See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/275250200

Market-Timing the Business Cycle

Article  in  Review of Financial Economics · March 2015

DOI: 10.1016/j.rfe.2015.03.003

CITATIONS READS

5 561

1 author:

Rolando Pelaez
University of Houston - Downtown
69 PUBLICATIONS   388 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Rolando Pelaez on 09 January 2020.

The user has requested enhancement of the downloaded file.

 Review of Financial Economics 26 (2015) 55–64

Contents lists available at ScienceDirect

Review of Financial Economics

journal homepage: www.elsevier.com/locate/rfe

Market-timing the business cycle

Rolando F. Peláez ⁎
College of Business, University of Houston-Downtown, Houston, TX 77002, USA

a r t i c l e i n f o a b s t r a c t

Article history: As a group, professional portfolio managers have been largely unable to outperform the market buy-and-hold
Received 10 August 2014 benchmark. Likewise, professional forecasters have been unable to predict recessions reliably. The paper
Received in revised form 7 January 2015 contributes to the literature in two significant respects. First, the Recession Probability Model herein correctly
Accepted 16 March 2015
forecasts out-of-sample the probability of a downturn and the binary state over a 45-year validation sample.
Available online 23 March 2015
This is important as it is around cyclical turning points that forecast errors are largest, and dependable forecasts
JEL classification:
are most useful. Reliable recession forecasts are essential for risk-management, planning capital outlays, and for
C35 portfolio management. Moreover, accurate forecasts of the turn allow policy-makers to mitigate the social cost of
C58 recessions. Second, the paper shows that it is extremely profitable to switch from equities to T-bills when the
G11 one-quarter-ahead probability of recession reaches a certain threshold. Several market-timing rules dominate
G17 the buy-and-hold in terms of the risk-adjusted measures of Treynor, Sharpe, and Jensen. One trading rule
achieves triple the terminal wealth of the buy-and-hold.
Keywords:
© 2015 Elsevier Inc. All rights reserved.
Market-timing
Forecasting
Recessions

1. Introduction Section 2 is a brief overview of the historical recession forecasting

record. Section 3 develops the forecasting model. Section 4 evaluates
Market-timing – switching from risky to safer assets in anticipation the model's forecasting accuracy over a validation sample of 181 quar-
of down markets – has a powerful and universal appeal. Market- ters from 1970Q1 to 2015Q1. The main thrust is Section 5. Specifically,
timers utilize the price–earnings ratio, dividend yield, credit spreads, it shows that market-timing rules linked to recession probability
interest rates, and a plethora of technical indicators, evidently with little forecasts, dominate the buy-and-hold in a risk-return framework over
success. Notwithstanding the wealth and effort devoted to the task, the a 45-year span. Section 6 concludes.
fact that most actively managed funds under-perform the market on a
risk-adjusted basis shows that predicting down markets is difficult.1
In a departure from the market-timing literature, this paper links 2. The recession forecasting record
the switch out of stocks to an out-of-sample probability forecast of
recession. Siegel (2014) notes that there are significant gains to switching Considering advances in economic theory, econometric techniques,
from stocks to short-term bonds in advance of a recession, and back to computing technology, and databases, surprisingly forecasters have
stocks before the end of the recession: “If one could predict in advance made little progress in reducing the size of forecast errors. In a large
when recessions will occur, the gains would be substantial. That is perhaps number of countries “the record of failure to predict recessions is virtually
why billions of dollars are spent trying to forecast the business cycle. But unblemished” (Loungani, 2000). Galvão (2006), attempted to predict
the record of predicting business cycle turning points is extremely poor.” recessions using structural break threshold Var models. Although fitted
probabilities were obtained using the complete information set,
i.e., past, present, and future data, one model predicted none and the
⁎ Tel.: +1 832 428 4949.
best predicted only 45% of recessions.
E-mail address: [email protected].
1
See among others, Treynor (1965), Treynor and Mazuy (1966), Sharpe (1966), Jensen Harding and Pagan (2010) examined the predictive power of several
(1968, 1972, 1978), Fama (1970, 1972, 1991), Henriksson and Merton (1981), Merton models and concluded that, “It is very difficult to predict a recession and it
(1981), Henriksson (1984), Chang and Lewellen (1984), Brinson, Hood, and Beebower is only after it is underway that the prediction probability will be high.”
(1986), Ippolito (1989), Malkiel (1995), Carhart (1997), Goetzmann, Ingersoll, and They note that some models forecast one or two recessions ex post,
Ivkovich (2000). More recently, Barras, Scaillet, and Russ (2010) show that during 1975
to 2006, only 0.6% of 2,076 actively managed U.S. open-end equity mutual funds were
but fail to forecast out-of-sample. The Director of the Division of
skilled in terms of Jensen's alpha being positive. However, some studies document superi- Statistics and Research at the Federal Reserve observed, “as I noted in
or returns to market timing, among others, Peláez (1998), and Chen and Liang (2007). the July chart show, the staff [of the FOMC] – and for that matter

http://dx.doi.org/10.1016/j.rfe.2015.03.003
1058-3300/© 2015 Elsevier Inc. All rights reserved.
56 R.F. Peláez / Review of Financial Economics 26 (2015) 55–64

virtually all economic forecasters – fail to reliably forecast recessions” a heightened sense of insecurity, the crisis-induced decline in mental
(Transcript, 2000, p. 11). and physical health, and the impact of the crisis on suicides. Chang,
Consensus forecasts are more indicative of the state-of-the-art than Stuckler, Yip, and Gunnell (2013) estimate that the number of suicides
are individual forecasts. In the U.S., consensus probability forecasts have increased globally by 4884 in 2009 compared to the number expected
never exceeded 0.5 before the start of recessions.2 Consensus forecasts given previous trends. The increase in suicide rates occurred mainly in
of both ASA/NBER and Blue Chip forecasters did not warn of the men in 27 European and 18 American countries. More recently,
1973–75 and 1981–82 recessions (McNees, 1987). March 2001 is the Reeves, McKee, and Stuckler (2014) find evidence of 10,000 economic
peak month in the NBER chronology, yet in the September 2001 survey suicides between 2008 and 2010 related to the Great Recession.
of Blue Chip forecasters only 13% of the respondents thought that the
economy had slipped into recession.
3. The Recession Probability Model (RPM)
The econometric model of the National Institute of Economic and
Social Research (NIESR) failed to forecast the 1974–75 and 1979–81
By convention, the start and end of recessions follows the chronolo-
recessions in the U.K. (Wallis, 1989). This inability was evident again
gy of peaks and troughs from the Business Cycle Dating Committee
with the recession of 1990Q3–1992Q2 during which real GDP in the
(BCDC) of the National Bureau of Economic Research (NBER). The
UK contracted by 3.7%. The highly respected NIESR has been largely
NBER sets both monthly and quarterly chronologies retrospectively
unable to forecast the onset either of recession, or of the subsequent
with an average lag of eleven months. A recession is, “… a significant
recovery (Osborne, Sensier, & Simpson, 2003).
decline in activity spread across the economy, lasting more than a few
Consensus forecasts did not predict the Great Recession. The Survey
months, visible in industrial production, employment, real income less
of Professional Forecasters (SPF) of the Federal Reserve Bank of Phila-
transfer payments, and real wholesale retail sales” (BCDC, 2001).
delphia missed it, and so did the IMF. The SPF release of May 13, 2008,
Regarding the start of recessions, it has issued nuanced statements. “A
predicted growth of 1.7% in 2008Q3, and 1.8% in 2008Q4. Likewise,
recession begins just after the economy reaches a peak of activity and
the May 2008, Outlook Survey of the National Association for Business
ends as the economy reaches its trough” (BCDC, 2001). More recently,
Economics (NABE), had nearly one-half (44%) of the forecasters
“A recession begins when the economy reaches a peak of activity and
expecting no recession in 2008. In the release of August 12, 2008, the
ends when the economy reaches its trough” (BCBC, 2008). The first
SPF forecasters still saw the risk of a contraction in 2008Q4 at only
statement allows the economy to stall before a widespread contraction
46.6%.3 By November 17, 2008, the Outlook Survey was sobering with
ensues. Thus with a peak at the end of a quarter, a contraction may begin
96% of the forecasters finally saying that the economy was in recession.
the following quarter. The second statement implies that the stall and
Transcripts of the FOMC meetings of 2006–07 show that the partic-
contraction occur simultaneously.
ipants considered recession and inflation possibilities, but judged infla-
tion as more pressing. From January 31, 2006 to June 29, 2006, the FOMC
raised the target funds rate in increments of 25 basis points from 4.25% 3.1. Regressors
to 5.25%. The conventional mortgage rate rose from 6.15% in January
2006, to 6.76% by July 2006, triggering resets on adjustable rate mort- Forecasters may choose from a large number of indicators — Stock
gages, and exacerbating the ongoing process of defaults and foreclo- and Watson (2010) examined 270. Given the number of potential
sures. The FOMC began to ease on September 18, 2007, and by April regressors and of possible specifications, it may be simpler to find a nee-
30, 2008, the funds target rate was 2%. It would remain at that level dle in a haystack than to find a robust model. Compounding the prob-
until October 8, 2008. To disabuse the illusion that it is easy to forecast lem, most macroeconomic series are revised, and economists are
recessions, please note that the staff forecast prepared for the FOMC unable to quantify the effects of important factors such as expectations,
meeting of December 2007 saw the economy avoiding recession. inventions, political snags, and animal spirits.
“Overall, our forecast could admittedly be read as still painting a pretty The modeling approach herein aims at parsimony and at excluding
benign picture: Despite all the financial turmoil, the economy avoids all revised regressors. Parsimony is important because if an irrelevant
recession” (Transcript, FOMC Meeting, December 11, 2007, pg. 14). In variable proxies for an omitted effect, biased forecasts ensue when the
June 2008, one staff economist reviewing the participants' economic irrelevant variable undergoes a break. Excluding the breaking variable
projections remarked, “Most of you think the economy will skirt recession” avoids the problem and yields a better model. Likewise, if regressors
(Transcript, FOMC meeting June 24–25, 2008, p. 22). As late as Septem- have small marginal effects forecast variance increases with the number
ber 15, 2008, the conventional wisdom held that the economy would of regressors. A parsimonious model designed to forecast recessions is
not fall into a recession (Greenspan, 2012). Blind to the approaching likely to outperform large-scale macro models such as the FRB/US.5
recession, and confident in the soundness of their value-at-risk models, Greenspan (2000) observed, “Our models do not forecast a recession
several TBTF actors self-destructed, precipitating the greatest financial because… if we have an economy whose growth rate is declining, our
crisis since the Great Depression. In September 2008, JPMorgan was models will have interest rates and costs of capital falling at speeds
forecasting growth through 2009 (Greenspan, 2013). sufficiently rapid to engender a rebalancing of the economy – often
Atkinson, Luttrell, and Rosenblum (2013) estimate that the 2007–09 before it falls into recession.”
crisis cost the U.S. economy $15–$30 trillion, or 100 to 190% of 2007 out- It is important to use unrevised data, as initial releases are biased,
put. This ignores the continuing unemployment4 and underemployment, and revisions are large and do not have zero means (Aruoba, 2008).
Sample-based series – practically all macroeconomic series – are revised
multiple times to reduce sampling error, and are benchmarked period-
2
A large literature documents a disappointing record (See, e.g., Braun & Zarnovitz, ically to revised population totals, to the extreme that the signs of the
1992; Chin, Geweke & Miller, 2000; Fels & Hinshaw, 1968; Loungani, 2000; McNees, original growth rates are often reversed. Since pre- and post-revision
1979, 1987; Moore, 1983; Osborne, Sensier and Simpson, 2003; Stock & Watson, 2003; coefficients necessarily differ, forecasts obtained with real-time data
Wallis, 1989; Zarnovitz, 1967, 1992). Greenspan (2000) warns that economists do not
must vary from those obtained with latest-revised data.6 Additionally,
have the capacity to forecast recessions. Three days before September 15, 2008, JPMorgan
was forecasting growth through 2009. The transcripts of the FOMC meetings of 2007, the
5
Greenbook, and the minutes of the FOMC meetings of 2008, show that as of September FRB/US is the large-scale quarterly econometric model of the U.S. economy developed
2008 the Federal Reserve saw the economy avoiding recession. at the Federal Reserve.
3 6
SPF releases from 1990Q3 are available at http://phil.frb.org/research-and-data/real- The U.S. index of leading economic indicators (LEI), predicts better ex post than in real
time-center/survey-of-profession. time. Diebold and Rudebusch (1991) attribute the deterioration to differences between
4
In November 2014, the number unemployed 15-weeks-and-over was 4,247,000, al- initial and revised data, and to the fact that the LEI undergoes compositional and method-
most double the number of 2,366,000 in November 2007. ological revisions after failing to predict recessions in real time.
 R.F. Peláez / Review of Financial Economics 26 (2015) 55–64 57

the sampling instrument, and sample coverage change over time. There Table 1
is also the issue of periodic sample rotation. When macroeconomic This table shows the specification of the Recession Probability Model. GIC and GSP are
quarterly symmetric growth rates of initial claims and of stock prices respectively, YS is
series undergo definitional/compositional revisions, the new and old the yield spread (10-year Treasury less 3-month T-bill), and NOI and EMI are quarterly av-
series are no longer comparable. In practice, leading indicators must erages of the non-seasonally adjusted New Orders and Employment diffusion indices from
be changed regularly to improve their performance (Hendry, 2003). the Institute for Supply Management.
We limit the potential regressor set to a few financial and real
Regressor Source Release schedule
variables. The financial series include interest rates, term spreads, credit
GICt − 1 U.S. Dept. of Labor Thursdays
spreads, and the growth rate of stock prices. The non-financial series
GSPt − 1 and GSPt − 2 Standard & Poor's Daily
include diffusion indices from the Institute of Supply Management YSt − 3 Board of Governors, FRS Daily
(ISM), indices of consumer expectations/sentiment from the University NOIt − 3 ISM Monthly (1st business day)
of Michigan/Reuters, and the growth rate of initial unemployment EMIt − 2 ISM Monthly (1st business day)
claims. Using the general-to-specific model selection approach,
the Akaike Information Criterion (AIC) is employed in a non-linear
maximum-likelihood setting to select a final model from a general Table 2
unrestricted model in which the aforementioned variables enter with The table shows the NBER monthly and (quarterly) chronologies. Announcement dates
since 1980 are available at http://www.nber.org/cycles/cyclesmain.html.
generous lags. General-to-specific (GETS) modeling was popularized by
Hendry (1995, 2000) and Hendry and Krolzig (2001). Greene (2008) Peak Announcement Trough Announcement
notes that econometricians have turned away from the traditional July 1953(II) May 1954(II)
simple-to-general approach and now prefer the general-to-specific August 1957(III) April 1958(II)
approach. April 1960(II) Feb 1961(I)
Dec 1969(IV) Nov 1970(IV)
Table 1 shows the selected Recession Probability Model (RPM). The
Nov 1973(IV) March 1975(I)
model eschews the “big four” indicators of the BCDC. The first two January 1980(I) June 3, 1980 July 1980(III) July 8, 1981
regressors are symmetric growth rates of initial claims, and of stock July 1981(III) January 6, 1982 Nov 1982(IV) July 8, 1983
prices; the remaining three are the yield spread, and two non- July 1990(III) April 25, 1991 March 1991(I) Dec 22, 1992
seasonally adjusted diffusion indices from the Manufacturing Report March 2001(I) Nov 26, 2001 Nov 2001(IV) July 17, 2003
Dec 2007(IV) Dec 1, 2008 June 2009(II) Sept 20, 2010
on Business of the Institute for Supply Management (ISM).7 Diffusion
indices measure the breadth of change, not the magnitude.
Notwithstanding the decline of U.S. manufacturing, both the employ- Estimating the model by logit sampling 1954Q1–2014Q4 yields:
ment and new orders diffusion indices are level-stationary. All series
are unrevised and are in the public domain.8
Coefficient Std.error t-value t-prob
An extensive literature documents the predictive power of the yield
spread, see among others, Berk (1998), Estrella and Mishkin (1998), Constant 33.7796 11.96 2.82 0.005
GICt − 1 1.94359 0.6395 3.04 0.003
Chauvet and Potter (2005), and Wright (2006). Rudebusch and
GSPt − 1 −1.48817 0.4873 −3.05 0.003
Williams (2008) show that a yield curve model is a better predictor of GSPt − 2 −1.18349 0.4862 −2.43 0.016
recessions than the consensus of professional forecasters. The growth YPt − 3 −4.84788 1.762 −2.75 0.006
rates of initial claims and of stock prices are among thirteen other EMIt − 2 −0.364705 0.1846 −1.98 0.049
components of the Leading Economic Index (LEI) of the Conference NOIt − 3 −0.338474 0.1952 −1.73 0.084

Board. The specification in Table 1 is slight improvement from that in Log-likelihood −11.5430219 No. of states 2
Peláez (2015) which yields a false alarm in 2007Q3.9 The LEI has zero No. of 244 No. of parameters 7
marginal predictive power beyond that of Peláez (2015) model. The observations
chronology of peaks and troughs from the Business Cycle Dating Baseline log-lik −112.0568 Test: Chi^2(6) 201.03 [0.0000]**
AIC 37.0860438 AIC/n 0.151991983
Committee (BCDC) of the National Bureau of Economic Research
Mean (SHD) 0.172131 Var (SHD) 0.142502
(NBER) delineates the state. See Table 2. Appendix A describes the
data in detail. Count Frequency Probability loglik

State 0 202 0.82787 0.82787 −4.287

3.2. Full-sample estimation results State 1 42 0.17213 0.17213 −7.256
Total 244 1.00000 1.00000 −11.54
Given a regressor set, x′ = (x1t, x2t,…, xnt), and a binary dependent
variable, Zt, defined as unitary from peak to trough, and zero otherwise, Summary statistics for explanatory variables.
the one-quarter-ahead conditional probability of recession is
P(Zt + 1|xt) = π(x). Likewise, P(Zt + 1|xt) = 1 − π(x) is the probability
#Obs Min Mean Max Std. dev
that the economy will expand. Taking logs of the odds ratio yields
the logit, State Zt = 0: Expansion
Constant 202 1 1 1 0
GICt−1 202 −8.1525 −0.77365 6.1283 2.3353
  −9.3038
 π ðxÞ  0
GSPt−1 202 1.0098 7.411 2.112
ln  ¼βx: ð1Þ GSPt−2 −9.3038
1−πðxÞ t 202 1.0272 7.411 2.2038
YPt−3 202 −1.3667 1.7203 3.8 1.0518
EMIt−2 202 27.5 50.277 67 7.1491
NOIt−3 202 25.833 56.405 71.167 8.0252
7
The ISM also publishes seasonally adjusted indices; the paper does not use those be-
State Zt = 1: Recession
cause they are revised.
8 Constant 42 1 1 1 0
The initial claims series is based on the universe of transactions. Not being sample-
GICt−1 42 −0.16051 3.8857 11.943 2.4325
based, it is not benchmarked, but it is seasonally adjusted, and the seasonal factors under-
GSPt−1 42 −10.865 −1.3237 3.656 3.1851
go revision.
9 GSPt−2 42 −10.865 −1.4485 2.5635 2.7805
Peláez (2015) does not consider market timing. His recession forecasting model lags
YPt−3 42 −1.43 0.27087 2.37 0.9265
the ISM production and employment diffusion indices three quarters, whereas the model
EMIt−2 42 31.333 45.893 57.5 6.4525
in Table 1 does not employ the production index, and lags the employment index two
NOIt−3 42 35 51.21 67.167 6.0873
quarters.
58 R.F. Peláez / Review of Financial Economics 26 (2015) 55–64

Table 3 Table 4
This table shows the results of testing the null of parameter constancy using Nyblom's test. This table shows annualized percent quarterly growth rates for 1973–74. TNFP is total
The results do not reject the null over 1954Q1–2014Q4. non-farm payrolls (U.S. Department of Labor). PIN is real personal income less transfer
payments (Conference Board). IPI is the industrial production index (Federal Reserve).
Test statistic P-value MTS is real manufacturing and trade sales (Conference Board). CCI is the Composite
Joint Test 0.838 0.794 Coincident Index of the previous four measures (Conference Board). GDP is real gross
Single Coefficients domestic product (U.S. Department of Commerce).
Constant 0.082 0.662
Date TNFP PIN IPI MTS CCI GDP
GICt − 1 0.141 0.402
GSPt − 1 0.111 0.515 1973Q1 5.3 1.3 11.3 11.0 6.1 10.2
GSPt − 2 0.095 0.591 1973Q2 3.6 3.2 3.2 −3.6 2.9 4.6
NOIt − 3 0.062 0.792 1973Q3 2.4 2.8 3.4 −0.2 2.4 −2.2
EMIt − 2 0.066 0.766 1973Q4 3.8 5.2 5.9 11.7 5.4 3.8
YPt − 3 0.057 0.822 1974Q1 1.9 −7.7 −3.7 −2.0 −1.5 −3.3
1974Q2 1.5 −4.6 0.4 −2.0 −0.5 1.1
1974Q3 0.6 −0.2 −1.7 −5.0 −0.5 −3.8
As expected, the probability of recession varies directly with the
growth rate of initial claims and inversely with the other regressors.
Collinearity between the employment and new orders diffusion indices
Table 5
accounts for their relatively low t-ratios. This table shows annualized percent quarterly growth rates for 2007–08. TNFP is total
Parameter instability is a leading cause of forecast failure. Nyblom's non-farm payrolls (U.S. Department of Labor). PIN is real personal income less transfer
(1989) test of the null of parameter constancy allows testing coefficients payments (Conference Board). IPI is the industrial production index (Federal Reserve).
individually and jointly. The test employs the cumulative sum of the MTS is real manufacturing and trade sales (Conference Board). CCI is the Composite
Coincident Index of the previous four measures (Conference Board). GDP is real gross
score function; it allows for breaks of unknown timing and for random
domestic product (U.S. Department of Commerce).
walk parameters. The test results in Table 3 do not reject the null of Annualized quarterly growth rates (%), 2007–08.
parameter constancy for the coefficients individually or jointly. This is
Year-Q TNFP PIN IPI MTS CCI GDP
noteworthy given the diversity of shocks to the economy and structural
changes over more than sixty years. 2007Q3 0.2 0.6 1.1 2.0 0.4 2.7
2007Q4 0.7 0.1 1.0 1.6 1.0 1.5
2008Q1 0.1 2.1 −1.4 −5.2 0.6 −2.7
4. Recursive forecasts 2008Q2 −1.4 −3.8 −5.7 −2.7 −2.7 2
2008Q3 −2.1 −3.2 −12.9 −13.8 −5.7 −2
2008Q4 −4.8 0.5 −17.3 −19.8 −7.4 −8.3
The earliest date for estimation is 1954Q1, since April 1953 is the
inception date of the 10-year Treasury.10 To simulate a real-time fore-
caster we initialize the model over 1954Q1–1969Q4, and obtain a
one-quarter-ahead probability forecast for 1970Q1. Thereafter the esti-
mation sample grows one observation at a time, but all forecasts for Table 6
t + 1 utilize only regressors and coefficients through t. Having to This table shows that all monthly levels of economic activity, except personal income,
make hard yes/no decisions, forecast users may prefer categorical fore- peaked on January 2008, and began to contract on February 2008.

casts. To assign a categorical forecast to a probability we need a decision Date TNFP PIN IPI MTS CCI
threshold, p⁎. The usual binary event classifier is that of the maximum Sep-07 137,752.0 10,613.3 100.7 1,146,560.0 107.3
probability rule in which p* = 0.5. Based on the most probable state, Oct-07 137,838.0 10,592.9 100.2 1,149,913.0 107.3
^tþ1 ≥0:5, and p
p ^tþ1 b 0:5 signal recession and expansion respectively. Nov-07 137,949.0 10,588.8 100.8 1,147,748.0 107.4
Since the NBER announces the turn with an average lag of one-year, Dec-07 138,042.0 10,622.4 100.8 1,142,148.0 107.5
Jan-08 138,056.0 10,641.3 100.5 1,146,546.0 107.7
a real-time forecaster typically knows the official chronology through
Feb-08 137,971.0 10,656.5 100.3 1,132,020.0 107.6
t − 4. Therefore, his forecast at time t pivots on his previous forecasts Mar-08 137,892.0 10,670.4 100.0 1,116,685.0 107.4
for t − 1 through t − 4. To foreclose an unfair advantage, for a moving Apr-08 137,677.0 10,606.0 99.2 1,131,060.0 107.2
window of t − 4 through t − 1, the categorical forecasts obtained
with the maximum probability rule replace the official chronology, i.e., if
p ^t jϕt−1 b 0:5;
^t jϕt−1 ≥0:5 then at the next step zt − 1 = 1.0. Likewise, if p increased modestly that month because the last part of the capital
at the next step zt−1 ¼ 0: Thus, categorical forecast errors, if any, are gains tax cuts that George W. Bush enacted in 2003 took effect on Janu-
endogenous within a moving window of four quarters. Fig. 1 shows ary 1, 2008.11 The fact that the model is right on both counts does not
the recursive forecasts (diamonds); the shaded bars correspond to the imply that the NBER peak months are wrong. Obviously, the peak date
NBER quarterly chronology in Table 2. and the onset of recession need not coincide as a broad contraction
The probabilities for 1973Q4 and 2007Q4 are close to zero. At may begin shortly after a peak.
first sight, the model erred since the respective peaks in the NBER chro- Incorrectly coding 1973Q4 as a recession quarter affects the coeffi-
nology are November 1973 (1973Q4), and December 2007 (2007Q4). cient vector going forward. Obtaining new forecasts coding 1973Q4 as
However, a closer look acquits the model. Table 4 shows that the econ- expansion, Zt = 0, and 1974Q1 as recession, Zt = 1, significantly
omy grew rapidly in 1973Q4, and that the recession began in 1974Q1. improves the model's discriminating ability. See Fig. 2, where forecasts
Likewise, Table 5 shows that the economy grew, albeit slowly in are highly polarized compared to those in Fig. 1.12 The model does not
2007Q4. Table 6 takes a closer look at monthly levels of economic predict recessions that do not occur. This is important, as false positives
activity. Only manufacturing and trade sales (MTS) decreased from may be as costly as false negatives. The single categorical error in 181
November to December 2007. Except for industrial production, the quarters is a recession forecast for 1990Q2, whereas July 1990 is the
economy did not contract in January 2008 either. A broad contraction
began in February 2008, when total nonfarm payrolls, industrial
production, real manufacturing and trade sales, and the composite coin- 11
Couples with taxable income up to $65,100 and singles with taxable income up to
cident index decreased. Personal income less transfer payments
$32,550 would pay a zero percent tax rate on capital gains from the sale of stocks, bonds,
mutual funds, vacation homes and other assets they held for a year or more.
10 12
See, http://research.stlouisfed.org/fred2/series/GS10 (Last accessed December 22, 2014). In Fig. 2, 1973Q4 and 2007Q4 are coded as expansion quarters.
 R.F. Peláez / Review of Financial Economics 26 (2015) 55–64 59

Fig. 1.

NBER peak month. For comparison's sake, Kauppi and Saikkonen (2008) 1990–91, and probability forecasts for all other recessions exceeded
tested the predictive performance of several dynamic quarterly models 0.5 only after they had started.
using the yield spread and lagged values of the indicator variable. They Although forecast accuracy is remarkable, it is worth noting that the
use a longer initialization sample, 1955Q2–1977Q4, and a shorter vali- individual regressors have low predictive power. The one-quarter
dation sample, 1978Q1–2003Q4, which includes only four recessions lagged growth rate of initial claims, GICt − 1, predicted the start of four
versus our seven. A dynamic autoregressive probit model outperformed recessions, yielded five false alarms, and forecasted expansion during
all other models; nevertheless, it completely missed the recession of thirteen quarters of recession. See Fig. 2.1.

Fig. 2.
60 R.F. Peláez / Review of Financial Economics 26 (2015) 55–64

a a GI Ct -

Fig. 2.1.

a a GS Pt - a GS P t -

Fig. 2.2.

Fig. 2.2 shows that over the last 45 years the stock market was positives and no false negatives. Finally, Fig. 2.4 shows that the two
consistently late to the party. The two lagged growth rates of stock diffusion indices have even less predictive power than the stock market.
prices, GSPt − 1 and GSPt − 2, missed all recessions and yielded four
false alarms.13 Fig. 2.3 shows in that terms of consistency and lead-
time, the yield spread is the best predictor with a score of four correct 5. Market timing

Assume an investor who holds an S&P 500 index fund, and who
13
Samuelson (1966) famously quipped, “The stock market has predicted nine of the last wishes to avoid capital losses during recessions. To that end, on
five recessions.” the last business day of each quarter, he forecasts the probability of
 R.F. Peláez / Review of Financial Economics 26 (2015) 55–64 61

a a Y Pt -

Fig. 2.3.

a a E M It- a N OI t -

Fig. 2.4.

^tþ1 ≥0:5; he switches to three-month T-

recession in the next quarter. If p have occurred since 1970, this minimalist strategy entails only seven
bills, and returns to stocks after a pre-determined number of months.14 round trips in 45 years. Table 7 shows the trade dates under a nine-
We consider five naïve trading rules in which the T-bill holding period is month rule. For example, using data through December 1969, the
pre-determined, non-discretionary, and ranges from 9 to 5 months. ^tþ1 ¼ 1:0; prompting a switch from
probability forecast for 1970Q1 is p
Since recession forecasts prompt the switch, and only seven recessions stocks to T-bills and a switch back to the index on September 30, 1970.
Under the buy-and-hold benchmark, an initial investment of $1000
on December 1969 grows to $88,596 on December 2014. Based on the
14
A higher (lower) threshold is warranted if the loss from switching out of stocks due to
actual monthly returns on the S&P 500, and on the monthly bond-
a false alarm is greater (lower) that from not switching when a recession forecast is equivalent yield on secondary market three-month T-bills, the same
correct. $1000 grows to $300,490 by December 2014. Fig. 3 traces the path of
62 R.F. Peláez / Review of Financial Economics 26 (2015) 55–64

Table 7 Table 8
The table lists the trade dates for the nine-month trading rule based on the recession Given an initial investment of $1000, the table shows the terminal wealth accruing to the
probabilities in Fig. 2. buy-and-hold (B&H) and to the switching portfolios.

Switch-out date Switch-back date Date B&H SW9M SW8M SW7M SW6M SW5M

December 31, 1969, Wed. September 30, 1970, Wed. Dec. 31,1969 $1000 $1000 $1000 $1000 $1000 $1000
December 31, 1973, Monday September 30, 1974, Monday Dec. 31, 2014 $88,596 $300,490 $233,311 $206,616 $204,095 $158,067
December 31, 1979, Monday September 30, 1980, Tuesday
June 30, 1981, Tuesday March 31, 1982, Wed.
March 30, 1990, Friday December 31, 1990, Monday
Table 9
December 29, 2000, Friday September 28, 2001, Friday
The table shows the results of estimating Eq. (3) sampling January 1970–December 2014.
December 31, 2007. Monday September 30, 2008, Tuesday
The last column shows the count of months when the switching portfolios held T-bills.

Rule ^
γ T-stat P-val It = 1 (months)
monthly terminal wealth for the buy-and-hold, and for the switching SW5 0.0165 7.3 1.00E-09 35
portfolio based on the dates in Table 7. SW6 0.0199 9.1 1.00E-09 42
Table 8 shows that ignoring expenses and transactions costs, all SW7 0.0173 7.9 1.00E-09 49
switching portfolios dominate the buy-and-hold in terms of terminal SW8 0.0173 7.8 1.00E-09 56
SW9 0.0194 8.6 1.00E-09 63
wealth. Hereto we have ignored expenses and transactions costs.
S&P 500 index funds incur expense ratios of approximately 50 basis
points; however, expense ratios on T-bill index funds are much switching portfolios exhibit positive and highly significant coefficients,
lower. Transactions costs on T-bill round-lots are negligible, and thus decisively rejecting the null.
buyers foreclose those by submitting tenders directly to the Federal Table 10 shows that all switching portfolios dominate the buy-and-
Reserve. To the extent that expense ratios on stock index funds exceed hold based on the Treynor (1965), Sharpe (1966), and Jensen (1968)
transactions costs and expense ratios on T-bills, the results in Table 8 performance measures. Every metric of performance shows the buy-
underestimate the return advantage of the switching portfolios. and-hold to be inferior to the switching portfolios.
Further, interest rates tend to drop during recessions, thus T-bills Assessing investment performance over shorter horizons provides
may generate small capital gains that we ignore to the detriment of additional information. Table 11 shows that the buy-and-hold did not
the switching portfolios. outperform SW9M during any of the five-year spans.
Switching is advantageous if T-bills outperform the S&P 500 during Fig. 4 shows both recursive and 24-month rolling alphas for SW9M.
out-of-market months. To test this we regress the monthly differential Negative rolling alphas occur during several months in the unusual span
return on a binary variable that takes a value of one, It = 1, when the of 1981–82 characterized by back-to-back recessions. As shown in
portfolio holds T-bills, and a value of zero, It = 0, when the portfolio Table 2, one recession began on January 1980 and ended in July 1980.
holds the S&P 500. Letting RSWt and RMt denote monthly returns on The stock market ignored this recession; from March to November
the switching and market portfolios respectively, 1980, all monthly excess returns on the S&P were positive. A manager
who switched out of stocks on December 31, 1979, would have had an
RSW t −RM t ¼ γIt þ εt : ð3Þ opportunity loss.
The trading rules provide a glimpse of possibilities. All rules overlook
^ ; measures the average difference in
The regression coefficient, γ information that a real-time manager could use in timing the return to
monthly returns between T-bills and the S&P during out-of-market stocks, e.g., macroeconomic data, monetary and fiscal policies, exoge-
months. Absent market-timing ability, γ^ ≤0. Table 9 shows that all nous shocks, market and financial developments, corporate profits,

Fig. 3.
 R.F. Peláez / Review of Financial Economics 26 (2015) 55–64 63

Table 10 Table 11
Based on risk-adjusted performance measures for January 1970–December 2014 the The table shows Jensen's alpha for the SW9M portfolio, and Sharpe's reward-to-variability
switching portfolios dominate the buy-and-hold. ratio for the switching and benchmark portfolios.

Measure SW9M SW8M SW7M SW6M SW5M B&H alpha Sharpe (SW9M) Sharpe (B&H)

(RWPt–RFt)1 0.625 0.578 0.556 0.553 0.501 0.399 1970–74 0.62 0.07 −0.14
STD2 4.007 4.094 4.151 4.19 4.248 4.467 1975–79 0.0 0.14 0.14
βeta3 0.812 0.845 0.868 0.883 0.899 1 1980–84 0.27 0.12 0.05
Sharpe4 0.156 0.141 0.134 0.132 0.118 0.089 1985–89 0.0 0.18 0.18
Treynor5 0.77 0.684 0.64 0.627 0.557 0.40 1990–94 0.2 0.13 0.08
Jensen (α)6 0.301 0.241 0.21 0.201 0.148 0 1995–99 0.0 0.41 0.41
alpha P-value7 0.00005 0.00047 0.0012 0.001 0.01 – 2000–04 0.33 0.003 −0.09
Rate of return8 12.75% 12.18% 11.91% 11.88% 11.30% 10.01% 2005–09 0.34 0.04 −0.04
1 2010–14 0 0.32 0.32
Mean monthly portfolio return in excess of the T-bill rate, percentage.
2
Standard deviation of monthly excess returns, percentage.
3
Slope of the portfolio characteristic line.
4
(RWPt–RFt)/STD, or reward-to-variability ratio.
5
(RWPt–RFt)/βeta.
6
Intercept of the portfolio characteristic line. Appendix A. Data sources
7
Probability of the observed α under the α = 0 null hypothesis.
8
Monthly-compounded rate of return (December 1969–December 2014). A.1. Initial claims

price–earnings ratios, earnings announcements, and a host of other The Federal–State Unemployment Insurance Program provides
factors. It is unlikely that on September 30, 2008, witnessing the finan- unemployment benefits to eligible workers as determined under
cial sector in meltdown mode, a manager would switch from T-bills to State law within guidelines established by Federal law. Pursuant to
stocks as in Table 7 and Fig. 3. Avoiding discretion, the trading rules this program, the U.S. Department of Labor, Employment and Train-
hint at the lower bound of achievable returns. ing Administration, obtains weekly initial claims (IC) data from the
State Unemployment Insurance agencies and releases the weekly
6. Conclusions advance IC data on Thursdays. A minor revision the following week
incorporates late reports. There is no need to benchmark, i.e., to re-
Using recession probability forecasts to switch out of stocks is a evaluate the relationship to the population because the IC data are
winning strategy that dominates the buy-and-hold during the 45-year not sample-based – the State agencies report the universe of transac-
span of 1970–2014. The strategy works because the stock market tions. The weekly series exhibits pronounced seasonality. Typically,
is a poor predictor of recessions, and because returns on equities on the last week of March, the Department of Labor (DOL) releases
typically slump after cyclical peaks. A tripling of terminal wealth weekly seasonal factors for the next 12 months and updates the
compared the buy-and-hold illustrates the economic significance weekly seasonal factors for the previous five years. January 7, 1967
of timing the turn in the cycle. The trading rules are divorced is the inception date of the 2014 calendar year weekly series posted
from discretion. A real-life portfolio manager has an informational at www.dol.gov. The weekly series was compacted to a monthly
advantage in deciding when to switch back to stocks. One caveat series by averaging. Prior to January 1967, the data are from the
is that a sudden massive shock could produce an unpredictable CITIBASE (1992) macroeconomic database, which cites the DOL as
recession. primary source.

Recursive and Rolling alphas: January 1973 - December 2014

1.75

Recursive 24-month rolling

1.50

1.25

1.00

0.75

0.50

0.25

0.00

-0.25
1973 1976 1979 1982 1985 1988 1991 1994 1997 2000 2003 2006 2009 2012 2015

Fig. 4.
64 R.F. Peláez / Review of Financial Economics 26 (2015) 55–64

A.2. Non-seasonally adjusted employment and production diffusion indices Greene, W. H. (2008). Econometric analysis (6th ed.). New Jersey: Prentice-Hall.
Greenspan, A. (2000, Dec. 19). Transcript. Federal Open Market Committee Meeting
(pp. 70) (Accessed November 2014. Available at http://www.federalreserve.gov/).
In 1931, the National Association of Purchasing Managers, recent- Greenspan, A. (2012, Aug. 9). Alan Greenspan on his legacy and the economy. Businessweek
ly re-named the Institute for Supply Management (ISM), began to (http://www.businessweek.com/articles/2012-08-09/alan-greenspan-on-his-legacy-
and-the-economy).
survey its members. The respondents indicate whether activity in Greenspan, Alan (2013). Interview, BBC News, October 20, 2013. Accessed October 2013.
the current month is better, unchanged, or worse relative to the previ- Available at http://www.bbc.co.uk/news/business-24602273
ous month. Adding the percentage reporting “better” and one-half the Harding, D., & Pagan, A. (2010). Can we predict recessions? NCER Working Paper #69.
National Centre for Econometric Research.
percentage reporting “unchanged” yields non-seasonally adjusted
Hendry, D. F. (1995). Dynamic econometrics. Oxford: Oxford University Press.
diffusion indexes.15 Hendry, D. F. (2000). Econometrics alchemy or science? Oxford: Oxford University Press.
Hendry, D. F. (2003). How economists forecast. In D. Hendry, & N. Ericsson (Eds.),
Understanding economic forecasts. Cambridge, Mass: The MIT Press.
A.3. Stock prices and the yield spread Hendry, D. F., & Krolzig, H. (2001). Automatic econometric model selection using PcGets 1.0.
London: Timberlake Consultants.
The daily closing levels for the S&P 500 index are from the Security Henriksson, R. (1984). Market timing and mutual fund performance: An empirical
investigation. Journal of Business, 57, 73–96.
Price Index Record (Standard & Poor's, (2001)) and the Federal Reserve Henriksson, R., & Merton, R. (1981). On market timing and investment performance II:
Bank of St. Louis. The 10-year constant maturity Treasury rate, and the Statistical procedures for evaluating forecasting skills. Journal of Business, 54,
secondary market yield on 3-month Treasury bills are from the H15 re- 513–533.
Ippolito, R. (1989). Efficiency with costly information: A study of mutual fund perfor-
lease of the Federal Reserve. Both are available online from the Federal mance, 1965–84. Quarterly Journal of Economics, 104, 1–23.
Reserve Bank of St. Louis. Jensen, M. C. (1968). The performance of mutual funds in the period 1945–1964. Journal
of Finance, 23, 389–416.
Jensen, M. C. (1972). Optimal utilization of market forecasts and the evaluation of investment
A.4. Returns on the S&P 500 and on Treasury bills performance. Mathematical Methods in Finance. North Holland: Amsterdam.
Jensen, M. C. (1978). Some anomalous evidence regarding market efficiency. Journal of
The monthly return series with dividends reinvested is from Financial Economics, 6, 95–101.
Kauppi, H., & Saikkonen, P. (2008). Predicting U.S. recessions with dynamic binary
Standard & Poor's. The monthly yield on secondary market, 3-month response models. The Review of Economics and Statistics, 90, 777–791.
Treasury bills is from the Federal Reserve. Loungani, P. (2000). How accurate are private sector forecasts? Cross-country evidence
from consensus forecasts of output growth. Working Paper #00/77. International
Monetary Fund.
References Malkiel, B. G. (1995). Returns from investing in equity mutual funds 1971 to 1991. Journal
of Finance, 50, 549–572.
Aruoba, S. B. (2008). Data revisions are not well behaved. Journal of Money, Credit, and McNees, S. (1979). The forecasting record of the 1970s. New England Economic Review,
Banking, 40, 319–340. 33–53.
Atkinson, T., Luttrell, D., & Rosenblum, H. (2013). How bad was it? The costs and McNees, S. (1987, Mar/Aprr). Forecasting cyclical turning points: The record in the past
consequences of the 2007–09 financial crisis. Staff Papers #20. Federal Reserve Bank three recessions. New England Economic Review, 31–40.
of Dallas. Merton, R. (1981). On market timing and investment performance: An equilibrium
Barras, L., Scaillet, O., & Russ, W. (2010). False discoveries in mutual fund performance: theory of value for market forecasts. Journal of Business, 54, 363–407.
Measuring luck in estimated alphas. Journal of Finance, 65, 179–216. Moore, G. (1983). Business cycles, inflation, and forecasting. Cambridge, Mass: Ballinger.
BCBC (2008). Determination of the December 2007 peak in economic activity. Accessed Nyblom, J. (1989). Testing for the constancy of parameters over time. Journal of the
on February 2013, Available at: http://www.NBER.org/cycles/dec2008/html American Statistical Association, 84, 223–230.
BCDC (2001). The business-cycle peak of March 2001. Accessed on March 2014, Available Osborne, D., Sensier, M., & Simpson, P. (2003). Forecasting and the UK business cycle. In D.
at: http://www.NBER.org/cycles/November2001/html Hendry, & N. Ericsson (Eds.), Understanding economic forecasts. Cambridge, Mass: The
Berk, J. M. (1998). The information content of the yield curve for monetary policy: A MIT Press.
survey. De Economist, 146, 303–320. Peláez, R. (1998). Economically significant stock market forecasts. The Financial Review,
Braun, P., & Zarnowitz, V. (1992). Twenty-two years of the NBER-ASA quarterly economic 33, 65–76.
outlook surveys: Aspects and comparisons of forecasting performance. NBER Working Peláez, R. (2015). A recession-and-state forecasting model. Southern Economic Journal.
paper #3965. Cambridge, Mass: National Bureau of Economic Research. http://dx.doi.org/10.1002/soej.12060.
Brinson, G., Hood, L., & Beebower, G. (1986). Determinants of portfolio performance. Reeves, A., McKee, M., & Stuckler, D. (2014). Economic suicides in Europe and North
Financial Analysts Journal, 43, 39–44. America's great recessions. The British Journal of Psychiatry, 1–2. http://dx.doi.org/
Carhart, M. (1997). On persistence in mutual fund performance. Journal of Finance, 52, 10.1192/bjp.bp. 114.144766.
57–82. Rudebusch, G., & Williams, J. (2008). Forecasting recessions: The puzzle of the enduring
Chang, E., & Lewellen, W. (1984). Market timing and mutual fund investment perfor- power of the yield curve. Working Paper #2007-16. Federal Reserve Bank of San
mance. Journal of Business, 57, 57–72. Francisco.
Chang, S., Stuckler, D., Yip, P., & Gunnell, D. (2013). Impact of 2008 global economic crisis Samuelson, P. (1966, September 19). Science and stocks. Newsweek, 92.
on suicide: Time trend study in 54 countries. British Medical Journal, 347(f5239), Sharpe, William F. (1966). Mutual fund performance. Journal of Business, 39(1), 119–138.
1–15. Siegel, Jeremy (2014). Stocks for the long run (5th ed.). New York: McGraw Hill.
Chauvet, M., & Potter, S. (2005). Forecasting recessions using the yield curve. Journal of Standard & Poor's (various issues) (2001). Security Price Index Record. New York: N.Y.
Forecasting, 24, 77–103. Stock, J., & Watson, M. (2003). How did leading indicator forecasts perform during the
Chen, Y., & Liang, B. (2007). Do market timing hedge funds time the market? Journal of 2001 recession? Federal Reserve Bank of Richmond. Economic Quarterly, 89, 71–90.
Financial and Quantitative Analysis, 42, 827–856. Stock, J., & Watson, M. (2010). Indicators for dating business cycles: Cross-history selec-
Chin, D., Geweke, J., & Miller, P. (2000). Predicting turning points. Staff Report # 267. tion and comparisons. American Economic Review: Papers and Proceedings, 100, 6–19.
Federal Reserve Bank of Minneapolis. Transcript (2000). Federal Open Market Committee Meeting, December 19, 2000 (pp. 11)
Diebold, F., & Rudebusch, G. (1991). Forecasting output with the composite leading index: (Accessed December 2012. Available at http://www.federalreserve.gov/).
A real-time analysis. Journal of the American Statistical Association, 86, 603–610. Treynor, J. L. (1965). How to rate management of investment funds. Harvard Business
Estrella, A., & Mishkin, F. S. (1998). Predicting U.S. recessions: Financial variables as Review, 43, 63–75.
leading indicators. The Review of Economics and Statistics, 80, 45–61. Treynor, J., & Mazuy, K. (1966). Can mutual funds outguess the market? Harvard Business
Fama, E. F. (1970). Efficient capital markets I: A review of theory and empirical work. Review, 44, 131–136.
Journal of Finance, 25, 383–417. Wallis, K. F. (1989). Macroeconomic forecasting: A survey. Economic Journal, 99, 28–61.
Fama, E. F. (1972). Components of investment performance. Journal of Finance, 27, Wright, J. (2006). The yield curve and predicting recessions. Finance and Economics
551–567. Discussion Series, 2006–07. Federal Reserve Board.
Fama, E. F. (1991). Efficient capital markets II. Journal of Finance, 46, 1575–1617. Zarnovitz, V. (1967). An appraisal of short-term economic forecasts. National Bureau of Eco-
Fels, R., & Hinshaw, E. (1968). Forecasting and recognizing business cycle turning points. nomic Research New York, N.Y.
New York, NY: NBER. Zarnovitz, V. (1992). Business cycles: Theory, history, indicators and forecasting. NBER Stud-
Galvão, A. (2006). Structural break threshold VARs for predicting US recessions using the ies in Business Cycles. Chicago: The University of Chicago Press, 27.
spread. Journal of Applied Econometrics, 21, 463–487.
Goetzmann, W., Ingersoll, J., & Ivkovich, Z. (2000). Monthly measurement of daily timers.
Journal of Financial and Quantitative Analysis, 35, 257–290. Further Reading

15 Peláez, R. (2007). Ex ante forecasts of business-cycle turning points. Empirical Economics,

The observations for the month of August from 1951 through 1957 are missing from
32, 239–246.
the ISM database. Thus, observations for the third quarter of those years were obtained
by averaging the observations for July and September.

View publication stats