction, discount
rates would not have to rise and thereby penalize
borrowers with “legitimate” credit demands.
Officials of Federal Reserve Banks, however,
generally believed it neither practical nor desirable
for the Fed to affect the private allocation of credit.
In a 1925 memorandum, Benjamin Strong asked
rhetorically how the Fed should respond to calls for
action against real estate and stock market speculation or, for example, to “too much enthusiasm in
automobile production”:
Wheelock
Where does our responsibility lie? Must we
accept parenthood for every economic
development in the country? That is a hard
thing for us to do. We would have a large
family of children. Every time any one of
them misbehaved, we might have to spank
them all. There is no selective process in
credit operations. If we undertake “direct
action” in one case, we would be saddled
with the responsibility for direct action in
all cases. Have we infallible good judgment
as well as sufficient knowledge to play the
role of parent?…Of one thing I am sure…and
that is that we have no direct responsibility
to deal with isolated situations, and must rely
for the development of our policy upon
estimates of the whole situation. (quoted in
Chandler, 1958, p. 428)
In Strong’s view, the Fed should be concerned
with the stock market, or any other particular
market, only to the extent that it bears on the behavior of the economy as a whole.
To the extent that a rising stock market meant
that monetary policy should become tighter, Strong
favored raising the discount rate and conducting
open market sales, rather than placing special restrictions on banks that made stock market loans. Such
restrictions, he argued, would not limit the flow of
credit to the stock market: “the money will go into
the stock exchange anyway” (quoted in Chandler,
1958, p. 430). Even if the Fed lent only to banks
that made no stock market loans, Strong claimed,
reserves supplied through the discount window
(or via open market purchases) could still end up
enabling banks in the aggregate to increase stock
market loans: “If we create an addition to the volume
of credit by our open-market operations or by our
discounts, the banks which get it [i.e., the credit]
pass it along through all the channels through which
credit circulates in our banking system—and we
cannot control what happens to it. Some of it will go
in one direction and some of it will go in another,
and the nature and the use of our funds is perfectly
impossible to control” (quoted in Chandler, 1958,
pp. 431-32).
Direct Action
At its meeting on January 11, 1928, the Federal
Reserve’s open market committee decided to implement a more restrictive monetary policy, defined
as “somewhat firmer money conditions,” so as to
M AY / J U N E 2 0 0 2
7
Wheelock
“check unduly rapid further increases in the volume
of credit” (quoted in Chandler, 1971, p. 38). The
Reserve Banks also began to increase their discount
rates and, for the most part, these initial restrictive
actions were supported by the Federal Reserve
Board. One Board member, however, dissented from
all moves to tighten policy. In explaining his vote
against a discount rate increase in 1928, Edward
Cunningham stated that “I feel that increases in the
discount rate for the purpose of restricting stock
market activities should only be resorted to when
other means within the power of the Board have
failed to accomplish the objective. I am not in favor
of penalizing agriculture and business because of
the indirect use of credit for investments in brokers
loans” (quoted in Chandler, 1971, p. 43).
Despite further discount rate increases and open
market sales in 1928, Fed officials were frustrated
by their apparent inability to control the flow of
credit to the stock market. Chandler (1971, pp. 5253) summarizes the quandary the Fed found itself
in at the beginning of 1929:
By late January 1929 the Federal Reserve’s
policy of restriction had been in effect about
a year. Monetary and credit conditions had
changed markedly during the period. Member bank borrowings at the Federal Reserve
had nearly doubled, rising to nearly $900
million, equal to 37 percent of total bank
reserves…The total volume of bank credit
was barely above its level of a year earlier.
Interest rates had risen sharply…Call-loan
rates averaged above 7 percent in December
1928 and frequently reached considerably
higher levels. However, the Federal Reserve
had not achieved its objective of curbing
stock speculation. Share prices rose 38 percent in the year…Brokers’ loans reached
the unprecedented level of $6.4 billion;
this reflected an increase of 45 percent for
the year…Domestic business activity was
still at high and rising levels, but even here
there were warning signs in the form of
decreasing availability of mortgage money,
a downturn in construction, and increasing
difficulties in floating long-term bond issues.
In these circumstances, disagreements within the
System over how to respond to the stock market
became more heated.
Federal Reserve Board officials believed that
the Reserve Banks had not properly administered
8
M AY / J U N E 2 0 0 2
REVIEW
their discount windows and were permitting member banks to borrow reserves to support “speculative” lending, meaning primarily loans to stock
brokers and dealers and to customers for the purpose
of purchasing securities. Although Reserve Banks
rediscounted only loans and securities that were
eligible as defined by the Federal Reserve Act, Board
officials argued that commercial banks should be
forced to liquidate their speculative loans before
being permitted to rediscount (or borrow against)
eligible paper. On February 2, 1929, the Federal
Reserve Board sent a letter to each of the 12 Reserve
Banks in which the Board stated that
The Federal Reserve Act does not…contemplate the use of the resources of the Federal
reserve banks for the creation or extension
of speculative credit. A member bank is not
within its reasonable claims for rediscount
facilities at its Federal reserve bank when it
borrows either for the purpose of making
speculative loans or for the purpose of
maintaining speculative loans.
The letter went on to request that each Reserve
Bank report to the Board as to “a) how they keep
themselves fully informed of the use made of borrowings by their member banks, b) what methods
they employ to protect their institution against
the improper use of its credit facilities by member
banks, and c) how effective these methods have
been” (quoted in Chandler, 1971, pp. 56-57).
Although the Board’s instructions to the Reserve
Banks were vague—for example, the terms “speculative credit” and “speculative loans” were not
defined—the Reserve Banks made some effort to
comply with the Board’s request that they administer their discount windows more tightly. At the
same time, the Reserve Banks pressed for increases
in their discount rates, but were denied by the Federal
Reserve Board. Consequently, Reserve Bank officials
grew increasingly frustrated with the Board’s “direct
action” policy of tightly restricting access to the
discount window. As George Norris, Governor of the
Federal Reserve Bank of Philadelphia, complained
to one Board member:
This whole process of “direct action” is wearing, friction-producing, and futile. We are
following it honestly and energetically, but
it is manifest, beyond…doubt, that it will
never get us anywhere. It is like punching
at a mass of dough. You make a dent where
FEDERAL RESERVE BANK OF ST. LOUIS
Wheelock
Figure 3
Figure 4
Principal Sources of Federal Reserve Credit
M1 Money Stock and Total Bank Reserves
Monthly data, January 1928–February 1933
Monthly data, January 1928–February 1933
$ millions
3000
M1 ($ millions)
29000
2500
27000
2000
25000
1500
23000
1000
21000
500
19000
Reserves ($ millions)
2700
2500
2300
2100
M1
0
1928
1929
1930
1931
1932
1933
Sum of discount loans and Fed holdings of acceptances
Total Federal Reserve credit
15000
1928
The Federal Reserve Board eased its policy of
“direct action” in June 1929, when economic activity
had begun to slow and Fed officials were concerned
that credit had become too tight. In August, however,
the discount rate of the Federal Reserve Bank of
New York was increased in an effort to discourage
borrowing, and market interest rates remained high
until the stock market crashed in October. Rates
then fell sharply. The crash prompted the Federal
Reserve Bank of New York to make large open
market purchases while also lending heavily through
its discount window. The Federal Reserve, however,
did not respond aggressively to the sharp decline
in economic activity that followed the stock market
crash or to the banking panics that occurred over the
1700
1500
1929
1930
1931
1932
1933
SOURCE: Board of Governors of the Federal Reserve System
and Friedman and Schwartz (1963) for M1.
SOURCE: Board of Governors of the Federal Reserve System.
THE CRASH AND GREAT DEPRESSION
1900
17000
Fed holdings of U.S. government securities
you hit, but the mass swells up at another
point. As long as we maintain a discount rate
which is absurdly low, and out of proportion
to all other rates, the present conditions will
continue. Our 5 per cent rate is equivalent
to hanging a sign out over the door “Come
in,” and then we have to stand in the doorway and shout “Keep out.” It puts us in an
absurd and impossible position. (Quoted in
Chandler, 1971, p. 66)
Total Member Bank Reserves
next three years.18 One reason for the Fed’s inaction
is that Federal Reserve officials remained mired in
debate over whether the System should attempt to
channel credit to “appropriate” uses or pursue an
active stabilization policy. This section reviews that
debate, focusing on the arguments of Fed officials
who opposed the use of expansionary monetary
policy to revive the economy.
The Fed’s Response
The Federal Reserve Bank of New York purchased some $160 million of government securities
for its own account immediately following the stock
market crash, and the System purchased another
$150 million of securities before the end of 1929.
During 1930 and most of 1931, however, Fed purchases of government securities were insufficient
to offset net declines in discount window loans
and Fed purchases of bankers acceptances. Hence,
total Federal Reserve credit outstanding fell (see
Figure 3). As Friedman and Schwartz (1963) show,
the money stock began to fall in this phase of the
Depression (see Figure 4).
Monetary contraction accelerated in the fourth
quarter of 1931, when speculation that the United
18
Macroeconomic conditions during the Depression are summarized
in Wheelock (1992).
M AY / J U N E 2 0 0 2
9
REVIEW
Wheelock
States would abandon the gold standard triggered
bank runs and a gold outflow. Banks borrowed
heavily from the discount window to replace lost
reserves and Federal Reserve credit increased
sharply. The Fed did not make substantial open market purchases of government securities, however,
claiming that it lacked sufficient gold reserves.19
Easier monetary policy did come in 1932 when,
under pressure from Congress, the Fed purchased
some $1 billion of U.S. government securities
between March and August.20 The total increase
in Federal Reserve credit was less than $1 billion
because of declines in discount window loans and
Fed holdings of bankers acceptances, but member
bank reserves increased and the money stock
stopped falling. The banking crisis resumed in early
1933, however, triggering a series of state bank
suspensions and prompting President Franklin
Roosevelt to declare a national bank holiday and
suspend the gold standard when he took office in
March. The Federal Reserve, meanwhile, stayed in
the shadows as the new president took charge of
macroeconomic policy.
How Should the Fed Respond to a
Decline in Economic Activity?
Why the Federal Reserve failed to respond more
aggressively to the Great Depression has been the
subject of considerable research.21 Certainly, there
were officials in the System who advocated a more
vigorous response to the Depression. Officials of
the Federal Reserve Bank of New York, for example,
argued that recovery required low interest rates, a
strong bond market, and a sufficient supply of
reserves to free member banks from having to obtain
discount window loans.22 In July 1930, New York
Fed Governor George Harrison wrote to his counterparts at other Reserve Banks urging that the Fed
“do everything possible and within its power to
facilitate a recovery of business.” He went on to
advocate open market purchases: “In previous business depressions, recovery has never taken place
until there has been a strong bond market,” and,
moreover, “we cannot foresee any appreciable harm”
from making open market purchases (quoted in
Friedman and Schwartz, 1963, p. 370).
Outside of New York, however, many Fed officials
were convinced that Federal Reserve credit should
contract with declines in economic activity and
loan demand. Those officials claimed that in the
absence of demand for loans by business and agri10
M AY / J U N E 2 0 0 2
cultural borrowers, reserves created by an expansion
of Federal Reserve credit would be used to finance
speculation. Many argued that open market purchases during recessions in 1924 and 1927 had
been a mistake and had contributed to the financial
speculation that they saw as responsible for the
subsequent economic depression. For example,
Adolph Miller, a member of the Federal Reserve
Board who had voted against open market purchases
in 1927, testified in 1931 that the operation “was
the greatest and boldest operation ever undertaken
by the Federal Reserve System, and, in my judgment,
resulted in one of the most costly errors committed
by it or any banking system in the last 75 years…
That was a time of business recession. Business
could not use and was not asking for increased
money at that time” (U.S. Senate, 1931, p. 134).
Miller’s view was not unique among System
officials. In response to a written question from the
Senate Banking Committee in 1931 about open
market purchases in 1924 and 1927, officials of
the Federal Reserve Bank of Richmond wrote that
“we think United States securities should not have
been purchased in these periods, and the aim should
have been to decrease rather than augment the
total supply of Federal Reserve credit” (U.S. Senate,
1931, p. 100). Officials of the Federal Reserve Bank
of Philadelphia responded similarly, arguing that
Federal Reserve credit should never be extended
except at the initiative of member banks. Other
Reserve Banks replied that the open market purchases of 1924 and 1927 had been justified, but
were too large.
19
Burgess (1936, pp. 285-86) argues that the Fed was constrained by a
lack of gold reserves, but Friedman and Schwartz (1963, pp. 399-406)
dismiss the Fed’s excuse. See also Chandler (1971, pp. 182-91).
20
The Glass-Steagall Act of 1932 permitted the Fed to back Federal
Reserve notes with U.S. government securities, which greatly eased
the Fed’s gold reserve requirement.
21
See Friedman and Schwartz (1963), Wicker (1966), Brunner and
Meltzer (1968), Wheelock (1991, 1992), and Meltzer (1994).
22
During the Depression, Fed officials interpreted historically low levels
of borrowed reserves and interest rates as indicating that monetary
conditions were exceptionally easy. No one in the System, and almost
no one outside the Fed, recognized that a falling price level and widespread banking panics, let alone a decline in the money stock, meant
that monetary policy was in fact exceptionally tight. Friedman and
Schwartz (1963) argue that the Fed would have responded much more
aggressively to the Depression had Benjamin Strong not died in 1928.
During the 1920s, however, Strong advocated basing the volume of
open market operations on the levels of borrowed reserves and market
interest rates; the Fed’s anemic response to the Depression does not
seem inconsistent with that framework (see Wheelock, 1990, 1991,
1992).
FEDERAL RESERVE BANK OF ST. LOUIS
Those officials who were critical of open market
purchases during the 1920s tended to argue that
Federal Reserve credit should be extended only at
the initiative of member banks, through the discount
window or by sales of bankers acceptances to the
Fed. Open market purchases of government securities, by contrast, constituted “artificial” easing,
which Federal Reserve Bank of St. Louis Governor
William McChesney Martin Sr. argued was “unwise”
and possibly “hazardous” (quoted in Chandler, 1971,
p. 142). The Federal Advisory Council argued similarly in 1930 that “the present situation will be best
served if the natural flow of credit is unhampered
by open-market operations” (quoted in Friedman
and Schwartz, 1963, p. 373). Such operations,
claimed Chairman Richard Austin of the Federal
Reserve Bank of Philadelphia, “lays us open to the
apparent undesirable charge that the action is not
justified by the demand for credit but for some other
purpose, it may be for boosting business, making a
market for securities, or some other equally criticizable cause that certainly will come back to plague
us” (quoted in Chandler, 1971, p. 136).
Several Fed officials argued that monetary policy
could do little to bring about recovery from the
Depression. Officials of the Federal Reserve Bank
of Philadelphia, for example, concluded that “the
correction must come about through reduced production, reduced inventories, the gradual reduction
of consumer credit, the liquidation of security loans,
and the accumulation of savings through the exercise of thrift” (quoted in Chandler, 1971, p. 137).
And, in response to a proposal for open market purchases, Governor James McDougal of the Federal
Reserve Bank of Chicago replied that the market
already had an “abundance” of funds. Further, he
argued that the Fed should “maintain a position
of strength, in readiness to meet future demands…
rather than to put reserve funds into the market
when not needed” (quoted in Friedman and
Schwartz, 1963, p. 371).
In the view of McDougal and several other Fed
officials, open market purchases would have little
positive impact on economic activity and could in
fact interfere with economic recovery by delaying
the liquidation of loans and speculative investments
that, in their view, was necessary for recovery to
begin. Moreover, in the absence of an obvious
demand for Federal Reserve credit, as evidenced
by discount window borrowing or sales of bankers
acceptances to Reserve Banks, McDougal and others
believed that reserves created by open market purchases could result in a dangerous misallocation
Wheelock
Figure 5
Government Security Holdings as a Percent
of Total Adjusted Federal Reserve Credit*
Annual data, 1915-2001
Percent
100
90
80
70
60
50
40
30
20
10
0
1910
1920
1930
1940
1950
1960
1970
1980
1990
2000
*Total adjusted Federal Reserve credit = Sum of discount loans and
advances to depository institutions and Federal Reserve holdings of
U.S. government and other securities.
SOURCE: Board of Governors of the Federal Reserve System.
of credit. They were not able to prevent open
market purchases altogether, but their resistance
undoubtedly slowed the Fed’s response to the Great
Depression.
LESSONS
Ironically, the Fed’s unwillingness to purchase
a large volume of government securities early in
the Depression ultimately may have contributed to
such purchases becoming the dominant source of
Federal Reserve credit. As the Depression continued
and banking panics worsened, commercial banks
became increasingly unable and unwilling to come
to the discount window. Some banks lacked eligible
collateral for discount window loans, while others
feared that borrowing would trigger deposit withdrawals by giving the appearance of weakness.23
By 1932, discount window borrowing and Federal
Reserve purchases of bankers acceptances had fallen
to minimal levels, where they stayed throughout
the remainder of the decade. As Figure 5 illustrates,
Fed holdings of U.S. government securities had become by far the most important source of Federal
Reserve credit. Since 1934, the size of the Fed’s
23
Chandler (1971, pp. 225-33) concludes that borrowing was reduced
to some extent by a lack of eligible collateral and a heightened reluctance to borrow. Wheelock (1990) finds that discount window borrowing declined more during 1930-33 than could be explained simply
by the decline in economic activity.
M AY / J U N E 2 0 0 2
11
REVIEW
Wheelock
portfolio of U.S. government securities has always
comprised over 90 percent of total adjusted Federal
Reserve credit outstanding (the sum of discount
loans and Fed holdings of U.S. government and
other securities).
Monetary policy lay dormant from the mid-1930s
to 1951. Neither the size of the Fed’s government
security portfolio nor total Fed credit outstanding
changed substantially between 1934 and 1941.24
During World War II and for several years subsequently, the Fed’s open market operations were
directed entirely at maintaining low and stable yields
on U.S. Treasury securities, while discount window
borrowing and Fed acceptance purchases remained
minimal. An agreement between the Federal Reserve
and the Treasury in March 1951 (the “Accord”) freed
the Fed from rigid support of U.S. Treasury security
prices, enabling the System to pursue broader
policy objectives (Hetzel and Leach, 2001a). Under
William McChesney Martin Jr., who became Chairman
of the Federal Reserve’s Board of Governors following the Accord, the Fed initiated an active monetary
policy designed to limit inflation and the amplitude
of business cycles (Hetzel and Leach, 2001b). To
achieve those goals, the Fed has relied primarily on
open market operations in U.S. government securities to manipulate the volume of bank reserves and
influence market interest rates. Although at times
the quantity of discount window borrowing has
been an operational target for open market policy,
discount loans have been a far less important source
of Federal Reserve credit since 1951 than they were
before 1934, as have Fed purchases of bankers
acceptances (see Figure 5).25
Whereas the Great Depression was a defining
moment in the conduct of U.S. monetary policy
(Calomiris and Wheelock, 1998), it perhaps had even
more impact on the regulation of the financial system and the government’s role in credit allocation.
A host of federal loan corporations and other agencies to allocate credit, such as the Reconstruction
Finance Corporation, were founded or expanded.
The widely held view that stock market speculation
and commercial bank involvement in the underwriting, sale, and financing of security purchases
had caused the Depression led to fundamental
reforms of securities markets and the banking
system, including the Glass-Steagall Act of 1933,
which prohibited the commingling of commercial
and investment banking.
The Federal Reserve also was given expanded
powers to influence the allocation of credit. The
12
M AY / J U N E 2 0 0 2
Federal Reserve Act was amended in 1933 to authorize the Fed to set minimum margin requirements
for stock market loans, while giving the Federal
Reserve Board clear authority to deny discount window loans to banks that made speculative loans. At
the same time, the definition of acceptable collateral
for discount window loans was broadened and the
Fed was authorized under certain circumstances to
make loans to nonmember banks, groups of banks,
and even to individuals, partnerships, and corporations (Hackley, 1973).
Since the Accord, the Federal Reserve has effectively insulated its monetary policy from credit
allocation. Discount window lending has been a
small fraction of total Federal Reserve credit, and
the Fed largely discontinued the purchase of bankers
acceptances in the 1950s, though authorization to
purchase acceptances was not eliminated until 1998.
Moreover, discount window loans and other Federal
Reserve transactions, such as foreign exchange market intervention and warehousing, are prevented
from affecting the monetary base by means of offsetting open market operations.
Goodfriend (1994) contends that pressure to
allocate credit could still be detrimental to monetary
policy. At times, Congress and the Administration
have called upon the Fed to lend to distressed firms
and governments, such as Penn Central Corporation
in 1970 and New York City in 1975 (see Schwartz,
1992). Although the Fed has usually resisted such
calls, Goodfriend (1994) argues that pressure put on
the Fed to conduct targeted credit policy threatens
the Fed’s independence, which he views as crucial
to the conduct of effective monetary policy. Arguably,
if the Fed were to rely more heavily on discount
window lending or to conduct open market operations in assets other than U.S. Treasury securities,
the System could face intensified pressure to alter
the composition of its asset portfolio. The experience
of the Fed during the Great Depression suggests that
a desire to affect the allocation of credit, even one
24
The Fed’s few open market purchases maintained a constant-size
portfolio as holdings matured. Gold inflows from abroad poured
reserves into the U.S. banking system, however, and banks amassed
high levels of reserves in excess of legal requirements. This “golden
avalanche” produced rapid growth of the money stock (Friedman and
Schwartz, 1963). The Fed did raise bank reserve requirements in 1936
and 1937, fearing the inflationary potential of excess reserves, and
lowered them under pressure from the Administration in 1938 when
the economy slipped into recession. See Calomiris and Wheelock
(1998) for discussion.
25
See Meulendyke (1989) for an overview of monetary policy since the
Accord.
FEDERAL RESERVE BANK OF ST. LOUIS
that originates from within the Federal Reserve
System, could undermine its monetary policy.
When direct lending to commercial banks was
an important source of Federal Reserve credit,
Federal Reserve officials became concerned that
banks were not employing the reserves they acquired
through the discount window appropriately. Elaborate collateral requirements were imposed by the
Federal Reserve Act, and at various times the Fed
reiterated that borrowing is a privilege, not a right.
Still, Fed officials became dissatisfied with the use
of credit and sought to impose tight controls to limit
borrowing. The result was an exceptionally tight
monetary policy that carried over into the Great
Depression, when many Fed officials feared that
aggressive monetary easing would only reignite
financial speculation.
The Federal Reserve is unlikely to repeat the
egregious error of contracting Federal Reserve credit
and the monetary base during a serious economic
downturn. However, if direct lending to financial
institutions or open market operations in assets
other than U.S. Treasury securities become important
in the implementation of monetary policy, the Fed’s
early history warns that new pressures to conduct
a credit policy could arise that might hamper the
conduct of monetary policy.
REFERENCES
Board of Governors of the Federal Reserve System. Banking
and Monetary Statistics, 1914-1941. Washington, DC, 1943.
___________. The Federal Reserve System: Purposes and
Functions. Washington, DC, 1994.
Broaddus, J. Alfred Jr. and Goodfriend, Marvin. “What Assets
Should the Federal Reserve Buy?” Federal Reserve Bank
of Richmond 2000 Annual Report, 2001.
Broz, J. Lawrence. The International Origins of the Federal
Reserve System. Ithaca, NY: Cornell University Press, 1997.
Brunner, Karl and Meltzer, Allan H. “What Did We Learn
from the Monetary Experience of the United States in
the Great Depression?” Canadian Journal of Economics,
May 1968, pp. 334-48.
Wheelock
Policy?” in Michael D. Bordo, Claudia Goldin, and
Eugene N. White, eds., The Defining Moment: The Great
Depression and the American Economy in the Twentieth
Century. Chicago: University of Chicago Press, 1998, pp.
23-66.
Chandler, Lester V. Benjamin Strong: Central Banker.
Washington, DC: The Brookings Institution, 1958.
___________. American Monetary Policy, 1928-1941. New
York: Harper and Row, 1971.
Congressional Budget Office. The Budget and Economic
Outlook. August 2001. <www.cbo.gov> (as posted July
11, 2001).
Dupont, Dominique and Sack, Brian. “The Treasury Securities
Market: Overview and Recent Developments.” Board of
Governors of the Federal Reserve System Federal Reserve
Bulletin, December 1999, pp. 785-806.
Dwyer, Gerald P. Jr. and Gilbert, R. Alton. “Bank Runs and
Private Remedies.” Federal Reserve Bank of St. Louis
Review, May/June 1989, 71(3), pp. 43-61.
Friedman, Milton and Schwartz, Anna J. A Monetary History
of the United States, 1867-1960. Princeton, NJ: Princeton
University Press, 1963.
Gilbert, R. Alton. “Did the Fed’s Founding Improve the
Efficiency of the U.S. Payments System?” Federal Reserve Bank
of St. Louis Review, May/June 1998, 80(3), pp. 121-42.
Goodfriend, Marvin. “Why We Need an ‘Accord’ for Federal
Reserve Credit Policy: A Note.” Journal of Money, Credit,
and Banking, August 1994, 26(3), pp. 572-80.
Greenspan, Alan. “The Paydown of Federal Debt.” Remarks
before the Bond Market Association, White Sulphur
Springs, West Virginia, 27 April 2001.
Hackley, Howard H. Lending Functions of the Federal Reserve
Banks: A History. Washington, DC: Board of Governors
of the Federal Reserve System, 1973.
Burgess, W. Randolph. The Reserve Banks and the Money
Market. New York: Harper and Brothers, 1936.
Hetzel, Robert L. and Leach, Ralph F. “The Treasury-Fed
Accord: A New Narrative Account.” Federal Reserve Bank
of Richmond Economic Quarterly, Winter 2001a, 87, pp.
33-55.
Calomiris, Charles W. and Wheelock, David C. “Was the
Great Depression a Watershed for American Monetary
___________ and ___________. “After the Accord:
Reminiscences on the Birth of the Modern Fed.” Federal
M AY / J U N E 2 0 0 2
13
REVIEW
Wheelock
Reserve Bank of Richmond Economic Quarterly, Winter
2001b, 87, pp. 57-64.
Hamilton, James D. “Monetary Factors in the Great
Depression.” Journal of Monetary Economics, March 1987,
19(2), pp. 145-69.
Kliesen, Kevin L. and Thornton, Daniel L. “The Expected
Federal Budget Surplus: How Much Confidence Should
the Public and Policymakers Place in the Projections?”
Federal Reserve Bank of St. Louis Review, March/April
2001, 83(2), pp. 11-24.
Meltzer, Allan H. “Why Did Monetary Policy Fail in the
1930s?” Working paper, 1994.
October 1992, 74(5), pp. 58-69.
U.S. Senate, Committee on Banking and Currency, 71st
Congress, 3rd Session. Operation of the National and
Federal Reserve Banking Systems. Washington, DC:
Government Printing Office, 1931.
West, Robert C. Banking Reform and the Federal Reserve,
1863-1923. Ithaca, NY: Cornell University Press, 1977.
Wheelock, David C. “Member Bank Borrowing and the
Fed’s Contractionary Monetary Policy During the Great
Depression.” Journal of Money, Credit, and Banking,
November 1990, 22, pp. 409-26.
___________. “New Procedures, New Problems.” Working
paper, 1997.
___________. The Strategy and Consistency of Federal Reserve
Monetary Policy, 1924-1933. Cambridge: Cambridge
University Press, 1991.
Meulendyke, Ann-Marie. U.S. Monetary Policy and Financial
Markets. New York: Federal Reserve Bank of New York,
1989.
___________. “Monetary Policy in the Great Depression:
What the Fed Did and Why.” Federal Reserve Bank of
St. Louis Review, March/April 1992, 74(2), pp. 3-28.
Schwartz, Anna J. “Understanding 1929-1933,” in Karl
Brunner, ed., The Great Depression Revisited. Boston:
Martinus Nijhoff, 1981, pp. 5-48.
White, Eugene N. The Regulation and Reform of the American
Banking System, 1900-1929. Princeton: Princeton
University Press, 1983.
___________. “The Misuse of the Fed’s Discount Window.”
Federal Reserve Bank of St. Louis Review, September/
Wicker, Elmus R. Federal Reserve Monetary Policy, 1917-1933.
New York: Random House, 1966.
14
M AY / J U N E 2 0 0 2
Unemployment
Insurance Claims and
Economic Activity
William T. Gavin and Kevin L. Kliesen
A
lthough the Federal Open Market Committee
(FOMC) monitors a large number of economic
series when deciding whether to alter the
current stance of its policy, it is generally accepted
that policymakers, as well as financial markets, pay
especially close attention to labor market indicators
during periods of economic uncertainty. The reason,
in short, is that changes in labor market activity
are thought to be useful predictors for changes in
real gross domestic product (GDP), the broadest
measure of economic activity.
The main indicators of activity in the labor
market include the civilian unemployment rate,
nonfarm payroll employment, and average weekly
hours, all of which are reported monthly in the
Employment Situation from the Bureau of Labor
Statistics (BLS). Indeed, the release of the monthly
employment report seemingly rivals the post-FOMC
meeting press release as the single most anticipated
economic event in the financial markets. Given its
significance, therefore, it is probably not too surprising that economists and market participants try to
anticipate changes in this and other labor market
indicators.
When it comes to forecasting monthly changes
in the unemployment rate or the number of new
nonfarm jobs created or destroyed, it appears that
many economists and market participants pay
particularly close attention to the report on initial
unemployment insurance claims. This report, which
is published by the Employment Training Administration (ETA), an agency within the U.S. Department
of Labor, attempts to measure, on a weekly basis,
labor flows from the ranks of the employed to the
ranks of the unemployed (initial claims). The report
also measures the total number of people currently
unemployed who are eligible to receive unemployment insurance benefits (continuing claims).
William T. Gavin is a vice president and economist and Kevin L.
Kliesen is an economist at the Federal Reserve Bank of St. Louis. The
authors thank Cynthia Ambler of the Department of Labor for providing information on the unemployment insurance program. Rachel J.
Mandal and Thomas A. Pollmann provided research assistance.
© 2002, The Federal Reserve Bank of St. Louis.
We begin with a brief review of the important
monthly labor market data and their usefulness to
economists, policymakers, and financial market
participants. We then examine whether these labor
market indicators are useful for predicting concurrent growth rates of real GDP. Finally, the paper will
examine whether there is significant information
to be gleaned from weekly changes in initial and
continuing unemployment claims for predicting
these monthly labor market indicators.
LABOR MARKET DATA
There are three major sources of data for the
labor market: the household survey, the establishment survey, and the reports of state agencies that
collect information about employment for the
unemployment insurance program. The former
two comprise the information that is found in the
monthly employment report, while the latter is the
source of information for the weekly unemployment
insurance claims data.
The Household Survey
The household survey collects information from
a small but representative sample of households.
Currently, about 60,000 households are surveyed
either in person or by telephone each month by the
Bureau of Census. This survey, although it comprises
less than 0.06 percent of the roughly 107 million
households in the United States, is meant to be a
representative sample of the U.S. civilian noninstitutional population, from which trends in labor market
activity can be inferred. From that survey, known
as the Current Population Survey (CPS), the BLS culls
information on the demographics of the job market,
such as race, age, sex, educational level, and detailed
information about those who are unemployed, such
as the duration of their unemployment.
The most important information from the CPS
is the unemployment rate, which is plotted in
Figure 1. Here, the monthly unemployment rates
are averaged to get a quarterly rate. Since there is
thought to be a significant cyclical relationship
between changes in the unemployment rate and
changes in aggregate output, also included is the
four-quarter growth rate of GDP.1 Visual evidence
1
The relationship between real GDP growth and the unemployment
rate is sometimes characterized by Okun’s law. Named after the late
economist Arthur Okun, the “law” says that for every percentage point
that real GDP growth is above (below) its potential growth, the unemployment rate will fall (rise) by one-half a percentage point. See Mankiw
(1998).
M AY / J U N E 2 0 0 2
15
REVIEW
Gavin and Kliesen
participation rate of women approached the rate of
men, the unemployment rate has gradually declined.
Figure 1
Unemployment Rate and Real GDP
The Establishment Survey
Percent
12
Unemployment
m
Rate
10
8
6
4
2
Real GDP
0
–2
–4
–6
–8
1965
1974
1983
1992
2001
NOTE: The unemployment rate is a quarterly average of
monthly rates. Real GDP is shown as a four-quarter growth
rate. Bars indicate periods of recession.
suggests that, during recessions, the unemployment
rate usually rises as real GDP declines. At other times,
though, the relationship does not hold very well,
suggesting that trends in the unemployment rate
are not a reliable indicator of GDP growth. Indeed,
as shown in Table 1, the correlation between the
four-quarter growth of real GDP and the contemporaneous value of the unemployment rate is negative,
but with a relatively low value (–0.27). One reason
why the two series might not be more closely correlated is that the unemployment rate lags the cycle.2
Another reason why changes in the underlying
trend of the unemployment rate appear unrelated
to the business cycle is the influence of microeconomic factors. These include changes in the benefits
associated with being unemployed, changes in the
demographics of the labor force, and cultural changes
in family structure and work habits. Regarding the
latter two factors, the large increase in the unemployment rate in the late 1960s and 1970s was associated with a growing number of young workers and
women entering the labor force. And since the unemployment rates for young workers and women
were higher than the average, this change in the
composition of the labor force was associated with
a rising trend in the unemployment rate. In the
1990s, as the baby boomers aged (fewer young
people entering the labor force) and the labor force
16
M AY / J U N E 2 0 0 2
The establishment survey, also known as the
Current Employment Statistics (CES) program,
includes labor input information from about 350,000
nonagricultural establishments that employ about
39 million people. (Establishments are not the same
as firms, but rather, they are distant parts of a firm
in different locations. For example, the Federal
Reserve Bank of St. Louis is a firm with establishments in St. Louis, Little Rock, Louisville, and
Memphis.) The time series we use on payroll jobs
and hours worked come from the CES. The data on
employment growth from the CES are considered
to be more accurate than the data from the CPS
because the establishment survey has much greater
coverage. Although the establishments surveyed are
not representative, they nonetheless are the largest
establishments and account for about 30 percent
of the workforce (compared with 0.06 percent for
the household survey).3
Figure 2 shows the four-quarter growth rate in
jobs as well as the four-quarter growth rate in GDP.
As can be seen visually in Figure 2, and statistically
in Table 1, there is a much closer correlation between
jobs growth and GDP growth than there is between
the unemployment rate and GDP growth. The correlation between the four-quarter growth of real GDP
and nonfarm payroll jobs is high, 0.79. Here the
cycles appear to coincide. However, there are sustained periods of productivity growth during which
the economy grows faster than the work force. Most
obvious in the chart are the decade of the 1960s
and the five years following 1995. It appears that
these periods of high productivity growth tend to
occur during expansions.
The other major series that comes from the
establishment survey is the index of hours worked.4
2
Further evidence of this assertion is that the average duration of
unemployment is included in the Conference Board’s list of lagging
indicators. Its weight places it seventh out of seven in terms of its
contribution to the index.
3
The results from this large sample are adjusted for the bias that exists
between the composition of the approximately 350,000 large establishments surveyed and the composition of the roughly five million smaller
establishments that are not included in the survey. This bias adjustment
process, as it is known, is being replaced with a completely different
methodology. See Getz (2000).
4
The index of aggregate hours worked is the product of average weekly
hours and employment of production or nonsupervisory workers.
See BLS Handbook of Methods.
FEDERAL RESERVE BANK OF ST. LOUIS
Gavin and Kliesen
Figure 2
Figure 3
Nonfarm Payroll Jobs and Real GDP
Nonfarm Hours Worked and Real GDP
Percent
Percent
12
12
10
10
Real GDP
8
6
6
4
4
2
2
Jobs
0
–2
–4
–4
–6
–6
1974
Hours
0
–2
–8
1965
Real GDP
8
1983
1992
–8
1965
2001
NOTE: Both payroll jobs and real GDP are shown as fourquarter growth rates. Bars indicate periods of recession.
1974
1983
1992
2001
NOTE: Both hours worked and real GDP are shown as fourquarter growth rates. Bars indicate periods of recession.
Table 1
Cross Correlations Between Real GDP Growth and Growth of Labor Market Variables
Continuing claims
Real GDP
Continuing
claims
Real GDP
1
–0.82
1
Hours
Initial claims
Hours
–0.86
Initial claims
Jobs
Unemployment
rate
0.90
–0.70
0.14
0.87
–0.79
0.79
–0.27
1
–0.69
0.94
–0.37
1
Jobs
–0.50
1
Unemployment rate
0.05
–0.48
1
NOTE: Correlations of four-quarter growth rates except for the unemployment rate, which is in levels.
Figure 3 shows that growth in hours worked also
moves closely with output growth over the business
cycle. Indeed, among the labor variables cited earlier,
the cross-correlations reported in Table 1 show that
the growth of hours worked has the highest correlation with the growth of real GDP (0.87 over the
sample period in Figure 3). Like jobs growth, the
movement in the growth of aggregate hours is procyclical and appears to coincide with output growth.
One of the reasons why the monthly labor
report from the establishment survey is considered
so important is because it provides early information
about GDP growth. To understand why this is so,
note that a given month’s report is released on the
first Friday of the next month. For example, on the
first Friday of each January, the Department of Labor
will release information about the labor market in
the previous December. The market will already
have received labor market data for October and
November. Labor data for the fourth quarter will
be available in the first week of January, but the
Department of Commerce will not release the
advance estimate of fourth-quarter GDP growth
until the last week of January.
M AY / J U N E 2 0 0 2
17
REVIEW
Gavin and Kliesen
Figure 4
Figure 5
Inital Claims and Real GDP
Continuing Claims and Real GDP
Percent
Percent
12
80
100
60
Real GDP
8
40
4
20
2
0
0
–2
–20
Initial Claims
–4
–60
1974
1983
1992
2001
NOTE: Both initial claims and real GDP are shown as fourquarter growth rates. Bars indicate periods of recession.
The initial release of payroll jobs and hours
worked is based on the establishment survey. According to the BLS, the most recent two months of estimates from the establishment data are considered
preliminary because not all of the surveys have been
returned and processed. Conceivably, then, the BLS
may report up to three different estimates (current
plus two subsequent revisions) of nonfarm job gains
or losses for any month. But even these are still only
preliminary, since the data on jobs and hours will
be revised the following year with the annual benchmark revisions. The purpose of the benchmark
revisions is to tie together the sample-based estimates
that underpin the monthly establishment data with
the actual “universe” counts of jobs, wages, and
earnings that are reported to employment security
agencies of the 50 states, the District of Columbia,
Puerto Rico, and the Virgin Islands. Thus, the third
source of information about the labor market is that
reported to the Department of Labor by the state
agencies that administer the federal-state unemployment insurance program.
Covered Employment and Wages
Program
This program, also known as the ES-202 program,
is a joint venture between the BLS and the state
employment security agencies. The purpose of the
M AY / J U N E 2 0 0 2
60
6
4
40
2
20
0
0
–2
–20
–4
–40
–6
80
Real GDP
8
6
18
12
10
10
–8
1965
Percent
Percent
Continuing
Claims
–6
–8
1965
–40
–60
1974
1983
1992
2001
NOTE: Both continuing claims and real GDP are shown as fourquarter growth rates. Bars indicate periods of recession.
program is to provide a comprehensive accounting
of nonagricultural employment and wage data by
industry at the national, state, and local levels. Thus,
coverage under the ES-202 program is nearly universal. In 1994, more than 96 percent of all wage
and salary civilian jobs were covered by the ES-202
program, while covered employees accounted for
nearly 93 percent of the wage and salary component
of national income. Those excluded from the program coverage include agricultural workers, the
military, and segments of state and local government
employees.
This statewide information is aggregated to
the national level by the ETA. Each week, the ETA
releases statistics for the number of individuals filing
new or continuing claims under the unemployment
insurance (UI) program. The UI program is a joint
arrangement between the federal government and
individual state governments. Its purpose is to provide temporary unemployment benefits to eligible
recipients. Though there are some common characteristics, each state operates under its own laws and,
accordingly, sets program eligibility requirements.
See the appendix for more detail on the program
and its eligibility requirements.
Figure 4 shows the widely reported series, initial
claims for unemployment insurance. Because growth
in initial claims is much more variable than real GDP,
their growth rates are shown on the right-hand
FEDERAL RESERVE BANK OF ST. LOUIS
scale in Figure 4. Initial claims are clearly countercyclical. Table 1 shows a high negative correlation
between the four-quarter growth rates of the two
series, –0.79. Despite this high correlation, the
National Bureau of Economic Research does not
place much weight on initial claims when it comes
to determining business cycle peaks and troughs.
In a question-and-answer section in the on-line issue
of The NBER’s Recession Dating Procedure dated
October 8, 2001, the following was posted:
Q: How do the movements of unemployment
claims inform the Bureau’s thinking?
A: A bulge in jobless claims would appear to
forecast declining employment, but we don’t
use forecasts and the claims numbers have
a lot of noise.
The weekly initial claims report also includes
a series on those individuals that continue to draw
unemployment compensation, otherwise known as
continuing claims. Figure 5 shows the four-quarter
growth of continuing claims and real GDP. Continuing claims are also much more variable than GDP,
and their growth rates are shown in Figure 5 on the
right-hand scale. Like initial claims, continuing
claims are also countercyclical. Visually, it is difficult
to distinguish the co-movement between initial and
continuing claims, although growth of the latter
appear to vary less. Either way, their correlations
with GDP growth are virtually identical, as seen in
Table 1.
THE LABOR MARKET AND GDP
From both a theoretical and empirical standpoint, the labor market is an important element in
the economy. Output production requires combinations of labor, land, capital, and other factors. About
two-thirds of the payments for factors go to the labor
component. From the point of view of data collection, perhaps our best measure of economic activity
is a measure of the number of people working.
As we saw in Figures 1 through 5, labor market
indicators move in tandem with output over the
business cycle. There is a considerable literature
showing that monthly data in general, and labor
data in particular, can be used to predict current
quarter GDP. Miller and Chin (1996) survey this literature and report their own research showing that
monthly information about hours worked helps
predict GDP growth. Recently, Koenig, Dolmas,
and Piger (2001) reported that monthly employment
Gavin and Kliesen
growth is a significant predictor of current-quarter
GDP growth. Some private sector economists have
even developed a “real time” model of aggregate
economic activity that uses both initial and continuing claims to predict monthly changes in real GDP.5
Accordingly, monthly data should be able to predict
quarterly GDP because the Bureau of Economic
Analysis (BEA) uses monthly labor market data as
an input to formulas that are used to estimate GDP
components.
We evaluate the predictive content of monthly
labor market data by adding these variables one at
a time to a univariate autoregressive model of real
GDP growth. We construct quarterly time series of
incoming monthly labor market data. We forecast
the current-quarter real GDP growth rate using
incoming monthly labor market data from the same
quarter. The general form of the forecasting model is
4
(1)
yt = c + β j LM kj , t + ∑ δ i yt − i + ε t ,
i =1
where yt=ln(GDPt /GDPt –1) ×400, the average annualized growth rate of GDP in the current quarter, and
LM kj is one of five labor market variables measured
at the end of each of the three months in the quarter.
The five labor market variables indexed by k include
the unemployment rate and the annualized growth
rates of payroll jobs, aggregate hours worked, initial
claims for unemployment insurance, and continuing claims for unemployment insurance. The labor
market variable is indexed by j to indicate which
month of the current quarter is being used in the
forecast. For example, at the end of the first month,
1
LM1,t
is the newly reported unemployment rate; at
the end of the second month, LM12,t is the average
of unemployment rates for the first two months of
the quarter; and, at the end of the third month, LM13,t
is the average of the three months of quarter t. Remember that the labor report for the third month
of a quarter arrives three to four weeks before the
first GDP report for that quarter. We consider each
labor market variable separately. We also include
four lags of GDP growth.
We use current vintage data in this forecasting
experiment.6 In this experiment, we begin by estimating a model using data from 1967:Q2 through
5
See Hatzius (2001).
6
Koenig, Dolmas, and Piger (2001) have shown that it is possible to
get a better forecast by using real-time vintage data. They show analytically that, if the revisions to data are not predictable, then the real-time
vintage data will yield a forecast model with a smaller out-of-sample
forecast error than one would get using current vintage data.
M AY / J U N E 2 0 0 2
19
REVIEW
Gavin and Kliesen
1991:Q3.7 This model is then used to forecast the
fourth quarter of 1991. We then update the forecasting model with 1991:Q4 data and use the newly
estimated model to forecast the second quarter.
That is, we update the model recursively and tabulate
the forecasts through 2001:Q3. We then calculate
the root-mean-squared error (RMSE) for that model
forecast.
We examine six models. The first is simply the
autoregressive (AR) component of the model excluding the labor market variable. The next five correspond to the labor market variables. The civilian
unemployment rate is measured in level form,
whereas the remaining four variables are measured
as annualized growth rates. Note that in the case of
these four variables, we calculate the first month’s
growth rate by taking the log ratio of the variable in
the first month to the average of the three months
in the previous quarter. In the second month we take
the log of the ratio of the average of the first two
months in the current quarter to the average of the
three months in the previous quarter. In the third
month we just look at the ratio of the three-month
averages. In all cases involving variables in the GDP
forecasting equation, we annualize growth rates.
To assess the statistical significance of the accuracy of the alternative model forecasts, we use two
tests developed for nested forecasting models. In
each case, we compare a model with lags of GDP
growth and a labor market variable with a model
that includes only lags of GDP growth. First, we use
an out-of-sample F test of the null hypothesis that
the model with the labor market variable has no
predictive content for real GDP growth once the
autoregressive model is taken into account. This
test, developed by McCracken (1999), is given by
MSEAR − MSE LM k
OOS-F = P
,
MSE LM k
where OOS-F is the out-of-sample F test, P is the
number of forecasts made, MSEAR is the meansquared error for the AR model forecasts, and MSELM
is the mean-squared error for the models that
include labor market variables. McCracken derives
the limiting distribution of this test statistic under
the null hypothesis and reports percentiles of the
OOS-F statistic. He derives tables under alternative
methods of updating the forecasting models. We use
a recursive scheme. The critical values of the test
statistic depend on which scheme is used and two
other factors: (i) the number of labor market variables
20
M AY / J U N E 2 0 0 2
included (in each case we have one in each model)
and (ii) the ratio (P/R) of the number of forecasts (P)
to the number of observations used to estimate the
model that was used to make the first forecast (R).
Percentiles of the distribution are listed in the tables.
We are comparing nested models so we use a onesided test. When the MSE of the forecasts from the
unrestricted model is larger than the MSE from the
restricted model, this test statistic is negative.
The second test we use is an out-of-sample test
for encompassing. (Encompassing is simply that, if
one forecast incorporates all of the relevant information, then adding information from the other forecast will not help predict the actual value.) We use
an encompassing test of the null hypothesis that
the AR model encompasses the model augmented
with the labor market variable. This test, developed
by Clark and McCracken (2000), is given by
MSE − MCPE
AR
ENC-CM = P
,
MSE
k
LM
where ENC-CM is the encompassing test proposed
by Clark and McCracken (2000) and MCPE is the
mean cross product of the forecast errors from the
restricted (AR) and unrestricted (LM) models.8 Clark
and McCracken derive the limiting distribution of
this test statistic under the null hypothesis and
report percentiles of the ENC-CM statistic. As in the
case of the OOS-F statistic, the limiting distribution
depends on the method used to update the forecasting models, the number of parameters restricted to
zero, and the ratio, P/R. The percentiles of the distribution are shown in Table 2. Again, we are comparing nested models, so we use a one-sided test.
The statistic will be negative only if the average
cross product is positive and larger than the meansquared error of the forecasts from the AR model.
The results of our evaluation are shown in
Table 2. The first column reports results for the AR
model (which excludes contemporaneous labor
market data). This is the benchmark model and is
nested in all the others. The RMSE of the forecast
from the AR model for the period from 1990:Q1 to
2001:Q3 is 2.17 percent, with an adjusted R2 of 6
percent for the last model estimated—that is, estimated over the period from 1967:Q2 to 2001:Q2.
7
Except in the case where the labor market variable is continuing claims.
These data begin in January 1968.
8
The MCPE is calculated by the following formula:
1 T
∑ eit e jt .
T t =1ˆ ˆ
FEDERAL RESERVE BANK OF ST. LOUIS
Gavin and Kliesen
Table 2
Evaluation of GDP Forecasts (1990:Q1 through 2001:Q3)
AR (4)
model
Unemployment
rate
Payrolls
jobs
Hours
worked
Initial
claims
Continuing
claims
2.17
2.17
1.81
1.95
2.18
2.23
Second
2.17
1.81
1.76
2.03
2.07
Third
2.17
1.78
1.66
1.93
1.93
0.06
0.57
0.63
0.41
0.47
First month
–0.31
20.32
10.75
–0.40
–2.64
Second
–0.28
20.31
24.34
6.39
4.48
Third
–0.28
22.29
32.65
12.48
12.25
First month
0.11
26.75
24.42
6.38
6.44
Second
0.00
31.49
35.05
13.98
12.17
–0.07
34.60
39.86
20.65
17.87
RMSEs (out-of-sample forecasts)
First month
2*
Adjusted R
0.06
McCracken out-of-sample F test†
Clark-McCracken nested encompassing test‡
Third
NOTE: *Adjusted R2 is for the full 1967:Q1 to 2001:Q2 sample using the 3-month models.
†The null hypothesis is that the AR model is more accurate than the model with the labor market variable. Here P/R=0.52. For P/R =0.4,
the 99th, 95th, and 90th percentiles for the OOS-F tests are 2.768, 1.298, and 0.814, respectively; and for P/R =0.6, the 99th, 95th, and
90th percentiles for the OOS-F tests are 3.719, 1.554, and 0.796, respectively.
‡The null hypothesis is that the AR model encompasses the model with the labor market variable. Here P/R=0.52. For P/R=0.4, the 99th,
95th, and 90th percentiles for the Clark-McCracken encompassing tests are 2.098, 1.079, and 0.685, respectively; and for P/R =0.6, the
99th, 95th, and 90th percentiles for the encompassing tests are 2.662, 1.312, and 0.791, respectively.
Bold values indicate that the null hypothesis is rejected at the 99th percentile.
The next column shows the results using the unemployment rate. As was suggested by Figure 1,
changes in the unemployment rate since 1990 do
not appear to help predict current-quarter real GDP
growth. The explanatory power of unemployment
was no better than with GDP alone, and the out-of
sample forecasts were slightly worse, although the
difference is small. The OOS-F statistics in the middle
section of Table 2 are all negative, thus we cannot
reject the hypothesis that the AR model is more
accurate than the model that includes the unemployment rate. The same is true for the ENC-CM statistics.
They are all below the 90th percentile value.
The next two columns show the models using
growth in jobs and an index of hours worked. Here,
there appears to be predictive information in the
growth of payroll employment in all three months.
Note, however, that the addition of the second month
does not result in a lower RMSE for the model that
includes payroll jobs. For the aggregate hours model,
adding information from the second and third
months lowers the RMSEs. These models also display a much higher in-sample explanatory power
than does the model that includes the unemployment rate. For the models that include payroll jobs
and hours worked, the OOS-F tests always reject
the hypothesis that the AR model is more accurate.
We can also reject the hypothesis that the AR model
encompasses these models.
Finally, the two series using the unemployment
insurance data lead to lower RMSEs only in the cases
with two and three months of claims data included.
Here the difference is large enough so that we can
reject the null hypothesis that the AR model is more
accurate than the models that include two or three
months of claims (both initial and continuing). The
RMSEs for the models that use initial or continuing
claims from the first month only are generally higher
than the RMSE from the AR model. The adjusted R2
values for the models with three months of initial
and continuing claims are 0.41 and 0.47, respectively.
Looking at the encompassing tests in the bottom
M AY / J U N E 2 0 0 2
21
REVIEW
Gavin and Kliesen
section of Table 2, we see that we can reject the
hypothesis that the AR model encompasses the
model augmented with initial claims, even when
the RMSEs are larger than the benchmark case.
In summary, we find that—consistent with previous empirical research—labor market data does
help to predict GDP growth. In the next section, we
examine the ability of weekly data on initial and
continuing claims to predict the monthly time series
on unemployment, payroll jobs, and the index of
hours worked.
PREDICTING MONTHLY LABOR
MARKET DATA USING UNEMPLOYMENT
INSURANCE DATA
In the previous section we saw that initial and
continuing claims for unemployment insurance
are not very useful for predicting real GDP growth
during the concurrent quarter. However, data on
monthly employment and hours worked did help
to predict GDP growth. Therefore, it would be useful
to be able to predict employment and hours worked
using the weekly claims data. Furthermore, many
economists and financial analysts use weekly claims
data to predict monthly changes in the unemployment rate. The payoff from this exercise is potentially quite large, since unexpected changes in the
unemployment rate can be a significant market
mover; moreover, these changes can sometimes
induce immediate changes in monetary policy.9 A
typical example of the analysis that posits a causality
between unemployment insurance claims and the
unemployment rate may be found in the following
Monetary Policy Report to the Congress:
Employment continued to decline in
December and January but much less than
in the preceding two months. Manufacturing and its related industries lost jobs at a
slower pace, and employment leveled off
in other private industries. The unemployment rate moved up to 5.8 percent in
December but then ticked down to 5.6 percent in January. The recent reversal of the
October and November spikes in new claims
for unemployment insurance and in the
level of insured unemployment also point
to some improvement in labor market conditions early this year. (Board of Governors
of the Federal Reserve System, February
2002, p. 20)
22
M AY / J U N E 2 0 0 2
A recent study by Montgomery et al. (1998) uses
monthly initial claims data to forecast the quarterly
unemployment rate. The study finds some support
for the predictive content of monthly initial claims.10
The contribution of initial claims was concentrated in periods when unemployment was rising.
McConnell (1998) reports a similar finding in which
she uses initial claims data to forecast payroll jobs
growth. In her study, initial claims helped to predict
payroll jobs, but only during periods of recession.
In this study we are looking at the ability of the
weekly data to predict the monthly series: not only
the unemployment rate, but the jobs and hours
worked data as well.
We use a model analogous to equation (1) to
evaluate the ability of the unemployment insurance
claims data to predict the monthly labor statistics:
12
(2)
LMtk = c + β jWeekly aj , t + ∑ δ i LMtk− i + ε t ,
i =1
where the dependent variable is one of three
monthly labor market series: the unemployment
rate, growth in payroll jobs, and growth in the index
of hours worked. Here, the growth rates are monthly.
There are two alternative weekly series used on the
right-hand side of equation (2): initial claims and
continuing claims. The data on initial claims are
released on Thursdays and apply to the previous
week that ended five days earlier. The data on continuing claims released at the same time apply to
the week that ended 12 days earlier.
We create five monthly series from each of these
latter two data series, initial and continuing claims.
The first weekly series is the data reported on the
first Thursday following the first Friday of the
month, the normal release date for the Employment
Situation. We take the logarithm of the ratio of this
weekly release to the average for the previous month.
The second weekly series is the logarithm of the
ratio of the average data reported on the first and
second Thursdays (following the first Friday of the
month) to the previous month’s average, and so
forth. We do not create a fifth series because there
is not always a fifth Thursday. Instead, we create a
series that we call the last week, which includes the
9
See Jordan (1992).
10
They started with seasonally adjusted data, but as usual in these
time-series models, they had to include the seasonal terms to get rid
of the residual correlation. We briefly examined ARIMA and Bayesian
VAR methods, but, overall, none generate more accurate forecasts
than the univariate regressions reported herein.
FEDERAL RESERVE BANK OF ST. LOUIS
Gavin and Kliesen
Table 3
Regression Output for Period from February 1968 to November 2001
Initial claims
β
t Statistic
Continuing claims
SEE
β
t Statistic
SEE
Unemployment rate
AR only
—
—
0.161
—
—
0.161
First week
0.003
2.11
0.161
0.021
4.27
0.158
Second
0.007
3.81
0.159
0.038
8.29
0.149
Third
0.008
4.92
0.157
0.038
9.27
0.146
Fourth
0.009
5.28
0.156
0.036
10.01
0.144
Last
0.009
5.36
0.156
0.036
10.33
0.143
—
—
0.171
—
—
0.171
Payroll jobs
AR only
First week
–0.006
–4.02
0.168
–0.030
–5.93
0.164
Second
–0.012
–6.82
0.162
–0.043
–8.95
0.156
Third
–0.014
–7.87
0.159
–0.045
–10.95
0.150
Fourth
–0.014
–8.20
0.158
–0.044
–12.34
0.145
Last
–0.014
–8.38
0.157
–0.043
–12.88
0.143
Hours worked
AR only
—
—
0.477
—
—
0.477
First week
–0.012
–2.63
0.474
–0.061
–4.13
0.468
Second
–0.024
–4.89
0.464
–0.087
–6.20
0.456
Third
–0.031
–6.17
0.456
–0.097
–7.86
0.444
Fourth
–0.034
–7.15
0.449
–0.103
–9.74
0.429
Last
–0.034
–7.32
0.448
–0.102
–10.09
0.426
NOTE: The values in the table are estimates of βj , its t statistic, and the standard error for equation (2).
average of the data released in the first four weeks
when there is no fifth week available.
The estimation results for the weekly initial
claims models using the full data set are shown in
Table 3. The estimation period includes the months
from February 1968 through November 2001. We
estimated OLS models for the three labor market
variables. Each model included a constant and 12
lags of the dependent variable as well as our weekly
series that use information about unemployment
insurance claims. There are three sections in Table 3.
The top section shows the results for the unemployment rate. The first row reports the standard error
of the equation (SEE) for the autoregressive (AR)
model which excludes the claims data. The estimate
of the coefficient on the weekly initial claims data
is reported in the first column of results with the t
statistic for that coefficient reported in the second
column. The third column reports the SEE for the
equation. The last three columns report the analogous results for continuing claims. The middle section includes the results for payroll jobs, and the
bottom section reports the results for hours worked.
Overall, the in-sample fit improved with the accumulation of information throughout the month.
Uniformly, the data on continuing claims do a better
job of predicting the labor market variables than
do the initial claims data. This condition is true in
spite of the extra-week delay in reporting information about continuing claims.
An Out-of-Sample Forecasting Exercise
To evaluate the predictive content of the claims
data, we conduct an out-of-sample forecasting
M AY / J U N E 2 0 0 2
23
REVIEW
Gavin and Kliesen
Table 4
Evaluation of Monthly Forecasts of Labor Market Indicators (Current Month Forecasts from
January 1990 to November 2001)
Unemployment rate
Initial
claims
Continuing
claims
Payroll jobs
Initial
claims
Continuing
claims
Hours worked
Initial
claims
Continuing
claims
RMSE (% at monthly rates)
AR model
0.135
0.106
0.371
First week
0.134
0.137
0.103
0.105
0.369
0.369
Second
0.133
0.136
0.105
0.108
0.374
0.371
Third
0.132
0.131
0.103
0.102
0.371
0.362
Fourth
0.133
0.130
0.102
0.097
0.371
0.357
Last
0.133
0.130
0.102
0.097
0.372
0.357
8.68
1.30
1.19
1.87
McCracken out-of-sample F test*
First week
1.64
–4.17
Second
4.37
–2.41
1.04
–5.52
–1.90
–0.29
Third
5.02
6.97
6.43
9.25
–0.07
6.98
Fourth
3.52
10.09
11.85
25.32
0.22
11.37
Last
2.89
9.18
10.11
25.65
–0.37
11.29
6.20
Clark-McCracken nested encompassing
test†
First week
1.87
2.52
9.03
16.30
1.85
Second
5.02
11.45
17.62
22.47
4.56
9.28
Third
7.39
18.60
23.64
36.99
7.80
15.84
Fourth
8.09
20.62
31.00
52.32
11.61
22.57
Last
8.05
20.50
31.74
53.58
12.35
22.88
NOTE: *The null hypothesis is that the AR model is more accurate than the model with the labor market variable. Here P/R =0.58. For
P/R =0.6, the 99th, 95th, and 90th percentiles for the OOS-F tests are 3.719, 1.554, and 0.796, respectively.
†The null hypothesis is that the AR model encompasses the model with the labor market variable. Here P/R =0.58. For P/R =0.6, the
99th, 95th, and 90th percentiles for the Clark-McCracken encompassing tests are 2.662, 1.312, and 0.791, respectively.
Bold values indicate that the null hypothesis is rejected at the 99th percentile.
experiment. Again, we are using current vintage
data to construct these out-of-sample forecasts. We
begin by estimating the model over the period from
February 1968 through December 1989. As before,
we update the model each month before we make
the next forecast, recursively computing the forecasts through November 2001. The RMSEs of the
forecasts are reported in the top section of Table 4.
The first row of results include the RMSEs from
the forecasts made by the autoregressive models.
Here the sample period includes the months from
January 1990 through November 2001.
The out-of-sample forecasting results are not
entirely consistent with the in-sample fit where
continuing claims always outperformed initial
24
M AY / J U N E 2 0 0 2
claims. Here, initial claims appears to do a better
job forecasting the unemployment rate and payroll
job growth early in the month and continuing claims
does better late in the month. In the second and
third sections of Table 4, we report the out-of-sample
tests for equality of MSEs and encompassing, respectively. Again, we compare the forecasts from the
full model with forecasts from the AR model that
is nested within each of the full models. Therefore, we use the tests for nested models that were
described above. The forecasting method was recursive, there is one restriction on the AR model, and
the P/R ratio for this experiment is 0.58, so we use
the percentiles for P/R=0.6 where the 99th percentile for the OOS-F test statistic is 3.719.
FEDERAL RESERVE BANK OF ST. LOUIS
Gavin and Kliesen
Table 5
Evaluation of Monthly Forecasts of Labor Market Indicators (Current Month Forecasts from
April 1991 to February 2001—Expansion Months Only)
Unemployment rate
Initial
claims
Continuing
claims
Payroll jobs
Initial
claims
Continuing
claims
Hours worked
Initial
claims
Continuing
claims
RMSE (% at monthly rates)
AR (12) model
0.128
0.095
0.374
First week
0.127
0.132
0.092
0.097
0.371
0.373
Second
0.126
0.133
0.097
0.103
0.375
0.379
Third
0.126
0.128
0.097
0.100
0.374
0.375
Fourth
0.127
0.128
0.097
0.095
0.373
0.369
Last
0.128
0.129
0.097
0.094
0.373
0.369
McCracken out-of-sample F test*
First week
1.13
–6.95
6.82
–6.05
1.93
0.70
Second
3.04
–8.64
–5.82
–19.20
–0.63
–3.49
Third
3.24
–1.29
–4.88
–11.43
0.06
–0.89
Fourth
0.56
–0.93
–4.64
–0.50
0.13
3.32
Last
0.19
–1.84
–6.05
0.76
0.13
3.18
1.25
–8.31
–5.94
1.71
3.93
Clark-McCracken nested encompassing
test†
First week
2.08
Second
5.86
8.69
0.24
–0.55
3.54
6.11
Third
7.37
13.60
4.20
8.14
6.04
10.05
Fourth
7.94
14.37
9.12
17.05
9.28
15.05
Last
7.92
14.26
9.25
18.96
9.80
15.12
NOTE: *The null hypothesis is that the AR model is more accurate than the model with the labor market variable. Here P/R =0.36. For
P/R =0.4, the 99th, 95th, and 90th percentiles for the OOS-F tests are 2.768, 1.298, and 0.814, respectively.
†The null hypothesis is that the AR model encompasses the model with the labor market variable. Here P/R =0.36. For P/R =0.4, the
99th, 95th, and 90th percentiles for the Clark-McCracken encompassing tests are 2.098, 1.079, and 0.685, respectively.
Bold values indicate that the null hypothesis is rejected at the 99th percentile.
Using a 1 percent critical region, the F tests
reject the null hypothesis that the AR forecast of
the unemployment rate is better than the initial
claims forecast for the second and third weeks, but
not for the fourth and last weeks. This hypothesis is
rejected for the continuing claims data in the models
where at least three weeks of data are available. We
can reject the null hypothesis for payroll jobs as well,
when we can use initial claims data for all but the
model with the first two weeks of data. As we found
with the unemployment rate, the null hypothesis is
rejected in the cases using at least three weeks of
continuing claims data. We cannot reject the null
hypothesis in the case of hours worked for initial
claims, but we can for cases including three or more
weeks of continuing claims data. The encompassing
tests are reported in the bottom panel of Table 4. In
all but a few cases involving models with just the
first-week data, we can reject the null hypothesis
that the AR model encompasses the models including the claims variables at the 99th percentile.
Our forecasting period included the recession
that began in July 1990 and ended in March 1991,
as well as the first nine months of the current recession. Both Montegomery et al. (1998) and McConnell
(1998) conclude that initial claims data can forecast
labor market variables, but only in times of recession
and rising unemployment. Therefore we calculated
the forecasting performance of these models during
the 10 years of expansion from April 1991 through
M AY / J U N E 2 0 0 2
25
Gavin and Kliesen
February 2001. These results are reported in Table 5.
Looking at expansion months only, we find much less
information in the claims data. However, there is still
some evidence that initial claims data help to predict
the unemployment rate and continuing claims data
help to predict growth in hours worked. Again, even
though the AR model often had a lower RMSE, we
could always reject the hypothesis that the AR model
encompassed the model that included the claims
data when we used at least three weeks of data.
CONCLUSION
Empirical evidence and economic theory suggest
that changes in labor market conditions will have
significant effects on aggregate output. Evidence
presented in this paper further suggests that incoming monthly data on nonagricultural payroll jobs
and the index of aggregate weekly hours help predict
changes in real GDP growth. Changes in the civilian
unemployment rate are less significant. This finding
suggests that predicting monthly changes in jobs
or hours growth would be helpful in predicting real
GDP growth. Many economists and financial market
analysts strive to do this by tracking initial claims
for state unemployment insurance benefits, which
are released weekly. This article has shown that
there is some statistically significant marginal information in the unemployment insurance claims data,
even during periods of expansion. However, information about continuing claims appears to be at
least as important as the information about initial
claims that usually appears in the headlines.
REFERENCES
Board of Governors of the Federal Reserve System.
Monetary Policy Report to the Congress. February 2002.
Clark, Todd E. and McCracken, Michael W. “Tests of Equal
Forecast Accuracy and Encompassing for Nested Models.”
26
M AY / J U N E 2 0 0 2
REVIEW
Working Paper RWP 99-11, Federal Reserve Bank of
Kansas City, November 2000.
Getz, Patricia M. “Implementing the New Sample Design
for the Current Employment Statistics Survey.” Business
Economics, October 2000, 35(4), pp. 47-50.
Hatzius, Jan. “Jobless Claims Imply Contraction, but Only
at a Glacial Pace.” Goldman Sachs Economics Goldman
U.S. Daily, 15 June 2001.
Jordan, Jerry L. “What Monetary Policy Can and Cannot
Do.” Federal Reserve Bank of Cleveland Economic
Commentary, 15 May 1992.
Koenig, Evan F.; Dolmas, Sheila and Piger, Jeremy M. “The
Use and Abuse of ‘Real-Time’ Data in Economic Forecasting.” Working Paper 2001-015A, Federal Reserve Bank
of Dallas, 2001.
Mankiw, N. Gregory. Principles of Economics. Orlando, FL:
Harcourt Brace, 1998.
McConnell, Margaret M. “Rethinking the Value of Initial
Claims as a Forecasting Tool.” Federal Reserve Bank of
New York Current Issues in Economics and Finance,
November 1998, 4(11), pp. 1-6.
McCracken, Michael W. “Asymptotics for Out of Sample
Tests of Causality.” Unpublished manuscript, Louisiana
State University, November 1999.
Miller, Preston J. and Chin, Daniel M. “Using Monthly Data
to Improve Quarterly Model Forecasts.” Federal Reserve
Bank of Minneapolis Quarterly Review, Spring 1996, 20(2),
pp. 16-33.
Montgomery, Alan L.; Zarnowitz, Victor; Tsay, Ruey S. and
Tiao, George C. “Forecasting the U.S. Unemployment
Rate.” Journal of the American Statistical Association,
June 1998, 93(442), pp. 478-93.
FEDERAL RESERVE BANK OF ST. LOUIS
Gavin and Kliesen
Appendix
METHODOLOGY OF THE
UNEMPLOYMENT INSURANCE
CLAIMS DATA
Data Series and Sources
Each week, state government employment
offices report the number of individuals filing
claims for unemployment insurance benefits. The
state offices then report the figures to the Office
of Workforce Security in the ETA.11 They are then
published in the Unemployment Insurance Weekly
Claims Report, which is issued by the ETA. Also
published in this report are continuing claims for
state unemployment insurance benefits (insured
unemployment), which is another closely monitored indicator.
Eligibility Requirements
Individuals who file for unemployment insurance benefits are not automatically eligible for
benefits. To qualify for benefits, the worker must
first demonstrate that they have a work history,
otherwise known as an “attachment to the labor
force.” In most states, this requirement is met by
having earned a minimum amount of money in a
job that is covered by the law. In some states, a
person is eligible if they have merely worked a
minimum amount of time in covered employment.
Covered employment excludes self employment,
small farms, and small domestic operations. Once
the person is deemed monetarily eligible, the reason for the claim is then examined. Although a
common reason stems from an unintended loss
of employment, some states disburse benefits to
individuals who are following a spouse to a new
job. If an unfavorable ruling results, the claimant
may appeal the decision.
Waiting Period Requirements
In general, individuals do not receive benefit
checks for two to three weeks after they are classi-
fied as eligible. Moreover, there is an additional
lag in those states that have a one-week waiting
period. This means that they cannot claim benefits
for that week. Most states require that claimants
file for benefits every two weeks. For every week
a person claims benefits, they are required to be
available and actively seeking work, and, among
other things, they cannot refuse a suitable job.
Type of Claims
The initial claims series that is reported weekly
comprises two types of claims: new and additional.
A new claim is defined as the first initial claim filed
in person, by mail, telephone, or other means to
request a determination of entitlement to and eligibility for compensation. This results in an agencygenerated document to determine monetary
eligibility. An additional claim is a subsequent
initial claim filed (i) during an existing benefit year
due to new unemployment and (ii) when a break
of one week or more has occurred in the claim
series due to intervening employment. Thus, these
claims are reported only when there has been
intervening employment since the last claim was
filed. Claims that follow breaks due to illness, disqualification, unavailability, or failure to report
for any reason other than job attachment are not
reported. Thus, if a person has multiple occurrences
of unemployment during their benefit year, the
first one is counted as a new initial and the others
are counted as additional initials. Both numbers
are incorporated into the published weekly counts
and thus represent new emerging unemployment
for that week.
11
The claims data are not derived from the ES-202 program, which is the
source of employment and wage data by industry at the national, state,
and county levels. Thus, they are not drawn from the sample of data
that is used to construct the establishment data in the monthly
Employment Situation report, nor are they used to calculate the
unemployment rate, which comes from the household survey.
M AY / J U N E 2 0 0 2
27
Gavin and Kliesen
28
M AY / J U N E 2 0 0 2
REVIEW
Did “Right-to-Work”
Work for Idaho?
Emin M. Dinlersoz and Rubén Hernández-Murillo
RIGHT-TO-WORK LAWS AND
ECONOMIC ACTIVITY
he right-to-work (RTW) law ensures that
workers are not forced to join unions or pay
union dues as a condition of employment.1
Despite many years of research, the impact of these
laws on a state’s economic performance is still a
controversial issue. Using a diverse set of data and
methods, a sizeable body of literature has concentrated on understanding whether the passage of
RTW laws matters.2 RTW laws continue to be an
important issue on states’ agendas and a source of
fierce campaigning by pro- and anti-union groups.
For instance, in September 2001, Oklahoma adopted
the RTW law after a lengthy period of campaigns
for and against it.
States with RTW laws usually offer additional
policies as part of a pro-business profile designed
to attract new firms and boost industrial development. This is the view taken by Holmes (1998), who
uses the RTW law as a proxy for the state’s businessfriendly climate. He studies the effects of probusiness policies on economic activity by examining
the performance of manufacturing industries across
state borders where one state has a RTW law and
the other does not. His analysis identifies a large,
positive impact of an overall favorable business
climate, but the effects cannot be traced to any
particular state legislation, such as a RTW law.
Many states passed RTW laws in the mid-1940s
to early 1950s. Since then, except for the 2001 adoption by Oklahoma, only two other states adopted
them: Louisiana in 1976 and Idaho in 1986. Indiana
adopted the law in 1957, but repealed it in 1965. It
is natural to think that economic conditions today
T
Emin M. Dinlersoz is an economist at the University of Houston.
Rubén Hernández-Murillo is an economist at ITAM, Mexico. The
authors thank Gordon Dahl, Roger Sherman, Lori Taylor, and seminar
participants at the October 2001 Federal Reserve System Conference
on Regional Analysis in San Antonio, Texas, for comments and suggestions. Barry Hirsch and David Macpherson provided useful suggestions
to compute the estimates of unionization rates. This article was written
when the authors were conducting research at the Federal Reserve
Bank of St. Louis.
© 2002, The Federal Reserve Bank of St. Louis.
are quite different from those that prevailed during
the earlier period when many states passed the law
en masse. An important question then is whether
the late adopters of this law have experienced any
real benefits.
Idaho’s Case
In this paper, we reassess the economic impact
of the RTW law by focusing on Idaho’s experience.3
Idaho adopted their RTW law in 1986, at a time
when the decline in unionization in the U.S. had
substantially run its course.4 Was the passage of
the law merely a gesture that simply reflected a
trend of decline in unionization, or did it have a
significant influence in making Idaho a more attractive location for business in the years following the
adoption? Our goal is to provide some evidence on
how Idaho’s unionization rate and industrial performance evolved over time, both before and after
the passage of the RTW law, thereby contributing
to the literature on the effect of business-friendly
policies on states’ industrial performance.
One important aspect of Idaho’s experience is
that the passage of the law itself was a long and
controversial process that took nearly two years.
The critical events related to the legislation process
are summarized in Abraham and Voos (2000). The
original bill was introduced to Idaho’s House in
January 1985, and the law was eventually passed
in November 1986, after a lengthy political and
bureaucratic process involving several confrontations between pro-law and anti-law groups, as well
as a veto and several delays. The law finally took
effect in 1987.
A detailed investigation of other business policies
adopted in Idaho around 1987 reveals that there
were no other major changes in Idaho’s business
1
Section 14(b) of the Taft-Hartley Act, passed in 1947 by Congress,
reaffirms states’ rights to pass RTW laws. These laws may or may not
apply to federal workers, depending on the specifics.
2
See Moore and Newman (1985) and Moore (1998) for a comprehensive
review of this literature.
3
Louisiana is also a candidate for such a study. However, the unavailability of long time series data before Louisiana’s adoption year (1976)
prevents the investigation of this case in detail.
4
Goldfield (1987) reports that between 1954 and 1978 the union
membership rate in the United States declined from 34.7 percent to
23.6 percent. See Goldfield (1987) for a comprehensive analysis of
the declining unionization in the United States. According to Hirsch,
Macpherson, and Vroman (2001), the union membership rate declined
from 29.3 percent in 1964 to 24.1 percent in 1977, and then to 13.6
percent in 2000.
M AY / J U N E 2 0 0 2
29
REVIEW
Dinlersoz and Hernández-Murillo
Figure 1
Washington
Oregon
Montana
Idaho
Nevada
Wyoming
Utah
RTW States
NRTW States
climate regarding incentives for new investments
or firm relocation.5
Idaho offers an interesting case study not only
because it is a late adopter, but also because three
of its six neighboring states have had the RTW law
for a long time and three have traditionally been
non-right-to-work (NRTW) law states.6 Figure 1
shows Idaho and its neighbors, which provide potential controls against which to judge Idaho’s performance. Clearly, these states are imperfect controls.
However, among all other states, Idaho’s neighbors
seem to be a natural choice for comparison for
the reason, if nothing else, that we can control for
common region-specific factors that do not vary
over time. Responses to nationwide economic fluctuations vary substantially across regions. Focusing
on a particular region minimizes this problem. In
analyzing the evolution of unionization rates, we also
consider the experience of states with an industry
mix that was similar to that of Idaho to account for
differences arising from the composition of industrial activity.
Our empirical analysis has two main parts. First,
we look at the evolution of the unionization rate
before and after the law. We find that there was a
large decline in unionization between 1981 and
1984, the year before the bill was introduced to the
legislature. The unionization rate then rebounded
somewhat until 1987, the year the law officially
took effect, but continued to decline persistently
thereafter. Idaho’s unionization rate gradually
became very similar to the average unionization
rate of other RTW states with a similar industrial
mix. When we compare Idaho’s unionization rate
also to that of its geographic neighbors, we find
that, particularly in the manufacturing sector,
Idaho’s unionization rate exhibits a significantly
faster decline.
30
M AY / J U N E 2 0 0 2
Second, we investigate the manufacturing
sector’s performance pre- and post-law. We observe
that in the post-law period, Idaho experienced a
significant and persistent annual growth in manufacturing employment and in the number of establishments, as opposed to virtually zero growth in
both of these variables in the pre-law period. The
difference between the pre-law and post-law growth
rates in Idaho was significantly larger compared
with other states in the region. In addition, we find
that the fraction of total manufacturing employment
in large manufacturing establishments increased
significantly in Idaho after the law was passed. The
average size of large manufacturing establishments
also grew substantially in the post-law period.7 Our
observations are consistent with the hypothesis
that Idaho became more attractive for large plants
because of declining unionization.
Overall, our findings indicate that the increase
in Idaho’s industrial growth rate is strongly related
to the decline in unionization. While we are tempted
to associate the patterns observed with the passage
of the law itself, the timing of the decline in the
unionization rate prevents such a definitive conclusion. The large decline in unionization started about
four years prior to the almost two-year-long bureaucratic process that eventually led to the passage of
the law. This prompts us to consider the hypothesis
that the passage of the law might actually have
been a consequence of the decline in unionization
and growing anti-unionism in Idaho, rather than a
cause. Consequently, while the declining unionization appears to be responsible for the strong post-law
growth trends in Idaho, we cannot fully ascribe the
initiation of the trends to the law itself. The passage
of the law, however, seems to have strengthened
and reinforced the trends.
Literature Review
One expects that a first-order effect of the
passage of a RTW law would be a reduction in the
5
We examined, in particular, the Directory of Incentives for Business
Investment and Development in the United States, published by the
National Association of State Development Agencies.
6
The RTW neighbors, Nevada, Utah, and Wyoming, adopted the law in
1951, 1955, and 1963, respectively. The time period between these
years and our first observation year (1975) is long enough to give us
some comfort that the potential effects of the RTW law must have
already been realized to a large extent in these states.
7
In general, larger establishments are more likely to be unionized and,
therefore, have more incentives to avoid unions. See Long (1993),
Galarneau (1996), and Lowe (1998) for evidence on this in Canada.
FEDERAL RESERVE BANK OF ST. LOUIS
union membership rate. There are several reasons
why this might be the case. As Ellwood and Fine
(1987) point out, the most obvious reason is that
the passage of the law makes unions less attractive
to workers because unions no longer have the ability
to enforce payments and fines. These effects depress
new union organizing and also deter the replacement of decertified unions. If a state’s labor force is
growing, then less union organizing means also a
reduction in the union membership rate.
Most earlier studies, surveyed by Moore and
Newman (1985) and Moore (1998), found a weak
relationship between the passage of RTW laws and
the level of the union membership rate. However,
this does not mean that unionization activity was
not influenced by RTW laws. Using 1951-77 data for
50 states on new union organizing activity (a measure of new membership flow into unions, rather
than the level of unionization), Ellwood and Fine
(1987) presented convincing evidence that the
passage of RTW laws led to a decline in new union
organizing of about 46 percent for the first five years
after the legislation and 30 percent during the next
five. This reduction in organizing disappears after a
decade. The level of union membership, as a result,
declines in most states by about 5 to 10 percent after
the 10 years, which may not have been detected
by the econometric methods used in the previous
studies. Further tests reveal that these findings are
robust to time-invariant differences across states.
Idaho’s experience provides a natural setting to
further assess the evolution of the union membership rates before and after the passage of the law.
Since we are looking at the same state both before
and after, time-invariant state-specific factors should
be irrelevant for the pattern of evolution in the union
membership rates.
As we mentioned before, an important concern
is whether declining union strength is a catalyst
for the passage of RTW laws, as opposed to being
a result of it. If the passing of RTW laws is a consequence rather than the cause, then the reduction
in union organizing should be visible during the
immediate years before the passage of the law.
Ellwood and Fine (1987) investigated this possibility
by analyzing the evolution of new union organizing
for seven states prior to the adoption of the law;
they detected no reduction in union organizing
during that period and concluded that the decline
in union organizing is likely to have been caused
by the passage of RTW laws.
According to the anecdotal evidence in Kendrick
Dinlersoz and Hernández-Murillo
(1996), one possible source of the events that led to
the eventual passage of the law was the “Bunker Hill”
incident. In 1984, employees of the Bunker Hill
mining company in Idaho voted for voluntary pay
cuts and other concessions to keep the company
from going out of business. The union headquarters
in Pittsburgh overruled this vote, resulting in a loss
of 1500 jobs. The Bunker Hill incident might have
initiated a change in attitude toward unions in Idaho.
If this is the case, then a growing anti-unionism in
the state might be the reason for the eventual passage of the law.
The rest of our paper is organized as follows.
We present evidence in the next section on the
evolution of unionization before and after the RTW
law was enacted, followed by evidence on the growth
in manufacturing.
PATTERNS OF UNIONIZATION IN
IDAHO
Unionization Across Industries
We used data from the Census Bureau’s Current
Population Survey (CPS) to estimate unionization
rates. We describe the characteristics of the data
and methodology in the appendix. The employment
and establishments data for our analysis of manufacturing comes from the Census Bureau’s County
Business Patterns data set and is also described in
the appendix.
We start our analysis by examining the evolution
of the unionization rate in Idaho. We compare the
trends in Idaho’s unionization rate with the average
trend in both RTW states and NRTW states that had
an industrial mix similar to that of Idaho in the years
prior to the passage of the law, 1977-86. For this we
construct a measure of dispersion using the employment shares in broadly defined industries.8 We identi8
–
We computed the following measure of distance (∆k ) to Idaho for each
of the 50 states in terms of industrial mix and performed the comparison for the closest “neighbors”:
∆k =
1 T N it it 2
Σ t =1Σ i =1( sk − s ) ,
T
where sitk is the employment share in industry i in state k in year t, N
is the number of industries, T is the total number of years in the sample
period, and sit is the index for Idaho, defined similarly. We used employment data from the following industry classifications: agricultural, mining, construction, manufacturing, transportation, wholesale trade,
retail trade, finance insurance and real estate services, and personal
services. The distribution of this measure had the following characteristics: the maximum value was 0.143, the mean was 0.019, and the
5th, 25th, 50th, 75th, and 90th percentiles were 0.002, 0.005, 0.014,
0.024, and 0.035. We selected states with a distance of less than 0.005.
M AY / J U N E 2 0 0 2
31
REVIEW
Dinlersoz and Hernández-Murillo
Figure 2
Figure 3
Evolution of Unionization in
Manufacturing Industries
Evolution of Unionization in
Manufacturing Industries
Idaho vs. RTW States
Idaho vs. NRTW States
Percent Unionization
Percent Unionization
50
50
45
Idaho
RTW 90% Upper Limit
45
Idaho
NRTW 90% Upper Limit
40
RTW 90% Lower Limit
RTW Average
40
NRTW 90% Lower Limit
NRTW Average
35
35
30
30
25
25
20
20
15
15
10
10
5
5
0
1977
79
81
83
85
87
89
91
93
95
97
fied 11 such states: 5 RTW states (Kansas, Nebraska,
Utah, Virginia, and Iowa) and 6 NRTW states
(California, Colorado, Minnesota, Oklahoma, Oregon,
and Washington).
So how do the patterns in unionization rates
differ across industries? The manufacturing sector,
being traditionally highly unionized, behaved quite
differently compared with the nonmanufacturing
sector. Figures 2 and 3, respectively, compare the
unionization rate in manufacturing to the average
of RTW and NRTW states. What is most interesting
about the trend for Idaho’s unionization is the relatively large decline that occurred between 1981 and
1984, prior to the passage of the law, and the pronounced recovery in the 1984-87 period, during
which much of the debate about the passage of the
law took place. We observe that the manufacturing
unionization rate in Idaho gradually converged to
the average unionization rate in RTW states. The
convergence took place mostly after 1987, and this
rate remained within the confidence bands and
below the average for RTW states that had similar
industrial composition prior to 1987. Figure 3 indicates that the manufacturing unionization rate in
Idaho remained within the confidence bands for
the average for NRTW states for most of the sample
period, but fell below the lower confidence band in
1994 and remained away from the average thereafter.
The patterns observed in Idaho’s manufacturing
unionization rate do not seem to result from business
32
M AY / J U N E 2 0 0 2
0
1977
1999
79
81
83
85
87
89
91
93
95
97
1999
cycles that affected all other states uniformly. However, since Idaho is a small state, its manufacturing
unionization rate may have been subject to fluctuations in the unionization rate of a small number of
industries, particularly in the period prior to the
passage of the RTW law. Examining Idaho’s unionization rates in narrowly defined manufacturing industries, we discovered that fluctuations in the years
prior to 1987 were closely related to fluctuations in
the food manufacturing industry.
Figure 4 shows the evolution of the overall
unionization rate in Idaho versus the average unionization rate in the five states with RTW laws and a
similar industrial mix.9 Idaho’s unionization rate
was around 17 percent in 1977; by 2000 it was down
to about 9 percent, a decline of almost 50 percent.
The average rate for RTW states also declined steadily,
starting in 1981. Throughout the period of analysis,
in 1983, 1984, and then again in 1987, 1989, 1991,
1992, and 1994, Idaho’s unionization rate was significantly different from the average RTW state’s
unionization rate, at the 90 percent confidence level.
In the years 1977-81 we observe that Idaho’s unionization rate was close to the upper confidence band.
In just three years, during the period 1981-84, the
9
Note that there was no change in other states’ RTW law status during
1977-2000. Idaho was the only state that changed status during this
period. Louisiana became a RTW state in 1976 and is included with
the other RTW states throughout the period. Excluding Louisiana did
not change our conclusions.
FEDERAL RESERVE BANK OF ST. LOUIS
Dinlersoz and Hernández-Murillo
Figure 4
Figure 5
Evolution of Unionization in
All Industries
Evolution of Unionization in
All Industries
Idaho vs. RTW States
Idaho vs. NRTW States
Percent Unionization
Percent Unionization
35
35
Idaho
30
RTW 90% Lower Limit
30
RTW Average
25
25
20
20
15
15
10
10
5
5
0
1977
79
81
83
Idaho
NRTW 90% Upper Limit
NRTW 90% Lower Limit
NRTW Average
RTW 90% Upper Limit
85
87
89
91
93
95
97
0
1977
1999
unionization rate fell from about 22 percent to
almost 9 percent, a decline of about 60 percent. The
decline observed for the average rate for RTW states
was not as pronounced.10 The pattern between
1984 and 1987 also exhibits a partial recovery in
the unionization rate. After the law took effect in
1987, however, we observe a persistent decline in
the unionization rate.
In Figure 5, we compare Idaho with the six
closest NRTW states. First, note that on average, a
NRTW state had a unionization rate of about 24
percent in 1977, compared with 17 percent for RTW
states. These figures were about 14 percent and 9
percent, respectively, in 2000. The difference in
unionization rates between the two groups of states
persisted throughout the sample period. In the years
1979-82, Idaho’s unionization rate is not statistically
distinguishable from the average unionization rate
in NRTW states. In the years following the 1981-84
decline, however, we can reject the equality of the
two rates. Idaho’s unionization rate hit the lower
confidence bound for the NRTW states’ average
around 1982 and consistently remained below that
bound for the rest of the analysis period. From the
patterns observed in Figures 4 and 5, Idaho’s unionization rate very early diverged from the NRTW
states’ average unionization rate and approached
the RTW states’ average. As shown in Figures 6 and
7, this behavior was largely due to the behavior
observed in the nonmanufacturing sector. In both
79
81
83
85
87
89
91
93
95
97
1999
figures, the dip during the 1981-84 period is visible
and highly pronounced, and even as early as 1982
the unionization rate in nonmanufacturing industries had converged to the average unionization
rate in RTW states and was statistically below the
NRTW states’ average. It is therefore likely that the
quick convergence in Idaho’s overall unionization
rate was unrelated to the passage of the RTW.11
Idaho’s Neighbors
To investigate the trends in the unionization
rate further, we concentrate on Idaho’s geographic
neighbors and run a simple state-by-state regression
of the form
10
As explained in the appendix, prior to 1983, unionization rates were
calculated based on samples that are roughly one-third of the samples
that are used after 1983. The estimated unionization rates are less
precise for the period before 1983 due to sample variability, especially
for smaller states, and in particular for 1981, when the sample sizes
were roughly one-third of the samples in 1977-80. Estimates of overall
and nonmanufacturing unionization rates were less sensitive to sampling problems than those for the manufacturing sector. Still, when we
discount 1981 and 1982, the decline observed in the manufacturing
unionization rate from 1980 to 1983 is reliably estimated.
11
As previously footnoted, the estimates of overall and nonmanufacturing unionization rates during the period 1977-86 were not likely to
be seriously affected by the small sample sizes used by the CPS before
1983, even accounting for 1981, as the sample sizes used in the estimation exceeded the thresholds described in the appendix for reliability
of the estimates. We are, however, silent on the driving factors of unionization in Idaho’s nonmanufacturing industries, as the focus of our
analysis is the manufacturing sector. We did verify, however, that the
1981-84 decline was not due to closures of large unionized firms.
M AY / J U N E 2 0 0 2
33
REVIEW
Dinlersoz and Hernández-Murillo
Figure 6
Figure 7
Evolution of Unionization in
Nonmanufacturing Industries
Evolution of Unionization in
Nonmanufacturing Industries
Idaho vs. RTW States
Idaho vs. NRTW States
Percent Unionization
Percent Unionization
30
30
25
Idaho
RTW 90% Upper Limit
Idaho
NRTW 90% Upper Limit
RTW 90% Lower Limit
RTW Average
NRTW 90% Lower Limit
NRTW Average
25
20
20
15
15
10
10
5
5
0
1977
79
81
83
85
87
89
91
93
95
97
0
1977
1999
79
81
83
85
87
89
91
93
95
97
1999
Table 1
Change in Unionization Rate by State and Industry
Overall
1977-86
1987-2000
–3.7
[–0.3]
–1.8
[0.06]
Idaho
–6.4
[1.8]
Washington
Manufacturing
1977-86
1987-2000
1977-85
1987-2000
27.37***
(0.00)
–4.7
[0.3]
–3.3
[0.06]
15.54***
(0.00)
–2.8
[0.3]
–1.2
[0.08]
17.97***
(0.00)
–2.8
[0.7]
3.2*
(0.08)
–5.8
[2.0]
–8.0
[0.8]
1.06
(0.31)
–6.3
[2.5]
–0.5
[0.7]
4.75**
(0.04)
–3.0
[0.7]
–1.7
[0.3]
2.98*
(0.09)
–4.1
[0.9]
–2.5
[0.3]
2.73
(0.11)
–2.4
[0.7]
–1.2
[0.3]
2.25
(0.14)
Oregon
–3.6
[1.0]
–2.1
[0.4]
1.71
(0.20)
–7.6
[0.7]
–5.7
[0.7]
3.14*
(0.09)
–1.6
[1.1]
–1.3
[0.4]
0.06
(0.81)
Montana
–4.2
[0.7]
–2.1
[0.4]
6.48***
(0.01)
–3.0
[2.7]
–5.3
[0.9]
0.65
(0.43)
–4.4
[0.6]
–1.8
[0.3]
11.36***
(0.00)
Nevada (RTW) –3.1
[0.9]
0.2
[0.4]
10.34***
(0.00)
–9.2
[3.9]
–0.5
[2.5]
3.46*
(0.07)
–2.9
[0.8]
0.2
[0.4]
10.22***
(0.00)
Utah (RTW)
–5.5
[1.4]
–3.7
[0.6]
1.26
(0.27)
–9.6
[1.7]
–2.6
[1.2]
11.49***
(0.00)
–4.7
[1.4]
–4.0
[0.7]
0.17
(0.68)
Wyoming
(RTW)
–3.5
[1.5]
–3.8
[0.2]
0.02
(0.88)
0.2
[2.1]
2.8
[3.3]
0.45
(0.51)
–3.7
[1.6]
–4.0
[0.2]
0.04
(0.83)
U.S.
F (Prob)
Nonmanufacturing
F (Prob)
F (Prob)
NOTE: Heteroskedasticity-autocorrelation consistent standard errors are in brackets. Figures in bold indicate significance at 1 percent.
``F” gives the F statistic for the test of equality of coefficients across two time periods. Probability values for the F statistic are in parentheses. *, **, and *** indicate significance of the F statistic at the 10, 5, and 1 percent levels, respectively.
34
M AY / J U N E 2 0 0 2
FEDERAL RESERVE BANK OF ST. LOUIS
(1) log ut=α PRE+β PRE D(t – t0 )+∆α POST (1– D)
+β POST (1– D)(t – t0 )+ε t ,
where t0=1977, t=1977,...,2000, and D is a dummy
variable that takes on a value of 1 if t<1987 and 0
otherwise. In this projection, α PRE is the intercept
term for the pre-law period, β PRE is the pre-law slope
coefficient, ∆α POST is the post-law increment in the
intercept, and β POST is the post-law slope coefficient.
The estimated values of β PRE and β POST are multiplied
by 100 and are presented in Table 1. With the log
specification, the figures in the table can be interpreted as the annual percent rate of change in
unionization. We also present the test results for the
equality of the growth rates across the two periods
β PRE=β POST.
We observe a persistent decline in unionization
rates. When all industries are considered, columns
1 and 2 reveal that, in general, the magnitude of
the decline was higher in the 1977-86 period in all
states in the region and in the United States, except
for Wyoming. In Idaho, the rate of decline in overall
unionization slowed down from 6.4 percent in the
pre-law period to 2.8 percent in the post-law period.
The difference between these two rates, however, is
statistically significant only at the 10 percent level.
Note also that, in both periods, Idaho’s rates of
decline were higher than the U.S. rates and most of
those for its neighboring states.
When manufacturing is considered separately,
columns 4 and 5 provide a different view. In fact,
the decline in Idaho’s manufacturing unionization
rate accelerated somewhat in the post-law period,
surpassing both the U.S. and its neighboring states,
which, for the most part, exhibited a slowdown in
the rate of decline. The difference between Idaho’s
unionization rates in manufacturing pre-law and
post-law is not statistically significant because of
the relatively high standard deviation for the pre-law
period. Overall, the slowdown in the rate of decline
of unionization did not apply to Idaho’s manufacturing and was primarily driven by nonmanufacturing
industries, as can be seen in the last two columns.
The findings in this section suggest that Idaho’s
unionization rate declined substantially over the
sample period, approaching the average unionization rate in RTW states. While the decline in the
unionization rate, especially in manufacturing, is
persistent after 1987, a substantial part of the decline
appears to have happened before 1987. The pattern
between 1984 and 1987, during which much of
the debate about the passage of the law took place,
exhibits a partial recovery in the unionization rate.
After the law took effect in 1987, we observe a con-
Dinlersoz and Hernández-Murillo
Figure 8
Evolution of Key Indicators in Idaho’s
Manufacturing Industry
Relative Magnitude Compared with 1987
1.6
1.4
1.2
1.0
0.8
0.6
0.4
Employment
Number of Establishments
0.2
0.0
1975 77
Unionization
79
81
83
85
87
89
91
93
95
97 1999
tinuing decline in the unionization rate, especially
in manufacturing. Particularly during the period
prior to 1987, large fluctuations in Idaho’s unionization rate in manufacturing seem to be related to
the behavior of individual industries.
MANUFACTURING
We now turn to the industrial organization consequences of declining unionization in Idaho. We
focus on two main indicators. First, we look at the
growth in employment and the number of establishments in manufacturing industries and compare
Idaho with its neighbors, in both the pre- and postlaw periods. If the passage of the law has had an
important positive effect on manufacturing growth,
then we expect to observe an acceleration in the
growth rate of employment and the number of
establishments in Idaho. Second, we look at the
changes in the importance of large establishments
in manufacturing in Idaho, again, for both periods.
As Holmes (1998) argues, large manufacturing
establishments are more likely to be attracted to
RTW states because larger plants are more likely to
be unionized. This argument suggests that we might
expect an influx of new large establishments into
Idaho or an expansion of existing establishments.
Employment Growth
Figure 8 is a preliminary look at the evolution
of the three key variables in Idaho’s manufacturing,
M AY / J U N E 2 0 0 2
35
REVIEW
Dinlersoz and Hernández-Murillo
Table 2
Manufacturing Growth Rates in Idaho and Its Neighbors (Simple Time Averages, Percent
Annual Growth)
Employment
Idaho
No. of establishments
1975-86
1987-96
1975-86
1987-96
1975-86
1987-96
0.76
[6.38]
3.71
[2.56]
1.27
[4.20]
3.99
[3.16]
–0.39
[6.73]
–0.21
[3.17]
(1.36*)
Washington
1.57
[5.34]
(1.98**)
2.18
[6.23]
2.86
[4.16]
(0.25)
Oregon
1.18
[6.86]
1.67
[2.64]
–0.33
[6.80]
3.51
[4.52]
–1.42
[9.66]
–2.10
[6.79]
6.01
[3.43]
3.01
[2.93]
0.96
[10.42]
3.57
[3.55]
2.19
[6.50]
(0.59)
0.58
[5.69]
–0.17
[4.96]
(–0.32)
2.69
[4.16]
(0.21)
–1.00
[5.64]
(–0.53)
(0.41)
3.32
[3.39]
–0.51
[6.33]
(0.56)
(0.24)
3.26
[2.64]
0.55
[4.30]
(0.57)
2.09
[4.34]
5.46
[6.52]
(–0.15)
Wyoming (RTW)
–0.98
[7.54]
(0.07)
4.84
[5.02]
0.29
[6.94]
(0.43)
1.19
[2.52]
1.94
[5.43]
(–0.40)
Utah (RTW)
–1.04
[7.48]
(–0.74)
1.35
[3.47]
6.15
[9.22]
1.96
[2.44]
2.32
[4.23]
(0.71)
Nevada (RTW)
(0.08)
(–0.59)
(0.21)
Montana
Average establishment size
–0.53
[9.63]
0.87
[7.26]
(0.38)
NOTE: Standard deviations in brackets. Figures in parentheses are the t statistics associated with the difference of the variable’s average
across two periods of analysis. * and ** indicate significance at the 10 and 5 percent levels, respectively, for a one-sided test; t statistics
are based on unpaired comparisons with unequal variances.
where we have normalized each variable by its 1987
value. Before 1987, there is considerable fluctuation
in both employment and the number of establishments, with no visible growth trend. Unionization
exhibits a decline, but is also subject to wide fluctuations, as discussed before. The pattern after 1987
is remarkably stable for all three series. Employment
and the number of establishments grew steadily in
that period by about 40 percent compared with their
1987 level, and unionization declined by more than
60 percent.
Table 2 shows the simple average annual growth
rates in employment, number of establishments,
and average establishment size in manufacturing
for Idaho and its neighbors. Consider employment
and the number of establishments first. From 1975
to 1986, Idaho’s manufacturing employment grew
at a rate of 0.76 percent annually on average. The
36
M AY / J U N E 2 0 0 2
average growth rate in the number of establishments
was around 1.27 percent per year. However, there
is a large standard deviation associated with both
of these figures, a reflection of the fluctuating
manufacturing growth in the state in that period,
as depicted in Figure 8. Idaho’s NRTW neighbors
did not fare much better. Washington and Oregon
appear to have experienced higher growth rates, but
the standard deviations are so high that the differences with respect to Idaho are not statistically significant. Idaho’s RTW neighbors appear to have
fared much better in this period, except for Wyoming.
Overall, it seems that the period before the law was
a period of weak growth, especially for NRTW states.
This pattern changes dramatically in the postlaw period. Idaho’s growth rates were much higher
compared with those in the pre-law period. Furthermore, the difference between the two periods’ growth
FEDERAL RESERVE BANK OF ST. LOUIS
Dinlersoz and Hernández-Murillo
Table 3
Manufacturing Growth Rates: Results from State-by-State Regressions
Employment
No. of establishments
1977-86
1987-2000
F (Prob)
1977-85
1987-2000
F (Prob)
Idaho
–0.03
[0.4]
3.7
[0.2]
58.97
(0.00)
0.6
[0.3]
4.1
[0.2]
82.71
(0.00)
Washington
1.1
[0.5]
0.4
[0.8]
0.48
(0.49)
2.4
[0.2]
2.1
[0.2]
0.93
(0.34)
Oregon
0.02
[0.6]
1.2
[0.2]
3.17
(0.09)
2.0
[0.2]
1.6
[0.2]
2.04
(0.17)
Montana
–1.3
[0.6]
1.2
[0.1]
15.74
(0.00)
1.8
[0.6]
2.6
[0.2]
2.58
(0.12)
Nevada (RTW)
5.0
[0.7]
4.3
[0.6]
0.49
(0.49)
4.9
[0.5]
6.1
[0.2]
4.39
(0.05)
Utah (RTW)
3.3
[0.3]
3.1
[0.1]
0.28
(0.60)
2.9
[0.2]
4.0
[0.3]
8.85
(0.00)
Wyoming (RTW)
–0.03
[1.3]
2.7
[0.2]
3.99
(0.06)
2.1
[0.4]
3.0
[0.3]
2.07
(0.16)
NOTE: Heteroskedasticity-autocorrelation consistent standard errors are in brackets. Figures in bold indicate significance at 1 percent.
``F” gives the F statistic for the test of equality of coefficients across two time periods. Probability values for the F statistic are in
parentheses.
rates turns out to be statistically significant, unlike
the case with neighboring states. Idaho’s post-law
growth rates also exceeded those of its NRTW neighbors and were similar to those of its RTW neighbors
(although the pairwise comparisons are not always
statistically significant due to large standard errors).
Overall, the patterns of change in the growth of
employment and the number of establishments
point to a post-law acceleration of growth in Idaho,
but not in any of the neighboring states.
Table 3 shows the results of a regression analogous to equation (1). The dependent variable is
either the logarithm of employment or the number
of establishments in manufacturing. The most notable result from this table is the exceptionally large
growth rate of Idaho in the post-law period for both
variables. The annual employment growth rate was
about 3.7 percent post-law, compared with an almost
zero annual average growth pre-law. The growth rate
in the number of establishments was about seven
times larger compared with that in the pre-law
period. Idaho did much better after the RTW law was
passed, compared with most other states in the
region, both in employment and the number of establishments. The differences in these growth rates
across the two periods have high statistical signifi-
cance for Idaho, but not for most of the other
states.
Manufacturing Employment Share
Before turning to the analysis of establishment
size, we report how the share of manufacturing as
a fraction of total private employment evolved in
Idaho. Again we compare Idaho against other states
that had a similar industrial mix in the period prior
to 1987. This analysis indicates that Idaho experienced a substantial change in industrial mix, especially after the passage of the RTW law.12 Figure 9
compares Idaho’s manufacturing share with the
average manufacturing share in the six NRTW states
we identified earlier. First, note that manufacturing’s
average employment share in NRTW states declined
throughout the sample period, which is an indication of the steady decline in the manufacturing
sector in the United States, especially during the
last quarter of the twentieth century. Idaho’s manufacturing share was far below the NRTW average
12
Constructing a distance measure analogous to that of footnote 9, we
observed that Idaho also experienced a substantial change during
our sample period in the composition of its manufacturing industry.
For brevity, we omit this analysis.
M AY / J U N E 2 0 0 2
37
REVIEW
Dinlersoz and Hernández-Murillo
Figure 9
Figure 10
Evolution of Manufacturing Share
Evolution of Manufacturing Share
Idaho vs. NRTW States
Idaho vs. RTW States
Share
Share
0.29
Idaho
NRTW 90% Upper Limit
NRTW 90% Lower Limit
NRTW Average
0.29
0.27
0.27
0.25
0.25
0.23
0.23
0.21
0.21
0.19
0.19
0.17
0.17
0.15
1975
77
79
81
83
85
87
89
91
93
during the 1975-82 period and declined at a much
faster rate than the average share in NRTW states.
This trend slowly started to change around 1982;
from 1984 onward, the manufacturing share in
Idaho was above the NRTW average and declined
much more slowly, which is consistent with accelerated growth in Idaho’s manufacturing employment
in this period. By 1987, Idaho’s share exceeded the
NRTW average, and the difference gradually became
statistically significant. By the end of the analysis
period, we can reject the hypothesis that Idaho had
a manufacturing share similar to an “average” NRTW
state with, initially, a similar industrial composition.
The comparison with the RTW states’ average share
in Figure 10 is consistent with this finding. While
Idaho’s share was much lower than the average RTW
states’ share before 1982, it gradually became closer
to the average afterward.13
Average Establishment Size
Considering the results in Table 2 regarding the
change in average establishment size, defined by
the number of employees per establishment, we
do not observe any definitive pattern. In all states,
the difference in the average growth rates in this
variable across the two periods was insignificant.
This, however, does not necessarily mean that Idaho
did not become an attractive location for larger
plants or that existing plants had less incentive to
38
M AY / J U N E 2 0 0 2
0.15
1975
1995
77
Idaho
RTW 90% Upper Limit
RTW 90% Lower Limit
RTW Average
79
81
83
85
87
89
91
93
1995
expand. It is well-known that there has been an
ongoing nationwide trend toward smaller establishments.14 It is possible that the increasing fraction
of small plants in Idaho masked the increasing
importance of larger establishments. To investigate
this possibility, we look at the evolution of two
measures: (i) the fraction of manufacturing employment in large establishments and (ii) the average
size of large establishments. Following Holmes
(1998), we define an establishment as “large” if it
has at least 100 employees.15 If large establishments
became more important in Idaho’s manufacturing
sector after the law, then the first measure is expected
to be higher in the post-law period. Similarly, if
existing large establishments expanded, or if new
large establishments that chose Idaho as a location
after the law were larger than their pre-law counterparts on average, then we should see an increase in
the second measure, too.
As Table 4 clearly indicates, the two variables
13
The observations in this section also apply if we consider all RTW
and NRTW states, not just those with an industrial mix similar to that
of Idaho.
14
See, for example, Davis (1990) and Davis and Haltiwanger (1990). The
trend toward smaller establishment sizes might also be responsible
for declining unionization, as explored by Even and Macpherson
(1990).
15
This choice is somewhat ad hoc, but as reported by Holmes (1998),
70 percent of all manufacturing establishments in 1992 were classified
in this category. Outside manufacturing, the figure was 38 percent.
FEDERAL RESERVE BANK OF ST. LOUIS
Dinlersoz and Hernández-Murillo
Table 4
Large Establishments in Manufacturing: Idaho and Its Neighbors
Average fraction of employment
in large establishments
Average establishment size
in large establishments
1975-86
1987-96
1975-86
1987-96
Idaho
0.66
[0.015]
0.68
[0.007]
324.6
[16.3]
348.2
[10.8]
Washington
0.70
[0.013]
0.69
[0.018]
444.1
[25.3]
Oregon
0.63
[0.016]
0.61
[0.007]
157.9
[7.5]
Montana
0.51
[0.035]
0.43
[0.028]
261.6
[22.6]
Nevada (RTW)
0.51
[0.020]
0.47
[0.035]
236.5
[21.4]
Utah (RTW)
0.68
[0.015]
0.68
[0.011]
355.3
[30.0]
Wyoming (RTW)
0.43
[0.035]
0.42
[0.020]
198.9
[16.6]
(2.97)
(4.03)
(–1.06)
(0.49)
(–3.69)
148.4
[3.7]
(–3.84)
(–5.84)
224.8
[11.1]
(–4.96)
(–2.84)
236.6
[16.3]
(0.01)
(0.01)
(–0.31)
451.6
[42.1]
358.3
[11.0]
(0.32)
184.3
[8.9]
(–2.48)
NOTE: Standard deviations in brackets. Figures in parentheses are the t statistics associated with the test for the equality of the variable’s
average across two periods of analysis. Figures in bold indicate significance at 1 percent; t tests are based on unpaired comparisons
wth unequal variances.
measuring the importance of large establishments
in the manufacturing sector experienced a significant increase in Idaho after the law was passed,
but this did not occur in any of the neighboring
states. There was about a 3 percent increase in the
average fraction of employment in large establishments after the law, and the average establishment
size for large establishments grew by about 24
employees, or by 7 percent. These results are consistent with the view (i) that Idaho became an attractive location for large establishments after the RTW
law was passed and (ii) that the importance of large
establishments in the manufacturing sector increased.
CONCLUSION
We have examined the impact of RTW laws on
a state’s industrial performance using Idaho’s recent
experience. We have presented evidence that, even
as a late adopter of the law, Idaho experienced a
strong decline in unionization and an acceleration
in manufacturing growth. Evidence from Idaho’s
neighbors suggests that a similar pattern was not
experienced by other states in the region, which
indicates that a regional boom is not a likely explanation. We are cautious, however, in associating the
increase in manufacturing growth with the passage
of the law. The exact starting time of the decline in
unionization and the narrow time frame of fluctuations in the unionization rate before the passage of
the law suggest that the relation is not clear cut. The
initial decline in unionization and its subsequent
rebounding between 1984 and 1987 can potentially
be related also to evolving expectations about the
eventual ruling on the RTW law—because the
bureaucratic process and political battles for the
passing of the RTW law took almost two years, with
several developments in favor of and against unionM AY / J U N E 2 0 0 2
39
Dinlersoz and Hernández-Murillo
ism. Adding to our skepticism is the Bunker Hill
incidence mentioned earlier, which, by itself, may
have been a turning point for the attitudes toward
unions in Idaho. In summary, while we are tempted
to associate the growth patterns and the decline in
unionization with the passage of the law, we cannot
rule out the possibility that the RTW law was a result
of growing anti-unionism in Idaho and may not have
been the cause of growth, per se.
In terms of policy implications, one has to be
cautioned before prematurely claiming that Idaho’s
exceptional growth pattern would apply to every
state considering the adoption of the law. Idaho’s
experience would definitely be more valuable than
the evidence from other RTW legislation in the past
because it took place in an environment where
unionization had already lost considerable ground.
As the analysis presented here suggests, even the
process leading to the passage of the law may be
quite important for the timing of events and the
patterns of growth in key variables. Examining the
behavior of union organizing activity through certification elections, as well as analyzing the effects
on wages, can provide a more detailed analysis of
the impact of the RTW law on unionization. The
recent experience of Oklahoma, together with
Idaho’s, can be used for this purpose. The ongoing
work by Dinlersoz and Hernández-Murillo (2001)
aims to provide more evidence in this direction.
REFERENCES
Abraham, Steven E. and Voos, Paula B. “Right-to-Work
Laws: New Evidence from the Stock Market.” Southern
Economic Journal, October 2000, 67(2), pp. 345-62.
Davis, Steven J. “Size Distribution Statistics from County
Business Patterns Data.” Working paper, University of
Chicago, 1990.
___________ and Haltiwanger, John. “The Distribution of
Employees by Establishment Size: Patterns of Change in
the United States: 1962 to 1985.” Unpublished manuscript,
University of Chicago, 1990.
Dinlersoz, Emin M. and Hernández-Murillo, Rubén. “A
Recent Assessment of the RTW Laws’ Effect in the Wake
of Idaho’s Experience.” Working paper, 2001.
Ellwood, David T. and Fine, Glenn. “The Impact of Right-
40
M AY / J U N E 2 0 0 2
REVIEW
to-Work Laws on Union Organizing.” Journal of Political
Economy, April 1987, 95(2), pp. 250-73.
Even, William E. and Macpherson, David A. “Plant Size and
the Decline of Unionism.” Economics Letters, April 1990,
32(4), pp. 393-98.
Galarneau, Diane. “Unionized Workers.” Perspectives on
Labour and Income, Spring 1996, 8(1), pp. 43-52.
Goldfield, Michael. The Decline of Organized Labor in the
United States. Chicago: The University of Chicago Press,
1987.
Hirsch, Barry T.; Macpherson, David and Vroman, Wayne G.
“Estimates of Union Density by State.” Monthly Labor
Review, July 2001, 124(7), pp. 51-55.
Holmes, Thomas J. “The Effect of State Policies on the
Location of Manufacturing: Evidence from State Borders.”
Journal of Political Economy, August 1998, 106(4), pp.
667-705.
Kendrick, David. “Right-to-Work—The Idaho Experience.”
National Institute for Labor Relations Research, delivered
at the Fraser Institute Conference on Right to Work,
Toronto, Canada, 21 June 1996.
Long, Richard J. “The Effect of Unionization on Employment
Growth of Canadian Companies,” Industrial and Labor
Relations Review, July 1993, 46(4), pp. 691-703.
Lowe, Graham S. “The Future of Work: Implications for
Unions.” Relations Industrielles/Industrial Relations, 1998,
53(2), pp. 1-25.
Moore, William J. and Newman, Robert J. “The Effects of
Right-to-Work Laws: A Review of the Literature.” Industrial
and Labor Relations Review, July 1985, 38(4), pp. 571-85.
___________. “The Determinants and Effects of Right-toWork Laws: A Review of the Recent Literature.” Journal
of Labor Research, Summer 1998, 19(3), pp. 445-69.
National Association of State Development Agencies.
Directory of Incentives for Business Investment and
Development in the United States: A State-by-State Guide.
Washington, DC: The Urban Institute Press, various
years.
FEDERAL RESERVE BANK OF ST. LOUIS
Dinlersoz and Hernández-Murillo
Appendix
DATA DESCRIPTION
Unionization Rates
Estimates of union membership rates by state
and by state industry were obtained using the May
files of the Census Bureau’s Current Population
Survey (CPS) for the period 1977-81, and from the
Merged Outgoing Rotation Groups CPS files for the
period 1983-2000, following the methodology of
Hirsch, Macpherson, and Vroman (2001). The 1982
CPS did not include any questions pertaining to
unions, and we set our estimate for 1982 to the
average of the estimates in 1981 and 1983. For
1983 and onward, each year included all 12 months
of the CPS, with each month including the outgoing
rotation groups that were asked the union questions. Prior to 1981, the May surveys administered
the union questions to all rotation groups; therefore the estimates before 1981 are based on samples that are one third of the size of the samples
used after 1983. The May 1981 CPS administered
the union questions only to the outgoing rotation
groups, making sample sizes roughly one-third of
the samples used in 1977-80. Union estimates for
1981 are, therefore, the least reliable.16
Due to the varying sample sizes, much of the
year-to-year variation in the estimated unionization
rates before 1983 can be attributed to sampling
error. This would be a more serious problem if
one wished to reliably estimate union earnings,
for example, as opposed to simply estimate union
membership rates as we did. The sample sizes of
major industry groups in Idaho (overall, manufacturing, and nonmanufacturing) were within the
standard measures used in the literature and the
Census Bureau’s guidelines (larger than 100 employ-
ees), except for manufacturing and particularly in
1981.
We were able to verify that our estimates of the
proportion of union members from the employed
population closely matched those of Hirsch,
Macpherson, and Vroman (2001) at the national
and state levels. Our estimates of state-industry
rates use the same methodology, but there were
no available series to verify accuracy.
Data on Industries
The data on industries come from the Census
Bureau’s County Business Patterns data series for
the years 1975-96. The data covers all taxpaying
establishments with one or more paid employees.
The employment figures are taken from the midMarch period of every year. An establishment is
defined as a single location where business is conducted or where services or industrial operations
are performed. Establishment size designations are
measured by paid employment in the mid-March
pay period. Establishment counts for 1983 and
onward are based on a determination of active
status as of anytime during the year. For the years
prior to 1983, establishment counts are based on
whether the establishment was active in the fourth
quarter. The data is available at the national, state,
and county levels. Further details on this data set
can be obtained from the Census Bureau’s Web
site, <www.census.gov>.
16
Every household that enters the CPS is interviewed each month for
4 months, then ignored for 8 months, then interviewed again for 4
more months. The union questions are asked only to households
in their fourth and eighth interview. These are the outgoing rotation
groups.
M AY / J U N E 2 0 0 2
41
Dinlersoz and Hernández-Murillo
42
M AY / J U N E 2 0 0 2
REVIEW
Predicting Exchange
Rate Volatility:
Genetic Programming
Versus GARCH and
RiskMetrics
Christopher J. Neely and Paul A. Weller
t is well established that the volatility of asset
prices displays considerable persistence. That is,
large movements in prices tend to be followed
by more large moves, producing positive serial
correlation in squared returns. Thus, current and
past volatility can be used to predict future volatility.
This fact is important to both financial market
practitioners and regulators.
Professional traders in equity and foreign
exchange markets must pay attention not only to
the expected return from their trading activity but
also to the risk that they incur. Risk-averse investors
will wish to reduce their exposure during periods of
high volatility, and improvements in risk-adjusted
performance depend upon the accuracy of volatility
predictions. Many current models of risk management, such as Value-at-Risk (VaR), use volatility
predictions as inputs.
The bank capital adequacy standards recently
proposed by the Basel Committee on Banking
Supervision illustrate the importance of sophisticated risk management techniques for regulators.
These norms are aimed at providing international
banks with greater incentives to manage financial
risk in a sophisticated fashion, so that they might
economize on capital. One such system that is widely
used is RiskMetrics, developed by J.P. Morgan.
A core component of the RiskMetrics system is
a statistical model—a member of the large ARCH/
GARCH family—that forecasts volatility. Such ARCH/
I
Christopher J. Neely is a research officer at the Federal Reserve Bank
of St. Louis. Paul A. Weller is a professor of finance at the Henry B.
Tippie College of Business Administration of the University of Iowa.
This paper is a revised and expanded version of a chapter entitled
“Using a Genetic Program to Predict Exchange Rate Volatility,” in
Genetic Algorithms and Genetic Programming in Computational Finance,
edited by Shu-Heng Chen, published by Kluwer Academic Publishers.
We would like to thank Janis Zvingelis for excellent programming
assistance. Charles Hokayem provided research assistance.
© 2002, The Federal Reserve Bank of St. Louis.
GARCH models are parametric. That is, they make
specific assumptions about the functional form of
the data generation process and the distribution of
error terms. Parametric models such as GARCH are
easy to estimate and readily interpretable, but these
advantages may come at a cost. Other, perhaps
much more complex models may be better representations of the underlying data generation process.
If so, then procedures designed to identify these
alternative models have an obvious payoff. Such
procedures are described as nonparametric. Instead
of specifying a particular functional form for the
data generation process and making distributional
assumptions about the error terms, a nonparametric
procedure will search for the best fit over a large
set of alternative functional forms.
This article investigates the performance of a
genetic program applied to the problem of forecasting volatility in the foreign exchange market. Genetic
programming is a computer search and problemsolving methodology that can be adapted for use in
nonparametric estimation. It has been shown to
detect patterns in the conditional mean of foreign
exchange and equity returns that are not accounted
for by standard statistical models (Neely, Weller, and
Dittmar, 1997; Neely and Weller, 1999, 2001; Neely,
2001). These achievements suggest that a genetic
program may also be a powerful tool for generating
predictions of asset price volatility.
We compare the performance of a genetic
program in forecasting daily exchange rate volatility
for the dollar–Deutsche mark and dollar-yen
exchange rates with that of a GARCH(1,1) model and
a related RiskMetrics volatility forecast (described
in the following section). These models are widely
used by both academics and practitioners and thus
are good benchmarks with which to compare the
genetic program forecasts. While the overall forecast performance of the two methods is broadly
similar, on some dimensions the genetic program
produces significantly superior results. This encouraging finding suggests that more detailed investigation of this methodology applied to volatility
forecasting would be warranted.
THE BENCHMARK MODEL
Before discussing the genetic programming
procedure, we will review the benchmark GARCH
and RiskMetrics volatility models. Engle (1982)
developed the autoregressive conditionally heteroskedastic (ARCH) model to characterize the observed
serial correlation in asset price volatility. Suppose
M AY / J U N E 2 0 0 2
43
REVIEW
Neely and Weller
we assume that a price Pt follows a random walk,
Pt +1 = Pt + ε t +1 ,
(1)
where εt+1∼N(0,σ t2 ). The variance of the error term
depends upon t, and the objective of the model is
to characterize the way in which this variance
changes over time. The ARCH model assumes that
this dependence can be captured by an autoregressive process of the form
(2)
σ t2 = ω + α 0ε t2 + α1ε t2−1 + L + α m ε t2− m ,
where the restrictions ω ≥ 0 and α i ≥ 0 for i=0,1,…,m
ensure that the predicted variance is always nonnegative. This specification illustrates clearly how
current levels of volatility will be influenced by the
past and how periods of high or low price fluctuation will tend to persist.
Bollerslev (1986) extended the ARCH class to
produce the generalized autoregressive conditionally heteroskedastic (GARCH) model, in which the
variance is given by
(3)
σ t2 = ω + β1σ t2−1 + β2σ t2−2 + L + β kσ t2− k
+α 0ε t2 + α1ε t2−1 + L + α m ε t2− m .
The simplest specification in this class, and the one
most widely used, is referred to as GARCH(1,1) and
is given by
(4)
σ t2
=ω+
βσ t2−1 + α ε t2 .
When α+β<1, the variance process displays mean
reversion to the unconditional expectation of σ t2 ,
ω /(1– α – β ). That is, forecasts of volatility in the
distant future will be equal to the unconditional
expectation of σ t2 , ω /(1– α – β ).
The RiskMetrics model for volatility forecasting
imposes the restrictions that α+β=1 and that
ω=0.1 In addition, the parameter β is not estimated,
but imposed to be equal to 0.94 ( J.P. Morgan/Reuters,
1996). This value was found to minimize the meansquared error (MSE) of volatility forecasts for asset
prices. The RiskMetrics one-day-ahead volatility
forecast is
(5)
σ t2 = βσ t2−1 + (1 − β )ε t2 .
The GARCH model has been used to characterize
patterns of volatility in U.S. dollar foreign exchange
markets (Baillie and Bollerslev, 1989, 1991) and in
the European Monetary System (Neely, 1999). However, initial investigations into the explanatory power
of out-of-sample forecasts produced disappointing
44
M AY / J U N E 2 0 0 2
results (West and Cho, 1995). Jorion (1995) found
that volatility forecasts for several major currencies
from the GARCH model were outperformed by
implied volatilities generated from the Black-Scholes
option-pricing model. These studies typically used
the squared daily return as the variable to be forecast. However, the squared return is a very imprecise
measure of true, unobserved volatility. For example,
the exchange rate may move around a lot during
the day, and yet end up close to its value the same
time the previous day. In this case, the squared daily
return would be small, even though volatility was
high. More recently, it has been demonstrated that
one can significantly improve the forecasting power
of the GARCH model by measuring volatility as the
sum of intraday squared returns (Andersen and
Bollerslev, 1998). This measure is referred to as
integrated, or realized, volatility. In theory, if the
true underlying price path is a diffusion process, it
is possible to obtain progressively more accurate
estimates of the true volatility by increasing the
frequency of intraday observation. Of course, there
are practical limits to this; microstructural effects
begin to degrade accuracy beyond a certain point.
GENETIC ALGORITHMS AND GENETIC
PROGRAMMING
Genetic algorithms are computer search procedures used to solve appropriately defined problems.
The structure of the search procedure is based on
the principles of natural selection. These procedures
were developed for genetic algorithms by Holland
(1975) and extended to genetic programming by Koza
(1992). The essential features of both algorithms
include (i) a means of representing potential solutions to a problem as character strings that can be
split up and recombined to form new potential solutions and (ii) a fitness criterion that measures the
“quality” of a candidate solution. Both types of
algorithms produce successive “generations” of
candidate solutions using procedures that mimic
genetic reproduction and recombination. Each new
generation is subjected to the pressures of “natural
selection” by increasing the probability that candidate solutions scoring highly on the fitness criterion
get to reproduce.
To understand the principles involved in genetic
1
The restriction, α+β=1, implies that shocks to the volatility process
persist forever; higher volatility today will lead one to forecast higher
volatility indefinitely. It therefore falls into the class of integrated
GARCH or IGARCH models.
FEDERAL RESERVE BANK OF ST. LOUIS
programming, it is useful to understand the operation of the simpler genetic algorithm. Genetic algorithms require that potential solutions be expressed
as fixed-length character strings. Consider a problem
in which candidate solutions are mapped into binary
strings s, with a length of five digits. One possible
solution would be represented as (01010). Associated
with this binary string would be a measure of fitness
that quantifies how well it solves the problem. In
other words, we need a fitness function m(s) that
maps the strings into the real line and thus ranks
the quality of the solutions. Next we introduce the
crossover operator. Given two strings, a crossover
point is randomly selected and the first part of one
string is combined with the second part of the other.
For example, given the two strings (00101) and
(11010) and a crossover point between elements two
and three, the new string (00010) is generated. The
remaining parts of the original strings are discarded.
The algorithm begins by randomly generating
an initial population of binary strings and then
evaluating the fitness of each string by applying the
fitness function m(s). Next, the program produces a
new (second) generation of candidate solutions by
selecting pairs of strings at random from this initial
population and applying the crossover operator to
create new strings. The probability of selecting a given
string is set to be proportional to its fitness. Thus a
“selection pressure” in favor of progressively superior
solutions is introduced. This process is repeated to
produce successive generations of strings, keeping
the size of each generation the same. The procedure
“evolves” new generations of improved potential
solutions.
Recall that genetic algorithms require that
potential solutions be encoded as fixed length
character strings. Koza’s (1992) extension, genetic
programming, instead employs variable-length,
hierarchical strings that can be thought of as decision trees or computer programs. However, the basic
structure of a genetic program is exactly the same
as described above. In particular, the crossover operator is applied to pairs of decision trees to generate
new “offspring” trees.
The application in this paper represents forecasting functions as trees and makes use of the
following function set in constructing them: plus,
minus, times, divide, norm, log, exponential, square
root, and cumulative standard normal distribution
function. In addition, we supply the following set
of data functions: data, average, max, min, and lag.
The data functions can operate on any of the four
Neely and Weller
data series that we provide as inputs to the genetic
program: (i) daily foreign exchange returns, (ii) integrated volatility (i.e., the sum of squared intraday
returns), (iii) the sum of the absolute value of intraday returns, and (iv) the number of days until the
next business day. For example, data (returns (t )) is
simply the identity function that computes the daily
return at t. The other data functions operate in a
similar fashion, but also take numerical arguments
to specify the length of the window—the number
of observations—over which the functions operate.
The numerical arguments that the functions take
are determined by the genetic program. Thus
average(returns(t))(n) generates the arithmetic average of the return observations t, t – 1,…, t – n+1.
The choice of elements to include in the function set is a potentially important one. While a
genetic program can, in principle, produce a very
highly complex solution from simple functions,
computational limitations might make such solutions very difficult to find in practice. Providing
specialized functions to the genetic program that
are thought to be useful to a “good” solution to the
problem can greatly increase the efficiency of the
search by encouraging the genetic program to
search in the area of the solution space containing
those functions. On the other hand, this might
bias the genetic program’s search away from other
promising regions. To focus the search in promising
regions of the solution space, we investigate the
results of adding three additional complex data
functions to the original set of functions. This is
described below.
The expanded set of data functions consists of
the original set plus geo, mem, and arch5. Each of
these functions approximates the forecast of a
known parametric model of conditional volatility.
Thus, the genetic program might find them useful.
The function geo returns the following weighted
average of ten lags of past data:
9
(6)
[
]
geo( data)(α ) ≡ ∑ α (1 − α ) data t − j .
j =0
j
This function can be derived from the prediction
of an IGARCH specification with parameter α, where
we constrain α to satisfy 0.01 ≤ α ≤ 0.99 and ten
lags of data are used. The function mem returns a
weighted sum similar to that which would be
obtained from a long memory specification for
volatility. It takes the form
9
(7)
mem( data)( d ) ≡ ∑ h j data t − j ,
j =0
M AY / J U N E 2 0 0 2
45
REVIEW
Neely and Weller
Figure 1
Example of a Hypothetical Forecast Function
(8/π )arctan(•)+ 4
*
0.1
Max(sum of squared intraday returns)(•)
5
for i = 1, 2, 3, and 4
for i = 5
5
4
(8) arch5( data)(h) ≡ ∑ h j data t − j , h = (h0 , h1,K, h4 ).
j =0
Figure 1 illustrates a simple example of a hypothetical tree determining a forecasting function.
The function first computes the maximum of the
sum of squared intraday returns over the last five
days. This number is multiplied by 0.1, and the
result is entered as the argument x of the function
(8/π )arctan(x)+4. This latter function is common
to all trees and maps the real line into the interval
(0,8). It ensures that all forecasts are nonnegative
and bounded above by a number chosen with reference to the characteristics of the in-sample period.
We now turn to the form of the fitness criterion.
Because true volatility is not directly observed, it is
necessary to use an appropriate proxy in order to
assess the volatility forecasting performance of
M AY / J U N E 2 0 0 2
S
Ri ,t = 100 ⋅ ln i +1,t
Si ,t
(9)
S
= 100 ⋅ ln 1,t +1
S5,t
σ I2,t = ∑ R i2,t .
(10)
where h0=1, h j ∝ (1/j!)(d+j –1)(d+j –2)…(d+1)d for
j>0 and the sum of the coefficients h j is constrained
to equal 1 so that the output would be of the same
magnitude as recent volatility. The parameter d is
determined by the genetic program and constrained
to satisfy –1<d<1. Finally, the function arch5 permits a flexible weighting of the five most recent
observations, where the values for h j are provided
by the genetic program and constrained to lie within
{–5,5} and to sum to 1. Again, the constraint on
the sum of the coefficients ensures that the magnitude of the output will be similar to that of recent
volatility. The function has the form
46
the genetic program. One possibility is to use the
ex post squared daily return. However, as Andersen
and Bollerslev (1998) have pointed out, this is an
extremely noisy measure of the true underlying
volatility and is largely responsible for the apparently poor forecast performance of GARCH models.
A better approach is to sum intraday returns to measure true daily volatility (i.e., integrated volatility)
more accurately. We measure integrated volatility
using five irregularly spaced intraday observations.
If Si,t is the i th observation on date t, we define
i =1
2
σ I,t
is the measure of integrated volatility on
Thus
date t.2 Using five intraday observations represents
a compromise between the increase in accuracy
generated by more frequent observations and the
problems of data handling and availability that arise
as one moves to progressively higher frequencies
of intraday observation.
In constructing the rules, the genetic program
minimized the mean-squared forecast error (MSE)
as the fitness criterion. There are potential inefficiencies involved in using this criterion on heteroskedastic data. However, a heteroskedasticitycorrected fitness measure proved unsatisfactory in
experiments. With three to five observations per day,
there were instances where the integrated daily
volatility was very small; the heteroskedasticity
correction caused the measure to be inappropriately
sensitive to those observations.3
2
More precisely, daily volatility is calculated from 1700 Greenwich
Mean Time (GMT) to 1700 GMT.
3
A perennial problem with using flexible, powerful search procedures
like genetic programming is overfitting—the finding of spurious
patterns in the data. Given the well-documented tendency for the
genetic program to overfit the data, it is necessary to design procedures
to mitigate this (e.g., Neely, Weller, and Dittmar, 1997). Here, we investigated the effect of modifying the fitness criterion by adding a penalty
for complexity. This penalty consisted of subtracting an amount
(0.002 × number of nodes) from the negative MSE. Nodes are data and
numerical functions. This modification is intended to bias the search
toward functions with fewer nodes, which are simpler and therefore
less prone to overfit the data. Unfortunately, this procedure produced
no significant changes in performance, so we will report results only
from the unmodified version.
FEDERAL RESERVE BANK OF ST. LOUIS
Neely and Weller
Table 1
Data Type and Source
Time
Source
Type of price
1000
Swiss National Bank
Triangular arbitrage on bid rates
1400
Federal Reserve Bank of New York
Midpoint of bid and ask
1600
Bank of England
Triangular arbitrage, unspecified
1700
Federal Reserve Bank of New York
Midpoint of bid and ask
2200
Federal Reserve Bank of New York
Midpoint of bid and ask
DATA AND IMPLEMENTATION
The object of this exercise is to forecast the
daily volatility (the sum of intraday squared returns)
of two currencies against the dollar, the German
mark (DEM) and Japanese yen ( JPY), over the period
June 1975 to September 1999. Thus, the final nine
months of data for the DEM represent the rate derived
from that of the euro, which superseded the DEM
in January 1999. The timing of observations was
1000, 1400, 1600, 1700, and 2200 GMT. Days with
fewer than three valid observations or no observation at 1700 were treated as missing. In addition,
weekends were excluded. The sources of the data
for both exchange rates are summarized in Table 1.
We provided the genetic program with three series
in addition to the integrated volatility series: daily
returns, sum of absolute intraday returns, and the
number of days until the next trading day.
The full sample is divided into three subperiods:
the training period June 1975 through December
1979; the selection period January 1980 through
December 30, 1986; and the out-of-sample period
December 31, 1986, through September 21, 1999.
The role of these subperiods is described below.
In searching through the solution space of forecasting functions, the genetic program followed the
procedures below.
1. Create an initial generation of 500 randomly
generated forecast functions.
2. Measure the MSE of each function over
the training period and rank according to
performance.
3. Select the function with the lowest MSE and
calculate its MSE over the selection period.
Save it as the initial best forecast function.
4. Select two functions at random, using weights
attaching higher probability to more highly
ranked functions. Apply the crossover oper-
ator to create a new function, which then
replaces an old function, chosen using weights
attaching higher probability to less highly
ranked functions. Repeat this procedure 500
times to create a new generation of functions.
5. Measure the MSE of each function in the new
generation over the training period. Take the
best function in the training period and evaluate the MSE over the selection period. If it
outperforms the previous best forecast, save
it as the new best forecast function.
6. Stop if no new best function appears for 25
generations, or after 50 generations. Otherwise, return to stage 4.
The stages above describe one trial. Each trial
produces one forecast function. The results of each
trial will generally differ as a result of sampling
variation. For this reason it is necessary to run a
number of trials and then to aggregate the results.
The aggregation methods are described in the following section.
RESULTS
The benchmark results are those from the
GARCH(1,1) and RiskMetrics models described in
the Benchmark Model section, estimated over the
in-sample period June 1975 to December 30, 1986.
We forecast daily integrated volatility (defined in
equations (9) and (10)) from these models, in and out
of sample, at horizons of 1, 5, and 20 days.4
We also forecast with a genetic program whose
training and selection periods coincide with the insample estimation period for the GARCH model. For
each case of the genetic program we generated ten
trials, each of which produced a forecast function.
4
Note that the forecasted variable at the 5-day (20-day) horizon is the
integrated volatility 5 (20) days in the future. It is not the sum of the
next 5 (20) days of integrated volatility.
M AY / J U N E 2 0 0 2
47
REVIEW
Neely and Weller
Table 2
: The Baseline Case
In-Sample Comparison of Genetic Program, GARCH, and RiskMetrics
Exchange
rate
Horizon
DEM
MSE
EW GP MW GP GARCH
R2
MAE
RM
EW GP MW GP GARCH
RM
EW GP MW GP GARCH
RM
1
0.50
0.53
0.50
0.49
0.30
0.33
0.33
0.33
0.18
0.15
0.16
0.14
DEM
5
0.56
0.59
0.56
0.52
0.31
0.34
0.37
0.34
0.12
0.11
0.10
0.10
DEM
20
0.61
0.63
0.67
0.56
0.33
0.34
0.46
0.37
0.08
0.04
0.04
0.05
1
0.56
0.58
0.60
0.62
0.32
0.32
0.38
0.37
0.22
0.20
0.14
0.08
JPY
5
0.65
0.65
0.73
0.66
0.36
0.37
0.43
0.38
0.06
0.04
0.02
0.04
JPY
20
0.66
0.67
0.71
0.69
0.38
0.39
0.51
0.40
0.05
0.03
0.01
0.02
JPY
NOTE: The in-sample mean-squared error (MSE), mean absolute error (MAE), and R2 from GARCH(1,1), RiskMetrics (RM), and genetic
program (GP) forecasts on DEM/USD and JPY/USD data at three forecast horizons: 1, 5, and 20 days. The GP forecast was generated
using five data functions and without a penalty for complexity. In columns 3, 7, and 11 we report the forecast statistics—MSE, MAE, and
R2—for the equally weighted (EW) genetic programming method. In columns 4, 8, and 12 we report the analogous statistics for the
median-weighted (MW) genetic programming forecast. Columns 5, 9, and 13 contain the results for the GARCH forecast. Columns 6,
10, and 13 contain RiskMetrics forecast statistics. The in-sample period was June 1975 to December 30, 1986.
Figure 2
One-Day-Ahead Forecasting Functions for
the DEM
(8/π )arctan(•)+ 4
log(•)
log(•)
Geo(Ndays)(•)
Geo(sum of squared intraday returns)(•)
–0.4744
The cases were distinguished by the following
factors: (i) forecast horizon—1, 5, and 20 days; (ii) the
number of data functions—five or eight. For each
case, we generated ten rules. The forecasts from each
set of ten rules were aggregated in two ways. The
equally weighted forecast is the arithmetic average
of the forecasts from each of the ten trials. The
median-weighted forecast takes the median forecast
from the set of ten forecasts at each date. We report
six measures of out-of-sample forecast performance:
48
M AY / J U N E 2 0 0 2
MSE, mean absolute error (MAE), R2, mean forecast
bias, kernel estimates of the error densities, and generalized mean-squared forecast error matrix tests.
Before discussing the results, we first present a
simple example of the forecasting rules produced
by the genetic program. Figure 2 illustrates a oneday-ahead forecasting function for the DEM. Its
out-of-sample MSE was 0.496. The function is interpreted as follows. The number –0.4744 at the terminal node enters as the argument of geo(sum of
squared intraday returns). Since the argument of
geo is constrained to lie between 0.01 and 0.99, it is
set to 0.01. The number generated by this function
then enters as the argument in geo(Ndays), where
Ndays refers to the “number of days to the next trading day.” We caution that this example was chosen
largely because of its relatively simple form; some
trials generated rules that were considerably more
complex, with as many as 10 levels and/or 100 nodes.
Table 2 reports in-sample results for the baseline
case with five data functions. The figures for MSE
for the DEM are very similar for the GARCH and
equally weighted genetic program forecasts at the
1- and 5-day horizons, but the genetic program is
appreciably better at the 20-day horizon. The
median-weighted forecast is generally somewhat
inferior to the equally weighted forecast, but follows
the same pattern over the forecast horizons relative
to the GARCH model. That is, its best relative performance is at the 20-day horizon. The RiskMetrics
forecasts also are generally comparable to GARCH
FEDERAL RESERVE BANK OF ST. LOUIS
Neely and Weller
Table 3
: The Baseline Case
Out-of-Sample Comparison of Genetic Program, GARCH, and RiskMetrics
Exchange
rate
Horizon
DEM
MSE
EW GP MW GP GARCH
R2
MAE
RM
EW GP MW GP GARCH
RM
EW GP MW GP GARCH
RM
1
0.35
0.38
0.33
0.32
0.30
0.34
0.33
0.32
0.09
0.08
0.12
0.10
DEM
5
0.38
0.42
0.36
0.34
0.31
0.35
0.35
0.33
0.06
0.06
0.08
0.07
DEM
20
0.41
0.42
0.44
0.37
0.31
0.31
0.43
0.35
0.01
0.01
0.02
0.02
1
1.35
1.35
1.29
1.33
0.42
0.44
0.47
0.47
0.14
0.13
0.16
0.11
JPY
5
1.48
1.48
1.56
1.44
0.43
0.45
0.52
0.49
0.03
0.02
0.04
0.06
JPY
20
1.48
1.48
1.43
1.46
0.45
0.46
0.55
0.51
0.02
0.02
0.05
0.05
JPY
NOTE: The out-of-sample MSE, MAE, and R2 from GARCH(1,1), RiskMetrics (RM), and genetic program (GP) forecasts on DEM/USD
and JPY/USD data at three forecast horizons: 1, 5, and 20 days. The GP forecast was generated using five data functions and without a
penalty for complexity. The out-of-sample period was December 31, 1986, to September 21, 1999. See the notes to Table 2 for column
definitions.
Table 4
Out-of-Sample Results Using the Data Functions Geo, Mem, and Arch5
Exchange
rate
Horizon
MSE
EW GP MW GP GARCH
R2
MAE
RM
EW GP MW GP GARCH
RM
EW GP MW GP GARCH
RM
DEM
1
0.37
0.44
0.33
0.32
0.29
0.37
0.33
0.32
0.12
0.05
0.12
0.10
DEM
5
0.36
0.37
0.36
0.34
0.30
0.30
0.35
0.33
0.05
0.04
0.08
0.07
DEM
20
0.38
0.38
0.44
0.37
0.30
0.30
0.43
0.35
0.01
0.01
0.02
0.02
JPY
1
1.27
1.31
1.29
1.33
0.43
0.44
0.47
0.47
0.18
0.15
0.16
0.11
JPY
5
1.45
1.46
1.56
1.44
0.46
0.46
0.52
0.49
0.04
0.03
0.04
0.06
JPY
20
1.49
1.62
1.43
1.46
0.44
0.50
0.55
0.51
0.04
0.00
0.05
0.05
NOTE: The out-of-sample MSE, MAE, and R2 from GARCH(1,1), RiskMetrics (RM), and genetic program (GP) forecasts on DEM/USD
and JPY/USD data at three forecast horizons: 1, 5, and 20 days. The GP forecast was generated using eight data functions including
geo, mem, and arch5 (for descriptions see equations (6) through (8) in the text) and without a penalty for complexity. The out-of-sample
period was December 31, 1986, to September 21, 1999. See the notes to Table 2 for column definitions.
forecasts at 1- and 5-day horizons, but a bit better
at longer horizons. For the JPY, the genetic program
produces equally weighted MSE figures that are in
all cases lower than for the GARCH and RiskMetrics
models. Similarly, the equally weighted genetic
programming rules have higher R2s over each
horizon than the GARCH and RiskMetrics models.
This result is not especially surprising given the
flexibility of the nonparametric procedure and its
known tendency to overfit in-sample.
Table 3 presents a more interesting comparison—out-of-sample performance over the period
December 31, 1986, through September 21, 1999.
The equally weighted genetic program MSE figures
are usually slightly larger than those of the GARCH
and RiskMetrics forecasts at all horizons for both
currencies. Similarly, the genetic programming R2s
are typically slightly smaller than those of the
GARCH/RiskMetrics forecasts. However, the equally
weighted genetic program has a lower MAE than
do the GARCH/RiskMetrics models at all horizons
and for both currencies.
Table 4 reports the out-of-sample performance
of the genetic program forecasts using the augmented
M AY / J U N E 2 0 0 2
49
REVIEW
Neely and Weller
Figure 3
1-Day DEM Forecast Error Densities
2.5
2
1.5
1
0.5
0
–2
–1
0
GP
1
2
GARCH
1-Day JPY Forecast Error Densities
2.5
2
1.5
1
0.5
0
–2
–1
0
GP
1
2
GARCH
NOTE: The kernel estimates of the densities of the 1-day
forecast errors (forecast minus realized volatility) for the
DEM and JPY for genetic program and GARCH(1,1) model
over the out-of-sample period, December 31, 1986, through
September 21, 1999. The dotted vertical line denotes zero.
set of data functions, which include geo, mem, and
arch5. For ease of comparison Table 4 repeats the
out-of-sample figures for the GARCH model. The MSE
and R2 statistics from this table are more equivocal
than those from Table 3. The equally weighted
genetic program MSE for the DEM cases are slightly
larger than those of the GARCH and RiskMetrics
forecasts at the 1- and 5-day horizons, but the genetic
50
M AY / J U N E 2 0 0 2
program performs somewhat better than GARCH at
the 20-day horizon. This performance is not, however, reflected in the R2, for which the GARCH/
RiskMetrics models are better at longer horizons.
For the JPY the situation is reversed; the equally
weighted genetic programming MSE is lower than
the GARCH/RiskMetrics figures at the 1-day horizon but larger at the 20-day horizon. The equally
weighted genetic program also has a slight edge in
R2 at the 1-day horizon. The figures for the MAE of
the genetic program are not very different from
Table 3 and are still substantially better than those
of the GARCH/RiskMetrics predictions.
To summarize: With MSE as the performance
criterion, neither the genetic programming procedure nor the GARCH/RiskMetrics model is clearly
superior. The GARCH/RiskMetrics models do
achieve slightly higher R2s at longer horizons but the
MAE criterion clearly prefers the genetic programming forecasts. In both tables, there is some tendency
for the median-weighted genetic programming forecast to perform less well than its equally weighted
counterpart. The out-of-sample RiskMetrics forecasts are usually marginally better than those of the
estimated GARCH model by MSE and MAE criteria
but marginally worse when judged by R2.
Comparing the genetic programming results in
Table 4 with those of Table 3 shows that expanding
the set of data functions leads to only a marginal
improvement in the performance of the genetic
program. Therefore further results will concentrate
on out-of-sample forecasts in the baseline genetic
programming case presented in Table 3, where only
five data functions were used. We present kernel
estimates of the densities of out-of-sample forecast
errors at the various horizons in Figures 3 through 5.5
The most striking feature to emerge from these
figures is the apparent bias in the GARCH forecasts
when compared with their genetic program counterparts. At all forecast horizons and for both currencies, there is a positive shift in the error distributions
of the GARCH forecasts that move the modes of the
forecast densities away from zero. However, the
relative magnitude of the bias in the mode does not
carry over to the mean. Table 5 shows that, though
both forecasts are biased in the mean, the magnitude
of the bias is considerably greater for the genetic
program. Tests for the bias—carried out with a
Newey-West correction for serial correlation—show
5
We choose to graph the density of the GARCH errors because the
density of the RiskMetrics errors will have a mean of approximately
zero by construction.
FEDERAL RESERVE BANK OF ST. LOUIS
Neely and Weller
Figure 4
Figure 5
5-Day DEM Forecast Error Densities
20-Day DEM Forecast Error Densities
2.5
2.5
2
2
1.5
1.5
1
1
0.5
0.5
0
0
–2
–1
0
GP
1
–2
2
–1
0
GP
GARCH
1
5-Day JPY Forecast Error Densities
20-Day JPY Forecast Error Densities
2.5
2.5
2
2
1.5
1.5
1
1
0.5
0.5
0
–2
–1
0
GP
1
2
GARCH
0
–2
–1
0
GP
2
GARCH
1
2
GARCH
NOTE: The kernel estimates of the densities of the 5-day
forecast errors (forecast minus realized volatility) for the
DEM and JPY for genetic program and GARCH(1,1) model
over the out-of-sample period, December 31, 1986, through
September 21, 1999. The dotted vertical line denotes zero.
NOTE: The kernel estimates of the densities of the 20-day
forecast errors (forecast minus realized volatility) for the
DEM and JPY for genetic program and GARCH(1,1) model
over the out-of-sample period, December 31, 1986, through
September 21, 1999. The dotted vertical line denotes zero.
that all the forecasts are biased in a statistically
significant way (Newey and West, 1987, 1994). The
evidence from Figures 3 through 5—that the modes
of the genetic programming error distribution are
closer to zero than those of the GARCH model—
indicates that the bias in the genetic programming
forecasts is being substantially influenced by a
small number of negative outliers.
The MSE and R2 evidence presented so far fails
to indicate a clear preference for any of the four
sets of forecasts. The best model varies by forecast
horizon and by forecast evaluation criterion. This
confused state of affairs leaves one wondering
whether these disparate results can be reconciled
to produce an unambiguous ranking of the two
methodologies. One method by which multi-horizon
M AY / J U N E 2 0 0 2
51
REVIEW
Neely and Weller
Table 5
Tests for Mean Forecast Bias
Equally weighted
GP
Exchange
Mean Predicted
rate
Horizon
σ2
σ2
Bias
Median-weighted
GP
Bias Predicted
p value
σ2
Bias
GARCH
Bias Predicted
p value
σ2
Bias
RM
Bias Predicted
p value
σ2
Bias
Bias
p value
DEM
1
0.43
0.29
–0.15
0.00
0.24
–0.19
0.00
0.46
0.03
0.00
0.43
0.00
0.92
DEM
5
0.43
0.23
–0.20
0.00
0.16
–0.27
0.00
0.49
0.05
0.00
0.43
0.00
0.92
DEM
20
0.43
0.19
–0.24
0.00
0.18
–0.25
0.00
0.59
0.16
0.00
0.43
0.00
0.92
JPY
1
0.56
0.33
–0.22
0.00
0.32
–0.24
0.00
0.57
0.02
0.06
0.55
0.00
0.88
JPY
5
0.56
0.37
–0.19
0.00
0.41
–0.15
0.00
0.59
0.04
0.07
0.55
–0.01
0.87
JPY
20
0.56
0.42
–0.14
0.00
0.44
–0.12
0.01
0.65
0.09
0.02
0.55
–0.01
0.89
NOTE: In column 3, mean volatility is the mean daily integrated volatility over the out-of-sample period December 31, 1986, through
September 21, 1999. Columns 4, 5, and 6 report the following statistics for the equally weighted genetic programming forecasts over
the same period: mean forecast of integrated volatility, the bias in the forecast (predicted volatility minus realized volatility), and the
p value for the test that the mean bias is zero. Columns 7 through 9 report the statistics for the median-weighted genetic programming
forecasts, and columns 10 through 12 report the analogous results for GARCH forecasts. The RiskMetrics statistics are in columns 13
through 15. The genetic program forecasts are based on the five-function model described in Table 3. The p values are computed with
Newey-West corrections for heteroskedasticity and serial correlation. The lag length was selected by the Newey and West (1994)
procedure.
Table 6
Test of Generalized Method of Second Forecast Error Moment Domination
Eigenvalues
GARCH-EW GP
DEM
JPY
GARCH-MW GP
RM-EW GP
RM-MW GP
GARCH-RM
–0.090
–0.148
0.012
–0.003
–0.021
–0.037
–0.017
0.003
–0.138
0.086
0.082
0.079
–0.084
–0.002
0.036
–0.369
–0.359
–0.028
–0.035
–0.295
0.145
0.127
0.055
0.059
0.200
0.199
0.203
–0.101
–0.102
0.144
NOTE: Table 6 provides sets of eigenvalues for the test of generalized method of second forecast error moment criterion. The first model
dominates the second model if all the eigenvalues in a set are nonpositive and at least one is negative. GARCH-EW GP denotes the
GARCH model versus the equally weighted genetic programming forecasts for the baseline case, as in Table 3. GARCH-MW GP
denotes the GARCH model versus the median-weighted genetic programming forecasts for the baseline case, as in Table 3. RM-EW
GP denotes the RiskMetrics model versus the equally weighted genetic programming forecasts for the baseline case. RM-MW GP
denotes the RiskMetrics model versus the median-weighted genetic programming forecasts for the baseline case. GARCH-RM
denotes the GARCH model versus RiskMetrics forecasts.
forecasts from two sources can be aggregated and
compared is the generalized forecast error second
moment (GFESM) method proposed by Clements
and Hendry (1993). Unfortunately, this method has
some drawbacks. For example, the GFESM can prefer
52
M AY / J U N E 2 0 0 2
model 1 to model 2 based on forecasts from horizon
1 to horizon h, even if model 2’s forecasts dominate
at every forecast horizon up to h. To remedy the perceived weaknesses in the GFESM, Newbold, Harvey,
and Leybourne (1999) proposed the generalized
FEDERAL RESERVE BANK OF ST. LOUIS
mean-squared forecast error matrix (GMSFEM)
criterion. This procedure prefers forecasting method
1 to method 2 if the magnitude of all linear combinations of forecast errors is at least as small under
method 1 as method 2.
To explain the GMSFEM more fully, let us introduce some notation. The one-by-three vector of
1-, 5-, and 20-day GARCH forecast errors at time t
GARCH GARCH GARCH
is etGARCH={et,1
,et,5 ,et,20 }, and the second
moment matrix of these forecast errors is Φ GARCH=
E(etGARCHetGARCH′ ). The RiskMetrics and genetic programming variables are defined analogously. The
GMSFEM says that the GARCH model is preferred
to the genetic programming forecasts if every linear
combination of GARCH forecast errors is at least as
small as every linear combination of genetic programming forecast errors. That is, if
(11)
(
)
d ′ Φ GARCH − Φ GP d ≤ 0
for all vectors d ≠ 0 .6
This condition is met if every eigenvalue of the
matrix (Φ GARCH – Φ GP ) is nonpositive and at least one
is negative. Clearly, the criterion prefers the genetic
programming forecast if every eigenvalue is nonnegative and at least one is positive.
Table 6 shows five sets of eigenvalues from
the (Φ GARCH – Φ GP ) matrix, using both the equally
weighted and median-weighted genetic program
forecasts, for both exchange rates. It confirms the
previous results. The only case in which there are
all negative (or positive) eigenvalues is the comparison of the RiskMetrics forecast to the medianweighted genetic programming forecast. In that
case, all the eigenvalues are negative, indicating
that the RiskMetrics forecasts dominate the medianweighted genetic programming forecasts under the
GMSFEM criterion. In every other set of eigenvalues
there are both positive and negative values. Neither
GARCH/RiskMetrics forecasts nor genetic programming forecasts dominate the other under the
GMSFEM criterion.
Neely and Weller
While the genetic programming rules did not usually
match the GARCH(1,1) or RiskMetrics models’ MSE
or R2, its performance on those measures was
generally close. But the genetic program did consistently outperform the GARCH model on MAE and
modal error bias at all horizons. The genetic programming solutions appeared to suffer from some
in-sample overfitting, which was not mitigated, in
this case, by an ad hoc penalty for rule complexity.
Our results suggest some interesting issues for
further investigation. The superiority of the genetic
program according to the MAE criterion is perhaps
surprising given that we used MSE as the fitness
criterion. This raises the possibility that further
improvement in the forecasting performance of the
genetic program relative to the GARCH model could
be achieved by using MAE as the fitness criterion.
Also, given that increasing the frequency of intraday observations has been shown to improve the
accuracy of forecasts based on the GARCH model
(Andersen et al., 2001), it is important to discover
whether the results of this investigation survive in
that context.
REFERENCES
Andersen, Torben and Bollerslev, Tim. “Answering the
Skeptics: Yes, Standard Volatility Models Do Provide
Accurate Forecasts.” International Economic Review, 1998,
39(4), pp. 885-905.
___________; ___________; Diebold, Francis and Labys,
Paul. “Modeling and Forecasting Realized Volatility.”
Working Paper W8160, National Bureau of Economic
Research, 2001.
Baillie, Richard and Bollerslev, Tim. “The Message in Daily
Exchange Rates: A Conditional-Variance Tale.” Journal
of Business and Economic Statistics, July 1989, 7(3), pp.
297-305.
___________ and ___________. “Intra-Day and Inter-Market
Volatility in Foreign Exchange Rates.” Review of Economic
Studies, May 1991, 58(3), pp. 565-85.
DISCUSSION AND CONCLUSION
We choose to use the problem of forecasting
conditional volatility in the foreign exchange market
to illustrate the strengths and weaknesses of genetic
programming because it is a challenging problem
with a well-accepted benchmark solution, the
GARCH(1,1) model. The genetic program did reasonably well in forecasting out-of-sample volatility.
Bollerslev, Tim. “Generalized Autoregressive Conditional
Heteroskedasticity.” Journal of Econometrics, April 1986,
31(3), pp. 307-27.
Chen, Shu-Heng, ed. Genetic Algorithms and Genetic
6
Note that the GMSFEM criterion implicitly favors models that do well
in terms of MSE, rather than in terms of MAE.
M AY / J U N E 2 0 0 2
53
Neely and Weller
Programming in Computational Finance. New York: Kluwer,
2002 (forthcoming).
REVIEW
(forthcoming in International Review of Economics and
Finance).
Clements, Michael P. and Hendry, David F. “On the Limitations of Comparing Mean Square Forecast Errors.” Journal
of Forecasting, December 1993, 12(8), pp. 617-76.
___________ and Weller, Paul A. “Technical Trading Rules
in the European Monetary System.” Journal of International
Money and Finance, June 1999, 18(3), pp. 429-58.
Engle, Robert F. “Autoregressive Conditional Heteroskedasticity with Estimates of the Variance of U.K. Inflation.”
Econometrica, July 1982, 50(4), pp. 987-1007.
___________ and ___________. “Technical Analysis and
Central Bank Intervention.” Journal of International
Money and Finance, December 2001, 20(7), pp. 949-70.
Holland, John. Adaptation in Natural and Artificial Systems.
Ann Arbor, MI: University of Michigan Press, 1975.
J.P. Morgan/Reuters. RiskMetrics Technical Document,
Part II: Statistics of Financial Market Returns. Fourth
Edition. New York: 1996.
Jorion, Philippe. “Predicting Volatility in the Foreign
Exchange Market.” Journal of Finance, June 1995, 50(2),
pp. 507-28.
___________; ___________ and Dittmar, Robert. “Is Technical
Analysis in the Foreign Exchange Market Profitable? A
Genetic Programming Approach.” Journal of Financial and
Quantitative Analysis, December 1997, 32(4), pp. 405-26.
Newbold, Paul; Harvey, David I. and Leybourne, Stephen L.
“Ranking Competing Multi-Step Forecasts,” in Robert F.
Engle and Halbert White, eds., Cointegration, Forecasting
and Causality. Chap. 4. Oxford: Oxford University Press,
1999.
Koza, John R. Genetic Programming: On the Programming
of Computers by Means of Natural Selection. Cambridge,
MA: MIT Press, 1992.
Newey, Whitney K. and West, Kenneth D. “A Simple Positive,
Semi-Definite, Heteroskedasticity and Autocorrelation
Consistent Covariance Matrix.” Econometrica, May 1987,
55(3), pp. 703-08.
Neely, Christopher J. “Target Zones and Conditional Volatility:
The Role of Realignments.” Journal of Empirical Finance,
April 1999, 6(2), pp. 177-92.
___________ and ___________. “Automatic Lag Selection in
Covariance Matrix Estimation.” Review of Economic
Studies, October 1994, 61(4), pp. 631-53.
___________. “Risk-Adjusted, Ex Ante, Optimal, Technical
Trading Rules in Equity Markets.” Working Paper WP
1999-015D, Federal Reserve Bank of St. Louis, 2001
West, Kenneth D. and Cho, Dongchul. “The Predictive
Ability of Several Models of Exchange Rate Volatility.”
Journal of Econometrics, October 1995, 69(2), pp. 367-91.
54
M AY / J U N E 2 0 0 2
@FedFRASER