Recession fears as self-fulfilling prophecies? Influence on stock returns and output

http://aum.sagepub.com/Australian Journal of Management

http://aum.sagepub.com/content/37/2/231The online version of this article can be found at:

DOI: 10.1177/0312896211423554

2012 37: 231 originally published online 27 April 2012Australian Journal of ManagementJohn G Powell and Sirimon Treepongkaruna

Recession fears as self-fulfilling prophecies? Influence on stock returns and output

Published by:

http://www.sagepublications.com

On behalf of:

Australian School of Business

can be found at:Australian Journal of ManagementAdditional services and information for

http://aum.sagepub.com/cgi/alertsEmail Alerts:

http://aum.sagepub.com/subscriptionsSubscriptions:

http://www.sagepub.com/journalsReprints.navReprints:

http://www.sagepub.com/journalsPermissions.navPermissions:

http://aum.sagepub.com/content/37/2/231.refs.htmlCitations:

What is This?

- Apr 27, 2012OnlineFirst Version of Record

- Aug 2, 2012Version of Record >>

by guest on October 11, 2013aum.sagepub.comDownloaded from by guest on October 11, 2013aum.sagepub.comDownloaded from by guest on October 11, 2013aum.sagepub.comDownloaded from by guest on October 11, 2013aum.sagepub.comDownloaded from by guest on October 11, 2013aum.sagepub.comDownloaded from by guest on October 11, 2013aum.sagepub.comDownloaded from by guest on October 11, 2013aum.sagepub.comDownloaded from by guest on October 11, 2013aum.sagepub.comDownloaded from by guest on October 11, 2013aum.sagepub.comDownloaded from by guest on October 11, 2013aum.sagepub.comDownloaded from by guest on October 11, 2013aum.sagepub.comDownloaded from by guest on October 11, 2013aum.sagepub.comDownloaded from by guest on October 11, 2013aum.sagepub.comDownloaded from by guest on October 11, 2013aum.sagepub.comDownloaded from by guest on October 11, 2013aum.sagepub.comDownloaded from by guest on October 11, 2013aum.sagepub.comDownloaded from by guest on October 11, 2013aum.sagepub.comDownloaded from by guest on October 11, 2013aum.sagepub.comDownloaded from by guest on October 11, 2013aum.sagepub.comDownloaded from by guest on October 11, 2013aum.sagepub.comDownloaded from by guest on October 11, 2013aum.sagepub.comDownloaded from by guest on October 11, 2013aum.sagepub.comDownloaded from by guest on October 11, 2013aum.sagepub.comDownloaded from by guest on October 11, 2013aum.sagepub.comDownloaded from by guest on October 11, 2013aum.sagepub.comDownloaded from by guest on October 11, 2013aum.sagepub.comDownloaded from by guest on October 11, 2013aum.sagepub.comDownloaded from by guest on October 11, 2013aum.sagepub.comDownloaded from by guest on October 11, 2013aum.sagepub.comDownloaded from by guest on October 11, 2013aum.sagepub.comDownloaded from by guest on October 11, 2013aum.sagepub.comDownloaded from

http://aum.sagepub.com/

http://aum.sagepub.com/content/37/2/231

http://www.sagepublications.com

http://www.agsm.edu.au

http://aum.sagepub.com/cgi/alerts

http://aum.sagepub.com/subscriptions

http://www.sagepub.com/journalsReprints.nav

http://www.sagepub.com/journalsPermissions.nav

http://aum.sagepub.com/content/37/2/231.refs.html

http://aum.sagepub.com/content/37/2/231.full.pdf

http://aum.sagepub.com/content/early/2012/04/27/0312896211423554.full.pdf

http://online.sagepub.com/site/sphelp/vorhelp.xhtml
































Australian Journal of Management37(2) 231 –260

© The Author(s) 2011 Reprints and permission: sagepub.

co.uk/journalsPermissions.navDOI: 10.1177/0312896211423554

aum.sagepub.com

Recession fears as self-fulfilling prophecies? Influence on stock returns and output

John G PowellIndependent, Wyevale, Canada

Sirimon TreepongkarunaThe University of Western Australia – Accounting and Finance, Perth, Australia

AbstractMarket participants must rely upon probability assessments of the current state of the economy, that is, their rational ex-ante estimates of recession fears, when making financial and investment decisions. This paper explores whether ex-ante recession fears, modelled using probit analysis of coincident indicators, affect stock returns and output during recessions as well as upturns. Ex-ante recession fears are unrelated to stock returns, an unexpected result that is explained by the lead-lag relationship between recession turning points and subsequent stock market recoveries. Ex-ante recession fears have important dampening effects on output during and, especially, prior to recessions, thus suggesting that recession fears can potentially become self-fulfilling.

Keywordsex-ante recession fears, ex-ante recession probability assessment, ex-ante stock market and output recession fear effects

1. Introduction

The risk that the economy is slipping into a recession is a major and sometimes predominant finan-cial market concern, as highlighted during the recent credit crisis, yet the onset of recessions or economic recoveries is almost never known with certainty until many months afterwards. Market participants must therefore rely upon probability assessments of the current state of the economy, that is, their rational ex-ante estimates of recession fears, when making financial and economic

Corresponding author:Sirimon Treepongkaruna, The University of Western Australia – Accounting and Finance, 35 Stirling Highway, Crawley WA, Perth 6009, Australia. Email: [email protected]

423554 AUM37210.1177/0312896211423554Powell and TreepongkarunaAustralian Journal of Management

Article

232 Australian Journal of Management 37(2)

decisions, since recessions can have a crucial effect upon the success of these decisions. This paper explores three simple but important questions related to ex-ante recession fears. Do ex-ante reces-sion fears affect the severity of stock market and output declines during recessions? Do recession fears also affect aggregate output and stock market returns even when the economy has not yet entered a recession? Can recession fears become self-fulfilling?

Results of the paper indicate that ex-ante recession fears have an important influence on output fluctuations during recessions, thus suggesting that it is the extent of the recession fear (and not just the recession event itself) that helps to determine the severity of recessions, although this effect appears to be diminishing in the most recent time period (post-1960). This more recent moderating of ex-ante recession fear effects on output during downturns perhaps reflects increased (or increasingly successful) central bank and government efforts to counteract recessions when it is perceived that they will be especially severe. Ex-ante recession fears actually have a much stronger effect on aggregate output outside recessions, thus indicating that recession fears dampen economic activity prior to the fears being realized via a recession. This latter result further sug-gests that recession fears can become self-fulfilling, since they can potentially tip the economy into recession through their dampening effect on output.

Surprisingly, however, ex-ante recession fears are not directly related to overall aggregate stock market returns, and specifically are unrelated to stock market returns during recessions. These results are unexpected and initially appear, for instance, to run counter to the experience during the recent financial crisis, when the stock market fell sharply as the crisis deepened. The spring 2009 stock market recovery can be used as an example that helps to explain these unex-pected results, however, since the stock market rose towards the end of the recession in apparent anticipation of the beneficial effects of a subsequent recovery from recession (see also DeStefano, 2004). In the year following the National Bureau of Economic Research’s (NBER’s) 1 December 2008 announcement of the December 2007 peak, for instance, the stock market gained over 30%. Indeed, the paper’s analysis indicates that a strategy of investing in the stock market when the start of a recession is announced until the subsequent announcement of the recession’s end would have been profitable during four of the last five recessions. This latter observation therefore helps to reconcile the paper’s findings with the prevailing view in the literature of the perceived impor-tance of recessions to the stock market (see, e.g., Resnick and Shoesmith, 2002; Siegel, 1998). Also unexpectedly, ex-ante recession fears have recently been associated with higher (not lower) stock market returns prior to recessions, thus possibly indicating that stock price run-ups at the end of upturns can contribute to the onset of recessions.

To model and estimate ex-ante recession fears (probability assessments of the current state of the economy that best utilize all relevant current information), this paper follows the recession forecasting literature and employs probit analysis of economic indicators of NBER recessions. Some leading indicators that have previously been used to forecast recessions are also found to be important coincident indicators of recessions, such as the slope of the yield curve, but other indica-tors, such as sales and employment, are shown to be coincident but not leading indicators of reces-sions, so estimation of ex-ante recession fears is related to but nevertheless distinct from forecasting recessions (see, e.g., Estrella and Mischkin, 1998; Marcellino, 2005; Stock and Watson, 1989). Intuitively, ex-ante recession fears will be influenced by risk factors associated with recessions, such as falling employment, declining output and stock market losses, but with account being taken of the information lead times required for most of these variables. Results indicate that the lagged term spread (the long-term government bond interest rate minus the short-term treasury bill rate), lagged stock returns and a lagged composite index of coincident indicators all play a consistent and important role in the formation of ex-ante recession fear probability assessments.

Powell and Treepongkaruna 233

Ex-ante recession fear probability estimates, obtained from coincident indicators, are employed as a recession fear explanatory variable in a second-stage regression analysis of returns and out-put, thus providing an indication of the role of ex-ante recession fears in financial and output markets (see, e.g., Heckman, 1979; Prabhala, 1997). This approach is employed to examine whether ex-ante recession fears have an effect on returns or output during a particular month, regardless of whether or not the particular month is eventually selected, by the NBER, into a recession sample (since recession fear effects are otherwise only observable if a recession actually occurs). Stock market returns are measured using the monthly real value-weighted Center for Research in Security Prices (CRSP) index, while output is measured using the trailing 12-month real industrial production growth rate, an increasingly utilized output measure that has been shown to be an important priced macroeconomic risk factor in cross-sectional asset pricing tests (see, e.g., Liu and Zhang, 2008). To take account of the potential role of regression variable per-sistence on statistical inference, a recursive bootstrapping technique is utilized that preserves the autocorrelation structure of both the regression model dependent and independent variables (see Efron and Tibshirani, 1986; see also Li and Maddala, 1997). In addition, the sensitivity of the analysis to alternative autoregressive regression model specifications is also investigated.

Initial univariate regression analysis indicates that ex-ante recession fears have an important effect on output but do not influence overall stock returns. Interestingly, ex-ante recession fears explain far more of the monthly variation in industrial production, on their own, than does the NBER classification of recessions that is only observable ex-post, many months afterwards. This latter result is a key finding of the paper, since it indicates that practitioners and academicians alike can focus on the strength of ex-ante recession fears when assessing potential impacts on output and the wider economy, rather than narrowly focusing on whether the economy is technically in a recession. A parallel implication is that indicators such as the term spread and the Conference Board Coincident Indicator are important in their own right, because they help to indicate the strength of current recession fears that in turn have a direct and considerable influence on output, especially prior to recessions.

To specifically explore when and how ex-ante recession fears have their greatest influence on stock returns and output, a multivariate regression analysis utilizing an interactive recession dummy explanatory variable is conducted, thus creating separate ex-ante recession fear explana-tory variables for recession, pre-recession and non-recession time periods (see, e.g., He and Ng, 1998). This is especially important for understanding the relationship between ex-ante recession fears and stock returns, since it is possible that ex-ante recession fears might only be important to the stock market as the economy enters a recession (see Boyd et al., 2005; Henkel et al., 2007; Nieto and Rubio, 2008). Boyd et al. (2005), for instance, examine whether investors react differ-ently to unemployment news surprises when they perceive that the economy is in a recession, versus whether they perceive an upturn is occurring, and find that an increase in unemployment is only bad news for the stock market during recessions. This paper’s results indicate that ex-ante recession fears reduce economic activity no matter what the current state of the economy, with recession fears actually having a much more important effect on output during non-recessions, and especially so just prior to recessions, thus suggesting that recession fears can potentially become self-fulfilling. The results imply that ex-ante recession fears do not explain stock returns during recessions, however, and instead indicate that the inverse empirical relationship between aggregate stock returns and recessions appears to be a phenomenon that is only directly identifiable ex-post. Further analysis reveals that stock indices fall during recessions but recover sharply afterwards, so when recession probability assessments peak towards the end of a recession then the stock market


is already looking forward to strong subsequent gains, thus obscuring any overall relationship between ex-ante recession fears and stock returns (see also DeStefano, 2004).

The study’s regression analysis procedure is motivated and outlined in the following section, along with the data sample used to examine the influence of ex-ante recession fears on stock returns and output. A second section contains the paper’s results, including an application of the paper’s approach to Australian markets, and a final section provides conclusions.

2. Model selection and data sample

When there is a risk that the economy is slipping or has slipped into a recession, considerable atten-tion is paid to business cycle indicators, such as employment or interest rate spreads, for clues as to the likelihood of a recession, and financial markets appear to respond to every macroeconomic news release or pronouncement. This section outlines how ex-ante recession probability assess-ments of the current state of the economy can be estimated from coincident business cycle indica-tors in order to examine whether ex-ante recession fears help to explain stock returns and output fluctuations, and also introduces the data used in this paper’s study.1

2.1 Ex-ante recession probability assessment model

As noted already, recessionary indicators, such as unemployment, falling output or declining profits, provide important clues as to the current state of the economy, albeit with an information release lag due to the lead times required for these coincident business cycle indicators. Shorter information release time lags therefore create the possibility that some coincident recession indicators are not leading indicators of recessions, as demonstrated in Section 3, so the process of estimating coincident ex-ante recession probability assessments is distinct from (but also obviously related to) the process of forecasting recessions. As is common in the literature, ex-ante recession fears are estimated in this study using probit regression analysis.2 A monthly data set is used when modelling ex-ante recession probability assessments, since recession turning points classified by the NBER often occur mid-quarter.

Perhaps the best-known recession indicator, due to the consistency with which it has fore-shadowed post-World War II recessions, is the term spread, defined as the spread between the long-term 10-year government bond yield minus the short-term 90-day treasury bill rate (S). The term spread has tended to spike downwards and turn negative (invert) at some point during the 12 or so months leading up to most recent recessions, whereas an upwards sloping term spread is present at most other times (Estrella and Hardouvelis, 1991; Harvey, 1989). Higher short-term rates increase the cost of funds for investing, as well as for consumer borrowing, thus providing an intuitive theoretical underpinning as to why an inverted term spread can foreshadow reces-sions if short rates rise relative to long rates. Interestingly, Rudebusch and Williams (2007) indicate that professional forecasters are unable to identify (predict) the current (future) state of the economy, but would actually be able to do so if they made better use of the yield curve when making forecasts. Theoretical and empirical considerations therefore suggest that the term spread, lagged by 12 months (by convention), should be included as an explanatory variable when modelling ex-ante recession fears.

Even though a fall in the term spread has preceded most recent recessions, the term spread failed to reliably indicate or foreshadow the 1991 recession, thus motivating researchers to identify other reliable recession indicators, such as the Conference Board Composite Index (see, e.g. Birchenhall et al., 1999; Estrella and Mishkin, 1998; Marcellino, 2005). Notably, Stock and Watson (1989)


construct an experimental coincidental recession probability indicator based on an index of a number of important business cycle series. Birchenhall et al. (1999) find that four important indica-tors, the lagged 12-month rates of change of industrial production, annualized real sales, annual-ized real personal income and non-farm payrolls, can be used to explain recession occurrences during the sample period from January 1953 to January 1995. These four indicators are used by the Conference Board to construct a Composite Coincident Indicator, and Hall (1991) notes the impor-tance of the four series in recent NBER dating decisions. The four series, appropriately lagged to reflect their information release delay, are thus explored, but due to multi-collinearity amongst the series a coincident indicator (CI) is constructed from the four series, as described in the following section, and utilized in the ex-ante recession probability assessment model of the current state of the economy.

An aggregate business cycle series that potentially overshadows the component coincident indicator series in terms of its overall role in business cycle dating is gross domestic product (GDP; see, e.g. Bry and Boschan, 1971). This suggests that appropriately lagged real GDP growth could have an important explanatory role in modelling ex-ante recession fears.3

In addition, Harvey (1989) notes that the stock market is traditionally assumed to contain impor-tant information about current and future economic growth because the economy determines the magnitude of cash flows that will accrue to stock market investors. Lagged real monthly stock returns (r) are therefore utilized in the ex-ante recession fear model. Lagged real monthly stock returns r are represented by the CRSP value-weighted total return index from 1926 onwards and the Schwert (1990) index for earlier periods, with each index being adjusted for inflation using the Consumer Price Index.

More recently, Öncü (2006) points out that the longer the economy spends in a recession then the more likely it is to switch to an upturn, whereas when the economy is initially in an upturn then it is less likely to switch to a recession. The duration variables R and U are therefore included in the recession probability assessment model of the current state of the economy, where R and U are defined as the duration in months since the last recession or upturn business cycle turning point, lagged by six months (the minimum number of months that is generally required to confirm that a prior turning point has occurred).

Another recent contribution to the literature is the use of the Livingstone expected business condition survey of GDP growth rate forecasts. Campbell and Diebold (2005), for instance, dem-onstrate that the Livingstone GDP growth rate forecast can be used to predict semi-annual stock returns (see also Goetzmann et al., 2007). The log rate of change of 12-month versus 6-month ahead Livingstone survey real GDP forecasts, lagged by 6 months (L), is therefore included as a final ex-ante recession probability assessment variable.4

The full ex-ante recession probability assessment probit model is

P BC y N S CI GDP r R Ut t t t t t t( | ) (= = + + + + + +− − − − −1 1 12 2 2 3 4 4 1 5 6 6α β β β β β β tt tL− −+6 7 6β ), (1)

where P denotes probability, BCt is a binary dependent variable equal to 1 if an NBER recession occurs in month t and 0 otherwise, N is the cumulative standard normal distribution function, a is a constant and the lagged business cycle explanatory variables ( yt ) are defined as above, with the subscript t–n indicating the explanatory variable is lagged by n months. Except for the term spread (S), where the convention is to use the 12-month lagged-term spread, all explanatory variables are employed as coincident indicators and are therefore lagged by the minimum amount required to reflect information release delays (see, e.g., Estrella and Mishkin, 1998).5


The probit ex-ante recession probability assessment model is first estimated using the most recent post-World War II sample period, a sample that has been extensively examined in the reces-sion forecast literature, and then tested for consistency using an earlier subsample as well as an extended 85-year sample that combines the two subsamples. The data set is constructed from mainly NBER sources, as described in the Section 2.3.

2.2 Ex-ante recession fear effects on stock returns and output

Following Heckman (1979), ex-ante recession probability assessments representing economic agents’ estimates of the unobserved current state of the economy, that is, their recession fears, are transformed into an Inverse Mill’s ratio explanatory variable to examine the relationship between ex-ante recession fears and output or return fluctuations (see Prabhala, 1997). A Heckman (1979) two-step ‘sample selection’ approach is utilized because, as mentioned already, economic agents must adjust their economic behaviour in reaction to their ex-ante recession fears, long before the effects of a recession can be observed ex-post once a time period has been selected by the NBER as being a part of the recession sample.6 The Inverse Mill’s ratio at month t, IMt, is defined as

IMy

N ytt

t

=++

φ α βα β( )

( ), (2)

where f represents the density function for a standard normal, a is a constant, the vector yt represents the leading indicator business cycle explanatory variables (see Equation (1)) and N is the cumulative standard normal distribution function. The Inverse Mill’s ratio is a decreasing func-tion of the coincident recession probability assessment N yt( ),α β+ since it equals zero when the current recession likelihood is a certainty and it equals infinity when there is a zero chance of a current recession.

Univariate regression analysis is employed to test whether ex-ante recession fears, represented by the monthly Inverse Mill’s ratio variable IMt, are related to real monthly stock returns rt and output IPt ; output is measured using the real industrial production growth rate (see, e.g., Boulier and Stekler, 2000; Liu and Zhang, 2008). The regression models used to test for overall ex-ante recession fear effects on real stock returns rt and output IPt are therefore

r IMt t t= + +β β ε0 1 (3)

and

IP IMt t t= + +β β ε0 1 , (4)

where et is the regression error term. A finding that the estimated Inverse Mill’s ratio coefficient b1 is significantly positive implies that increased ex-ante recession fears are associated with lower overall real stock returns or output growth (recall from Equation (2) that the Inverse Mill’s ratio is a decreasing function of the ex-ante recession probability assessment of the current state of the economy, thus explaining the Inverse Mill’s ratio coefficient b1 coefficient estimate interpretation).

Due to the potential influence of regression variable persistence on statistical inference (see Section 2.3), recursive bootstrapping that preserves the persistence properties of both the


dependent variable and independent variable series is utilized to take account of independent, as well as dependent, variable persistence (see Efron and Tibshirani, 1986; see also Li and Maddala, 1997).7 Details of the recursive bootstrapping procedure are outlined in the Appendix. The recursive bootstrapping procedure is used to estimate bootstrapped t-statistic cut-offs to determine if the Inverse Mill’s ratio explanatory variable IMt coefficient estimates for regression Equations (3) and (4) are statistically significant, even after controlling for independent and dependent vari-able persistence (see also Ferson et al., 2003; Granger et al., 2001).

As an additional sensitivity check, an AR2 regression model is also employed to estimate regression model (4), since AR2 regression models have been used in the prior literature to model monthly industrial production growth rates (see, e.g. Bernanke, 1983), with the AR2 regression specification being

IP IM IP IPt t t t t= + + + +− −β β β β ε0 1 2 1 3 2 . (5)

A further sensitivity check is to examine the influence of ex-ante recession fears on a GDP measure of output, since real quarterly GDP growth rates (GDPQ) are not persistent. The following modi-fication of regression model (4) is therefore used:

GDPQ IMt t t= + +β β ε0 1 , (6)

where regression model (6) is estimated using quarterly (not monthly) data.8

To explore whether ex-ante recession fears have their greatest impact when the economy moves into a recession, the univariate regression analysis is extended using separate ex-ante recession fear explanatory variables for recession, non-recession and pre-recession time periods (see, e.g. He and Ng, 1998). An Inverse Mill’s ratio IMt interaction variable is used to capture the influence of ex-ante recession fears on returns and output during recessions, as well as during upturns or just prior to recessions (see Boyd et al., 2005; Henkel et al., 2007; Nieto and Rubio, 2008).9 These effects are estimated using an interaction variable that employs the ex-post recession dummy variable BCt as follows:

r IM BC IMt t t t t= + + +β β β ε0 1 2 * (7)and

IP IM BC IMt t t t t= + + +β β β ε0 1 2 * , (8)

where BCt is a dummy variable equal to 1 if month t is identified as a recession and is 0 other-wise.10 A significantly positive Inverse Mill’s ratio explanatory variable coefficient b1 indicates that higher ex-ante recession probability assessments are associated with lower real returns or output growth rates during upturns. A significantly positive (negative) Inverse Mill’s ratio inter-action variable coefficient b2 indicates that higher ex-ante recession probability assessments are associated with return or output growth rate effects during recessions that are even stronger (weaker) than during upturns. [The estimated ex-ante recession fear effect during downturns is therefore equal to the sum of the two explanatory variable coefficient estimates, b1 + b2 .] The statistical significance of regression explanatory variable coefficient estimates is once again assessed using the paper’s recursive bootstrapping procedure (see the Appendix). Additional sen-sitivity checks, identical to the additional univariate regression sensitivity checks (see regression


Equations (5) and (6)), are also employed for output regression model (8), as reported below once the paper’s data sample is outlined.

2.3 Data sample and summary statistics

The paper utilizes an extensive 85-year sample that is constructed primarily from NBER and Federal Reserve Board sources, as noted below. The use of a long sample period is important, since the post-1955 Federal Reserve Board macroeconomic data sample has been extensively examined for statistical relationships, so it is helpful to also test, for consistency, using a subsam-ple period that has not previously been examined (see, e.g. Estrella and Mishkin, 1998). The study’s full sample period is 1921–2005, and is divided into two subsamples (pre- and post-1959), with 1921 being chosen as the start point for the full sample because it is the first date for which many of the NBER and Federal Reserve Board business cycle indicator variables are generally available on a monthly basis.

The four Conference Board Composite Coincident Indicator variables have long-term data series that are available in various forms (and with various base years) in the NBER Macrohistory Database and can be carefully spliced together. Monthly industrial production can be constructed back to January 1889 from these sources, and monthly personal income is available from January 1921 onwards, whereas monthly non-farm payroll data are not available until January 1929 and monthly sales are available from January 1939 onwards. Quarterly GDP is available from the NBER and Federal Reserve for the first quarter of 1921 onwards. NBER recessions (BC), the primary focus of interest for this study, are available from December 1854 onwards. The term spread (S), constructed by subtracting three month treasury bill rates from 10-year government bond yields, is available monthly from Federal Reserve sources starting in January 1920.11 The CRSP value-weighted total rate of return, adjusted for inflation (r), is used to measure stock returns from 1926 onwards and, as is common in the literature, Schwert (1990) index total returns, also adjusted for inflation, are used for earlier time periods.12

Summary statistics for all the business cycle variable series are provided in Table 1. The first column of Table 1 Panel A indicates that the economy spends 31.2% of months in recessions between January 1857 and December 2005, as revealed by the NBER recession dummy variable business condition indicator (BC) mean. Panel A of Table 1 also reveals that the real rate of growth of quarterly GDP (GDPQ), when annualized, exceeds 3% per annum, while the annual industrial production (IP), manufacturing and trade sales (Sale), and personal income (PI) real growth rates are all close to 3.5% per annum. Quarterly GDP growth (GDPQ) is defined as the log change in real quarterly GDP. Industrial production (IP), real manufacturing and trade sales (Sale), real per-sonal income (PI) and the non-farm payroll (NFPR) growth rate are all defined (by convention) as the 12-month log change in the values of the series. All the business cycle variable series examined are highly auto-correlated (see Panel B of Table 1), other than real stock returns (r) and the real quarterly GDP growth rate (GDPQ).

The business cycle variable correlations reported in Panel C of Table 1 indicate strong multi-collinearity amongst the four Conference Board Composite Coincident Indicator variables (IP, Sales, NFPR and PI), thus justifying a decision to combine them into a single composite index (CI) explanatory variable for the probit regression analysis (see Equation (1)). The monthly composite index (CI) is set equal to IP (the industrial production growth rate) for the time period 1889–1920, half of IP and PI (the real personal income growth rate) when personal income data becomes avail-able in January 1921, one third of IP, PI and NFPR (the non-farm payroll growth rate) from January 1929 to 1938 and one quarter of IP, PI, NFPR and Sales (the real Sales growth rate) from

Powell and Treepongkaruna 239T

able

1.

Des

crip

tive

stat

istic

s. P

anel

A p

rese

nts

sum

mar

y st

atis

tics

for

the

NBE

R r

eces

sion

indi

cato

r (B

C), t

he d

iffer

ence

bet

wee

n th

e 10

-yea

r T

-bon

d an

d 3-

mon

th T

-bill

rat

es (

S), t

he r

eal r

ate

of c

hang

e in

the

CR

SP v

alue

-wei

ghte

d to

tal r

etur

n in

dex

(r),

the

3-m

onth

log

chan

ge in

rea

l gro

ss d

omes

tic p

rodu

ct

(GD

PQ),

the

12-m

onth

log

chan

ge in

the

ann

ualiz

ed in

dust

rial

pro

duct

ion

inde

x (IP

), th

e 12

-mon

th lo

g ch

ange

in a

nnua

lized

rea

l sal

es (

Sale

), th

e 12

-mon

th

log

chan

ge in

non

-farm

pay

roll

(NFP

R), t

he 1

2-m

onth

log

chan

ge in

rea

l ann

ualiz

ed p

erso

nal i

ncom

e (P

I), t

he c

ompo

site

inde

x (C

I), w

hich

is d

efin

ed a

s an

eq

ually

wei

ghte

d in

dex

of IP

, Sal

e, N

FPR

and

PI, t

he d

urat

ion

in m

onth

s si

nce

the

last

rec

essi

on o

r up

turn

, lag

ged

by 6

mon

ths

(R a

nd U

) an

d th

e lo

g ra

te

of c

hang

e of

12-

mon

th v

ersu

s 6-

mon

th a

head

rea

l GD

P fo

reca

sts

from

the

Liv

ings

tone

sur

vey,

lagg

ed b

y 6

mon

ths

(L).

Pane

l B r

epor

ts a

utoc

orre

latio

ns

of t

he v

aria

bles

up

to s

ix la

gs, a

nd c

orre

latio

ns a

mon

g th

e va

riab

les

are

pres

ente

d in

Pan

el C

. The

sam

ple

peri

od t

hat

is c

omm

on t

o al

l the

var

iabl

es is

Ja

nuar

y 19

40–D

ecem

ber

2005

(763

obs

erva

tions

). A

ll da

ta a

re c

olle

cted

on

a m

onth

ly b

asis

, exc

ept f

or G

DPQ

, whi

ch is

col

lect

ed o

n a

quar

terl

y ba

sis

for

the

sum

mar

y st

atis

tics.

The

pai

rwis

e co

rrel

atio

n be

twee

n G

DPQ

and

the

rec

essi

on in

dica

tor

BC is

bas

ed u

pon

a qu

arte

rly

rece

ssio

n in

dica

tor

BCQ

, whi

ch e

qual

s on

e du

ring

the

qua

rter

if a

t le

ast

two

of t

hree

mon

ths

have

rec

essi

on in

dica

tors

BC

= 1

dur

ing

the

quar

ter,

and

is z

ero

othe

rwis

e.

BC

Sr

GD

PQIP

Sale

NFP

RPI

CIR

UL

Pane

l A: d

escr

iptiv

e st

atist

icsM

ean

0.31

20.

558

0.00

70.

008

0.03

60.

036

0.01

80.

035

0.02

94.

191

19.5

680.

013

Med

ian

0.00

00.

665

0.00

80.

008

0.04

30.

038

0.02

30.

037

0.03

40.

000

11.0

000.

013

Max

imum

1.00

04.

390

0.38

30.

176

0.48

30.

291

0.15

90.

237

0.26

265

.00

120.

000.

030

Min

imum

0.00

0−1

7.32

−0.2

85−0

.139

−0.4

10−0

.137

−0.1

64−0

.216

−0.3

810.

000

0.00

0−0

.018

Std.

Dev

.0.

463

1.66

10.

050

0.03

10.

113

0.05

70.

042

0.05

40.

074

9.36

725

.034

0.00

7O

bs.

1788

1788

1788

340

1392

763

912

980

1020

1788

1788

649

Pane

l B: a

utoc

orre

latio

n1

0.91

60.

913

0.08

70.

000

0.95

30.

910

0.98

40.

939

0.95

90.

952

0.96

90.

955

20.

833

0.80

5−0

.010

0.00

00.

873

0.83

70.

950

0.89

30.

895

0.90

40.

938

0.90

93

0.74

90.

732

−0.0

64−0

.113

0.77

90.

762

0.90

30.

858

0.81

80.

857

0.90

80.

864

40.

666

0.68

10.

014

0.00

00.

677

0.66

30.

847

0.80

80.

736

0.81

00.

877

0.81

85

0.58

20.

656

0.06

80.

000

0.57

70.

579

0.78

50.

763

0.65

40.

765

0.84

60.

772

60.

498

0.63

6−0

.021

0.20

60.

478

0.47

80.

719

0.71

30.

567

0.72

00.

816

0.72

7Pa

nel C

: cor

rela

tions

BC1.

000

S

−0.0

161.

000

r

−0.0

350.

084

1.00

0

GD

PQ−0

.485

0.03

70.

035

1.00

0

IP−0

.488

−0.1

50−0

.106

0.49

51.

000

Sa

le−0

.539

−0.0

44−0

.091

0.54

80.

876

1.00

0

NFP

R−0

.383

−0.2

98−0

.118

0.31

00.

828

0.73

41.

000

PI

−0.4

64−0

.240

−0.0

620.

468

0.84

70.

771

0.82

51.

000

CI

−0.5

18−0

.164

−0.1

020.

514

0.97

20.

936

0.87

90.

905

1.00

0

R0.

113

0.17

50.

104

−0.0

11−0

.542

−0.4

59−0

.599

−0.5

11−0

.554

1.00

0

U−0

.028

−0.3

85−0

.098

−0.1

150.

132

0.09

30.

343

0.23

70.

179

−0.3

841.

000

L

−0.1

240.

273

−0.0

530.

209

0.01

20.

058

−0.0

61−0

.039

0.00

70.

198

−0.2

771.

000


January 1939 onwards. Summary statistics for the composite index (CI) series are provided in the ninth column of Panels A and B of Table 1, and reveal that the composite index series properties closely match those of its component parts. Panel C of Table 1 also demonstrates that the recession dummy variable BC is strongly and negatively contemporaneously correlated with all the business cycle indicators, except spread S and return r (as might be expected) as well as the Livingstone forecast GDP growth rate L (an unexpected result). This latter result is unexpected because forecast GDP growth L would be expected to be lower during prolonged recessions, and higher during upturns, but it is consistent with Rudebusch and Williams’ (2007) finding that professional fore-casters are unable to forecast GDP growth because they do not make effective use of the predictive power of the term spread.

To further explore the relationship between recessions and the contemporaneous value of the business cycle indicators, Panels A and B of Table 2 report descriptive statistics separately for recession and non-recession subsamples.13 The term spread and real stock returns are much lower during recessions, and the rest of the indicators (other than the Livingstone growth forecast L) flip from being strongly negative during recessions to strongly positive during upturn months (compare the Panel A and B means in Table 2). The almost 1% per month difference between stock returns in upturns versus recessions is as expected (see, e.g. Resnick and Shoesmith, 2002; Siegel, 1998). The extremely small proportion of real return variability explained by ex-post recession classifications foreshadows, however, the regression results reported in Section 3, where real return differences explained by ex-ante recession fears are found to be insignificant (the real return col-umn adjusted R2 reported in Panel C of Table 2 is only 0.006). Panel C of Table 2, which outlines the contemporaneous differences in business cycle indicators between recessions and upturns, indicates that only the Livingstone GDP growth forecast L does not differ sharply between reces-sions and upturns.

3. Results

3.1 Ex-ante recession probability assessment model results

Results for the ex-ante recession probability assessment model of the current state of the econ-omy are provided in Table 3 for a January 1960–December 2005 sample period (see Panel A) as well as an earlier July 1921–1959 sample, plus a full sample that combines the two subsamples (see Panels B and C). The Table 3 results indicate that the lagged-term spread (St–12) and the lagged composite index (CIt–2) appear to play a consistently important role in the formation of ex-ante recession probability assessments (i.e. ex-ante recession fears), as do lagged stock returns (rt–1). Lagged GDP growth rates (GDPt–4) have insignificant explanatory power in the multivari-ate probit regression model, perhaps because recession turning points can occur mid-quarter, whereas GDP is measured quarterly, but they do have explanatory power in univariate probit regression analysis (results not reported).14 Table 3 indicates that the composite index (CIt–2) is very important for explaining coincident ex-ante recession fears, but (in unreported results) is found to have no predictive power for forecasting recessions, thus indicating that estimation of coincident ex-ante recession fears is distinct from the process of forecasting recessions.15

To further assess the fit of probit ex-ante recession probability assessment model (1), two good-ness-of-fit measures are reported in the final two columns of Table 3 (see also the McFadden adjusted R2). The second to last column provides a concordance goodness-of-fit measure. The concordance measure indicates that estimated probabilities provided by the ex-ante recession prob-ability assessment model (1), when rounded to either 1 or 0, are highly concordant with ex-post


Tab

le 2

. R

eces

sion

and

non

-rec

essi

on d

escr

iptiv

e st

atis

tics.

Pan

els

A a

nd B

pre

sent

sum

mar

y st

atis

tics

for

sepa

rate

rec

essi

on a

nd n

on-r

eces

sion

sam

ples

fo

r th

e N

BER

rec

essi

on in

dica

tor

(BC)

, the

diff

eren

ce b

etw

een

the

10-y

ear

T-b

ond

and

3-m

onth

T-b

ill r

ates

(S)

, the

rea

l rat

e of

cha

nge

in t

he C

RSP

va

lue-

wei

ghte

d to

tal r

etur

n in

dex

(r),

the

3-m

onth

log

chan

ge in

rea

l gro

ss d

omes

tic p

rodu

ct (

GD

PQ),

the

12-m

onth

log

chan

ge in

the

ann

ualiz

ed in

dust

rial

pr

oduc

tion

inde

x (IP

), th

e 12

-mon

th lo

g ch

ange

in a

nnua

lized

rea

l sal

es (

Sale

), th

e 12

-mon

th lo

g ch

ange

in n

on-fa

rm p

ayro

ll (N

FPR)

, the

12-

mon

th lo

g ch

ange

in r

eal a

nnua

lized

per

sona

l inc

ome

(PI),

the

com

posi

te in

dex

(CI)

whi

ch is

def

ined

as

an e

qual

ly w

eigh

ted

inde

x of

IP, S

ale,

NFP

R an

d PI

, the

dur

atio

n in

mon

ths

sinc

e th

e la

st r

eces

sion

or

uptu

rn, l

agge

d by

six

mon

ths

(R a

nd U

) an

d th

e lo

g ra

te o

f cha

nge

of 1

2-m

onth

ver

sus

6-m

onth

ahe

ad r

eal G

DP

fore

cast

s fr

om t

he L

ivin

gsto

ne s

urve

y, la

gged

by

6 m

onth

s (L

). Pa

nel C

rep

orts

the

BC

indi

cato

r re

gres

sion

coe

ffici

ents

, t-s

tatis

tics

(est

imat

ed u

sing

the

N

ewey

–Wes

t m

etho

d), a

djus

ted

R2 a

nd n

umbe

r of

obs

erva

tions

from

a r

egre

ssio

n of

the

var

iabl

e in

eac

h co

lum

n on

a c

onst

ant

and

the

BC in

dica

tor

(or

the

quar

terl

y BC

Q in

dica

tor

for

GD

PQ, w

here

BCQ

equ

als

one

duri

ng t

he q

uart

er if

at

leas

t tw

o of

thr

ee m

onth

s ha

ve r

eces

sion

indi

cato

rs B

C =

1 d

urin

g th

e qu

arte

r, a

nd is

zer

o ot

herw

ise)

. The

sam

ple

peri

od t

hat

is c

omm

on t

o al

l the

var

iabl

es is

Janu

ary

1940

–Dec

embe

r 20

05 (

763

obse

rvat

ions

). T

he

desc

ript

ive

stat

istic

s fo

r G

DPQ

are

bas

ed o

n qu

arte

rly

data

.

BC

Sr

GD

PQIP

Sale

NFP

RPI

CIR

UL

Pane

l A: d

escr

iptiv

e st

atist

ics w

hen

BC =

1M

ean

1.00

00.

005

0.00

1−0

.013

−0.0

54−0

.027

−0.0

28−0

.015

−0.0

478.

280

11.8

00.

010

Med

ian

1.00

00.

192

0.00

4−0

.008

−0.0

43−0

.026

−0.0

110.

001

−0.0

234.

000

0.00

00.

011

Max

imum

1.00

04.

390

0.37

50.

112

0.27

40.

074

0.06

90.

147

0.17

659

.00

120.

00.

021

Min

imum

1.00

0−1

7.32

−0.2

85−0

.125

−0.3

93−0

.137

−0.1

64−0

.216

−0.3

810.

000

0.00

0−0

.018

Std.

Dev

.0.

000

1.87

80.

061

0.03

80.

120

0.03

90.

051

0.05

90.

088

11.8

6222

.83

0.00

9O

bser

vatio

ns55

855

855

867

366

112

164

195

202

558

558

93Pa

nel B

: des

crip

tive

stat

istics

whe

n BC

= 0

Mea

n0.

000

0.80

80.

009

0.01

30.

068

0.04

70.

028

0.04

70.

048

2.33

723

.09

0.01

3M

edia

n0.

000

0.85

00.

011

0.01

00.

058

0.04

60.

026

0.04

30.

042

0.00

014

.50

0.01

3M

axim

um0.

000

4.16

00.

383

0.17

60.

483

0.29

10.

159

0.23

70.

262

65.0

011

4.0

0.03

0M

inim

um0.

000

−10.

03−0

.227

−0.1

39−0

.410

−0.1

12−0

.113

−0.1

19−0

.246

0.00

00.

000

−0.0

18St

d. D

ev.

0.00

01.

488

0.04

40.

027

0.09

10.

052

0.03

20.

046

0.05

57.

264

25.2

10.

007

Obs

erva

tions

1230

1230

1230

273

1026

651

748

785

818

1230

1230

556

Pane

l C: u

niva

riate

reg

ress

ions

Coe

ffici

ent

−0.8

03−0

.008

−0.0

26−0

.121

−0.0

75−0

.056

−0.0

62−0

.095

5.94

3−1

1.29

−0.0

03t-

stat

istic

−4.2

11−3

.039

−6.0

28−8

.215

−10.

28−5

.700

−5.9

59−6

.80

4.49

−4.2

7−1

.26

R20.

050

0.00

60.

110

0.22

30.

212

0.25

90.

206

0.26

40.

086

0.04

30.

016

Obs

erva

tions

1788

1788

340

1392

763

912

980

1020

1788

1788

649


recession classifications BCt , with concordance reaching almost 90% in the post-1960 subsample (i.e. the number of months where the classifications are the same is almost 90%).16 The final column of Table 3 outlines quadratic probability scores (QPSs), the probability analogue of mean-squared error, for ex-ante recession probability assessment model (1) (see, e.g. Anderson et al., 2007):

Table 3. Ex-ante estimate of the coincident probability of a recession using probit regression. This table reports results for various samples for multivariate probit regression model (1), specified as P BC y N S CIt t tt( | ) (= = + + +− −1 1 2 212α β β b b b b b3 4 4 1 5 6 6 6 7 6GDP r R U Lt t t t t− − − − −+ + + + ), where P denotes probability, BCt is a binary dependent variable equal to 1 if a recession is occurring in month t and 0 otherwise, N is the cumulative standard normal distribution function, a is a constant and the lagged business cycle explanatory variables are defined as: St−12 is the 10-year government bond yield minus the 90-day treasury bill rate at the end of month t−12, CIt−2 is the month t−2 Coincident Indicator, GDPt−4 is the quarterly log change in GDP (recorded monthly for each month in the quarter, and lagged by four months), rt−1 is the log change in the monthly real total return CRSP value-weighted stock index (lagged by one month), Rt−6 and Ut−6 are the duration in months since the last recession or upturn (lagged by six months), and Lt−6 is the log rate of change of 12-month versus 6-month ahead real GDP forecasts from the Livingstone survey (lagged by 6 months). The z-statistics are reported in brackets. McFadden R2 is also reported. (-l) in the first row indicates the independent variable is lagged by l times. Panels A−C and D−F are for sample periods 01/1960−12/2005, 07/1921−12/1959 and 07/1921−12/2005, respectively, whereas Panels G and H are for sample periods 01/1857−07/1921 and 01/1857−01/2005, respectively. Panels D−H are for a probit regression with one explanatory variable only (St−12). Concordance (Concord) is the proportion of months classified correctly by model (1) and QPS is a measure of fit defined by Equation (9).

Constant S(−12) r(−1) GDP(−1) CI(−2) R(−6) U(−6) L(−6) R2 Obs Concord QPS

Panel A: Jan 1960 to Dec 20050.803 −0.604 −4.615 −10.604 −26.269 −0.187 −0.003 −37.330 0.415464 552 0.89 0.15

(2.21) (−5.70) (−2.45) (−0.94) (−5.72) (−3.63) (−1.02) (−2.00) Panel B: Jul 1921 to Dec 1959−0.502 −0.432 −3.632 0.380 −6.400 0.031 0.022 0.251365 462 0.82 0.27

(−3.71) (−4.87) (−3.30) (0.22) (−5.51) (2.96) (5.24) Panel C: Jul 1921 to Dec 2005−0.292 −0.476 −3.458 0.413 −6.679 0.013 0.002 0.243854 1014 0.84 0.23

(−2.83) (−9.18) (−3.49) (0.22) (−5.16) (1.19) (0.88) Panel D: Jan 1960 to Dec 2005−0.382 −0.813 0.291137 552 0.87 0.18

(−4.04) (−9.01) Panel E: Jul 1921 to Dec 1959−0.221 −0.344 0.049413 462 0.76 0.37

(−2.15) (−4.99) Panel F: Jul 1921 to Dec 2005−0.288 −0.508 0.129541 1014 0.82 0.27

(−4.21) (−10.5) Panel G: Jan 1857 to Jul 1921−0.166 −0.158 0.025229 763 0.58 0.48

(−3.40) (−4.90) Panel H: Jan 1857 to Dec 2005−0.364 −0.337 0.116496 1776 0.71 0.37

(−10.85) (−14.9)


QPST

P BCtt

T

t= −=∑1 21

2( ) , (9)

where T is the number of observations and Pt is the ex-ante recession probability estimate for month t. The QPSs indicate a very good model fit, with Panel A of Table 3 indicating that the model fit in the most recent (post-1960) subsample is better than the fits obtained in the literature for comparable samples when forecasting recessions (see, e.g. the first column of Table IV of Anderson et al., 2007: 76). The QPSs also appear to indicate that, relatively speaking, it is becoming easier to identify business cycle states as they occur, given the lower QPS for the most recent (post-1960) subsample.

To graphically illustrate the ex-ante recession probability assessment model’s goodness of fit, the ex-ante recession probability estimates provided by the Panel A–C results in Table 3 are graphed in Figure 1 and reveal a pronounced, clear-cut tendency of the ex-ante recession fear estimates to track recessions during the sample periods (NBER recession periods are highlighted in bold). Also of interest is the tendency of the estimated ex-ante recession fears to spike upwards prior to recessions, thus suggesting that recession fears can become self-fulfilling, an observation that can be explored further when the multivariate regression results are outlined below.

Overall, the Table 3 results indicate that the ex-ante recession probability assessment model (Equation (1)) appears to capture ex-ante recession fear assessments. This point is further rein-forced in Figure 1 by overlaying the subsample ex-ante recession probability estimates with the full sample results. The results imply that the three estimated models (the subsample and full sam-ple models) are qualitatively and quantitatively similar (see Figure 1 and Panels A–C of Table 3),

(a)

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

1960 1965 1970 1975 1980 1985 1990 1995 2000 2005

RecessionCoincident recession probability obtained from Table3ACoincident recession probability obtained from Table3C

Date

Pro

babi

lity

(b)

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

1925 1930 1935 1940 1945 1950 1955

RecessionCoincident recession probability obtained from Table3BCoincident recession probability obtained from Table3C

Date

Pro

babi

lity

Figure 1. (a)This figure plots the actual recession (shaded area) and the ex-ante recession probabilities (line) that are obtained from Panels A and C of Table 3 during January 1960–December 2005. (b) This figure plots the actual recession (shaded area) and the ex-ante recession probabilities (line) that are obtained from Panels B and C of Table 3 during July 1921–December 1959.


so the economic model underlying ex-ante recession probability assessments of the current state of the economy appears to be relatively stable.

As noted already, data for the term spread is available for a longer sample period, thus allowing a key component of the ex-ante recession probability assessment model to be further checked for consistency using an even longer sample period. Panel G of Table 3 provides results for the ex-ante recession probability assessment model for an 1857–July 1921 sample period where the lagged-term spread (St–12) is the only explanatory variable. Panels D–F and H of Table 3 provide results for this univariate probit model in the more recent sample periods, as well as in an extended sample (1857–2005), respectively. The Panel D–G results of Table 3 indicate that the 12-month lagged-term spread, by itself, does not have consistently strong explanatory power, especially in the earlier subsample periods (January 1857–July 1921), and the term spread does not appear to provide a strong ex-ante indicator of recessions from the late 1930s until at least the early 1950s. This result is illustrated in Figures 2(a) and (b), which overlay the full probit model and univariate probit model results (see Panels A and B, as well as D and E, of Table 3). The Table 3 and Figure 2 results thus imply that the term spread should not be relied upon exclusively when explaining ex-ante recession probability assessments of the current state of the economy.

Despite probit ex-ante recession probability assessment model (1) appearing to capture reces-sion fear assessments via a pronounced, clear-cut tendency of the ex-ante recession fear estimates to track recessions, the statistical significance of the individual explanatory variables in model (1) can be brought into question, because many of the probit model variables are highly persistent (see Powell et al., 2009). Panels D–G of Table 3, in conjunction with Table 10 of Powell et al.

(a)

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

1960 1965 1970 1975 1980 1985 1990 1995 2000 2005

Recession

Coincident recession probability obtained from Table3A

Coincident recession probability obtained from Table3D

Pro

babi

lity

Date

(b)

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

1925 1930 1935 1940 1945 1950 1955

Recession

Coincident recession probability obtained from Table3B

Coincident recession probability obtained from Table3E

DateP

roba

bilit

y

Figure 2. (a) This figure plots the actual recession (shaded area) and the ex-ante recession probabilities (line) that are obtained from Panels A and D of Table 3 during January 1960–December 2005. (b) This figure plots the actual recession (shaded area) and the ex-ante recession probabilities (line) that are obtained from Panels B and E of Table 3 during July 1921–December 1959.


(2009), can be used to assess the influence of dependent and explanatory variable persistence on the statistical significance of the term spread as an ex-ante recession fear explanatory variable. Table 1 reveals that the first-order autocorrelation of the term spread St–12, as well as the recession dummy variable BCt , are both approximately 0.9, so simulated 95% z-statistic confidence inter-val cut-off bounds that take account of regression variable persistence, reported in Table 10 of Powell et al. (2009), for the estimated term spread (St–12) coefficient in Panel D of Table 3 are [–4.69, 4.77]. All of the estimated term spread (St–12) coefficient z-statistics reported in Table 3 fall outside this range, thus providing an indication that the term spread appears to have significant explanatory power even after taking account of probit model dependent and explanatory variable persistence.17

3.2 Ex-ante recession fears, real aggregate stock returns and output markets

An overall indication of the importance of ex-ante recession fears in output and financial markets is provided in Tables 4(a) and (b), which report the results for univariate ex-ante recession fear effect regression models (3) and (4). Table 4(a) reveals that there is a very strong overall negative relationship between ex-ante recession fears and industrial production growth (IPt), since the t-sta-tistics for the recession fear Inverse Mill’s Ratio (IMt) coefficient are well above the 95% boot-strapped t-statistic cut-off values reported in Panels D–F. To emphasize the strength of this result, it can be noted that the Table 4(a) adjusted R2 values imply that ex-ante recessions fears explain almost twice as much of the variability of industrial production growth rates (IP), on their own, than does the ex-post NBER recession indicator BCt (compare Panel C of Table 2 with Panel F of Table 4(a)). Although the approaches are not necessarily directly comparable (due to differing time series properties), ex-ante recession fears that are estimated using ex-ante recession probability assessment model (1) also appear to have much stronger power for explaining industrial pro-duction growth rates than do recession fears that are estimated using stochastic discount factors (compare Panels D–F of Table 4(a) with Table 3 of Nieto and Rubio (2008)).

Panels A–C of Table 4(b) further indicate that the relationship between ex-ante recession fears and industrial production growth estimated in Table 4(a) remains statistically significant when lags of the dependent variable are included as control variables as an alternative method to con-trol for serial correlation. As a further sensitivity check, real quarterly GDP growth rates (GDPQ) are also used in Panels D–F of Table 4(b) as an alternative dependent variable representing output in regression Equation (4), since quarterly GDP is a non-persistent and also a more extensive measure of output than industrial production; the results are qualitatively and quantitatively simi-lar to the industrial production results.18 Similarly, the results of Table 4(a) are unaffected when including standard macroeconomic control variables (the S&P500 earnings and dividend yields) in the regression analysis, as reported in Panels G–I of Table 4(b).

Surprisingly, there is no overall relationship between estimated ex-ante recession fears and aggregate share returns apparent in Table 4(a) (see Panels A–C). This result could be due to the empirical relationship being asymmetric during recessions versus non-recessions, however, thus obscuring the overall relationship between ex-ante recession fears and real stock returns. Boyd et al. (2005) find, for instance, that the stock market reacts negatively to unemployment rate increase surprises when there is a high likelihood of a recession, and positively otherwise, thus suggesting it is possible that the relationship between ex-ante recession fears and stock returns is asymmetric (see also Henkel et al., 2007; Nieto and Rubio, 2008).

Tables 5(a) and 5(b) provide regression results that test for differential, asymmetric ex-ante recession fear effects during upturns versus recessions. The Table 5(a) Inverse Mill’s ratio


recession fear (IM) coefficient estimate reveals that ex-ante recession fears consistently have a very strong and important effect on aggregate output when the economy has not yet entered a recession, thus implying that recession fears dampen economic activity even when the fears have not already been realized via a recession (see the recession fear IM coefficient estimate in Panels D–F of Table 5(a)). This result strongly suggests the possibility that recession fears can actually become self-fulfilling, since the result implies that sizable reductions in output are possible when recession fears spike upwards prior to a recession. [It could be argued, for instance, that recession fears created by the recent financial crisis eventually became self-fulfilling.] Recall also, from Figures 1(a) and (b), that estimated ex-ante recession fears tend to spike upwards prior to reces-sions, thus further suggesting that recession fears can become self-fulfilling via their effect on output. Table 6 tests this possibility further, with Panels D and E of Table 6 indicating that the influence of recession fears on output is especially strong just prior to recessions.

Table 5(a) also reveals that ex-ante recession fears have an important influence on output fluc-tuations during downturns in the overall sample (refer to the sum of the recession fear IM and recession fear interaction variable BC*IM coefficient estimates in Panel F), thus indicating that the

Table 4(a). Ex-ante recession fear effects on business cycle variables. This table reports results for univariate regression analysis equations (3) and (4). The independent variable is the monthly Inverse Mill’s ratio (IM) estimated using multivariate probit regression model (1) (see Table 3 and Equation (2)). The t-statistics are adjusted using Newey−West HAC standard errors and covariance matrix set at lag 4. The adjusted R2 is also reported. Bootstrapped 5% t-statistic cut-offs are provided in brackets below the independent variable t-statistic. The dependent variable for each regression is listed in the first column.

Dependent variable Constant IM Adjusted R2 Sample

Panel A: Jan 1960 to Dec 2005Return (r) 0.008 −0.001 −0.001265 552 (1.35) (−0.48) (−2.13/2.20) Panel B: Jul 1921 to Dec 1959Return (r) 0.010 −0.0002 −0.002171 462 (1.11) (−0.03) (−2.19/2.15) Panel C: Jul 1921 to Dec 2005Return (r) 0.007 0.0001 −0.000987 1014 (1.21) (0.03) (−2.10/2.08) Panel D: Q1 1960 to Q4 2005Industrial production (IP) −0.036 0.031 0.407590 552 (−4.46) (9.46) (0.62/3.19) Panel E: Jul 1921 to Dec 1959Industrial production (IP) −0.214 0.177 0.512069 462 (−8.11) (11.82) (−0.99/3.16) Panel F: Jul 1921 to Dec 2005Industrial production (IP) −0.154 0.115 0.413362 1014 (−8.03) (11.04) (−0.20/3.23)


Table 4(b). Ex-ante recession fear effects on business cycle variables sensitivity analysis. This table reports results for univariate regression analysis equations (5) and (6). The independent variable is the monthly Inverse Mill’s ratio (IM) estimated using multivariate probit regression model (1) (see Table 3 and Equation (2)). In Panels A−C, two lags (IP(−2) and IP(−1)) of the dependent variable industrial production are included as additional explanatory variables. The t-statistics are adjusted using Newey−West HAC standard errors and covariance matrix set at lag 4. The adjusted R2 is also reported. The dependent variable for each regression is listed in the first column, where IP indicates monthly industrial production, GDPQ indicates quarterly GDP, EY indicates the S&P500 earnings yield (obtained from Robert Shiller’s website) and DY is the S&P500 dividend yield (also from Shiller).

Dependent variable Constant IM IP(−1) IP(−2) Adjusted R2 Sample

Panel A: Jan 1960 to Dec 2005IP −0.003 0.003 1.296 −0.376 0.946507 552

(−2.25) (4.08) (23.12) (−6.39)Panel B:Jul 1921 to Dec 1959IP −0.018 0.0155 1.483 −0.588 0.952768 462

(−2.70) (3.06) (28.76) (−11.04)Panel C: Jul 1921 to Dec 2005IP −0.007 0.006 1.483 −0.565 0.952768 1014

(−2.37) (2.92) (31.73) (−11.36)Panel D: Q1 1960 to Q4 2005 GDPQ −0.003 0.005 0.170106 184

(−1.42) (5.39) Panel E: Q3 1921 to Q4 1959GDPQ −0.024 0.023 0.109160 154

(−3.16) (4.71) Panel F: Q3 1921 to Q4 2005 GDPQ −0.015 0.014 0.078620 338

(−3.29) (5.27)

Dependent variable Constant IM EY DY Adjusted R2 Sample

Panel G: Jan 1960 to Dec 2005IP −0.015 0.030 0.264 −1.195 0.434505 552

(−1.24) (9.61) (1.15) (−1.95) Panel H:Jul 1921 to Dec 1959IP −0.177 0.166 1.021 −1.995 0.538584 462

(−3.33) (11.39) (2.08) (−2.25)Panel I: Jul 1921 to Dec 2005IP −0.151 0.112 0.701 −1.204 0.431072 1014

(−6.57) (11.61) (1.84) (−1.56)

magnitude of ex-ante recession fears (and not just the recession event itself) affects the severity of output reductions during recessions. Panel D of Table 5(a) reveals, however, that this effect appears to be disappearing in the most recent (post-1960) time period (the sum of the ex-ante recession fear IM and ex-ante recession fear interaction variable BC*IM coefficient estimates is essentially zero, thus indicating that variations in ex-ante recession fears have no influence on output during down-turns in the post-1960 sample). This result might reflect increasing (or perhaps increasingly suc-cessful) post-World War II central bank and government intervention in the economy when ex-ante recession fears peak during recessions (see also Figure 1). The sensitivity analysis provided in Panels A–C of Table 5(b) indicates that the estimated asymmetric relationship between ex-ante


recession fears and output markets does not disappear when lagged industrial production growth rates are included as control variables as an alternative method to correct for dependent variable persistence in regression (see Equation (8)). Qualitatively and quantitatively similar results are also obtained when quarterly GDP growth (GDPQ) is employed as an alternative, non-persistent output dependent variable in Panels D and E of Table 5(b). The results are also unaffected when standard macroeconomic control variables (the S&P500 earnings and dividend yields) are added as control variables in the regression analysis, as reported in Panels G–I of Table 5(b).

The Table 5(a) Panels A and B results once again indicate that there is no apparent relationship between estimated ex-ante recession fears and aggregate real stock returns, either during reces-sions or during upturns (see the ex-ante recession fear IM and ex-ante recession fear interaction variable BC*IM coefficient estimates, respectively, in Panels A–C of Table 5(a)). The Table 5(a) results therefore further imply that the overall influence of recessions on stock market returns cannot be predicted or explained using information that is available ex-ante, but is instead only

Table 5(a). Recession fear effects on business cycle variables during recessions and non-recessions. This table reports results for multivariate regression analysis equations (7) and (8). The independent variables are constructed from the monthly NBER recession dummy variable BC and the monthly Inverse Mill’s ratio (IM) that is estimated using multivariate probit regression model (1) (see Table 3 and Equation (2)). The t-statistics are adjusted using Newey−West HAC standard errors and covariance matrix set at lag 4. The adjusted R2 is also reported. Bootstrapped 5% cut-offs are provided in brackets below the t-statistics for the Inverse Mill’s ratio independent variables. The dependent variable for each regression is listed in the first column, where IP indicates industrial production.

Dependent variable Constant IM BC*IM Adjusted R2 Sample

Panel A: Jan 1960 to Dec 2005Return (r) 0.010 −0.002 −0.004 −0.001961 552 (1.65) (−0.72) (−0.55) (−2.13/2.20) (−2.45/2.42) Panel B: Jul 1921 to Dec 1959Return (r) 0.014 −0.002 −0.009 −0.000432 462 (1.56) (−0.36) (−1.17) (−2.19/2.15) (−2.39/2.43) Panel C: Jul 1921 to Dec 2005Return (r) 0.010 −0.001 −0.007 0.000259 1014 (1.63) (−0.32) (−1.10) (−2.10/2.08) (−2.25/2.28) Panel D: Jan 1960 to Dec 2005IP −0.026 0.028 −0.023 0.438347 552 (−3.01) (8.01) (−2.63) (0.62/3.19) (−2.72/3.80) Panel E: Jul 1921 to Dec 1959IP −0.182 0.166 −0.063 0.544722 462 (−6.62) (10.56) (−2.75) (−0.99/3.16) (−3.06/3.42) Panel F: Jul 1921 to Dec 2005IP −0.132 0.107 −0.041 0.441013 1014 (−6.85) (10.00) (−3.10) (−0.20/3.23) (−2.50/2.94)


observable ex-post. Interestingly, however, Panels A–C of Table 6 indicate that ex-ante recession fears have recently been associated with higher (not lower) stock returns just prior to recessions, a result that is perhaps due to stock booms at the end of upturns being associated with factors that contribute to the onset of recessions, such as central bank monetary policy tightening to combat inflation.

The lack of an overall relationship between ex-ante recession fears and returns during reces-sions is unexpected, since financial markets often appear to react to stories about recession fears during downturns, such as during the recent financial crisis. This result also contrasts, somewhat, to Resnick and Shoesmith (2002) who demonstrate, using simulation analysis, that ex-ante

Table 5(b). Recession fear effects on business cycle variables during recessions and non-recessions, sensitivity analysis. This table reports results for multivariate regression analysis equations (7) and (8). The independent variables are constructed from the monthly NBER recession dummy variable BC and the monthly Inverse Mill’s ratio (IM) that is estimated using multivariate probit regression model (1) (see Table 3 and Equation (2)). The t-statistics are adjusted using Newey−West HAC standard errors and covariance matrix set at lag 4. The adjusted R2 is also reported. The dependent variable for each regression is listed in the first column, where IP indicates monthly industrial production, GDPQ indicates quarterly GDP, EY indicates the S&P500 earnings yield (obtained from Robert Shiller’s website) and EY is the S&P500 dividend yield (also from Shiller).

Dependent variable Constant IM BC*IM IP(−1) IP(−2) Adjusted R2 Sample

Panel A: Jan 1960 to Dec 2005IP −0.0002 0.002 −0.008 1.24 −0.323 0.949687 552 (−0.17) (2.81) (−4.16) (22.12) (−5.73) Panel B: Jul 1921 to Dec 1959IP −0.011 0.014 −0.016 1.431 −0.543 0.954540 462 (−1.68) (2.71) (−4.17) (26.61) (−9.60) Panel C: Jul 1921 to Dec 2005IP −0.001 0.004 −0.012 1.429 −0.514 0.953297 1014 (−0.39) (1.85) (−5.63) (30.09) (−9.58) Panel D: Q11960 to Q4 2005GDPQ −0.001 0.004 −0.005 0.205260 184 (−0.18) (3.87) (−2.11) Panel E: Q3 1921 to Q4 1959GDPQ −0.013 0.019 −0.021 0.143935 154 (−1.50) (3.50) (−2.67) Panel F: Q3 1921 to Q4 2005GDPQ −0.007 0.011 −0.014 0.118427 338 (−1.31) (3.68) (−3.79)

Dependent variable Constant IM BC*IM EY DY Adjusted R2 Sample

Panel G: Jan 1960 to Dec 2005IP −0.011 0.028 −0.019 0.223 −0.969 0.453395 552 (−0.97) (8.44) (−2.25) (0.99) (−1.65) Panel H: Jul 1921 to Dec 1959IP −0.168 0.156 −0.066 1.169 −1.796 0.573647 462 (−3.14) (10.13) (−3.06) (2.59) (−2.15) Panel I: Jul 1921 to Dec 2005IP −0.143 0.105 −0.042 0.754 −0.972 0.458665 1014 (−6.16) (10.33) (−3.39) (2.04) (−1.29)


identification of bull and bear markets using the term spread would have enhanced stock returns during a 1970–1999 sample period.19 To reconcile this ex-ante recession fear–return result with ex-post realized real return differences in recessions (see Panel C of Table 2), the relationship between NBER recession indicator lags (BCt–n, where n indicates the number of lags in months) and real stock market returns ( rt ) is explored using the regression equation

r BCt t n t= + +−β β ε0 1 . (10)

Table 7 reveals that, consistent with Panel C of Table 2, annualized realized real returns are almost 1% lower per month during recessions versus non-recessions. Within six months, however (i.e.

Table 6. Recession fear effects on business cycle variables prior to recessions. This table reports results for multivariate regression analysis equations (7) and (8). The independent variables are constructed from the monthly NBER pre-recession dummy variable PRER, which equals 1 during the 12 months prior to a recession turning point, and 0 otherwise, and the monthly Inverse Mill’s ratio (IM) that is estimated using multivariate probit regression model (1) (see Table 3 and Equation (2)). The t-statistics are adjusted using Newey−West HAC standard errors and covariance matrix set at lag 4. The adjusted R2 is also reported. The dependent variable for each regression is listed in the first column, where IP indicates industrial production.

Dependent variable Constant IM PRER*IM IP(−1) IP(−2) Adjusted R2 Sample

Panel A: Jan 1960 to Dec 2005Return 0.009 −0.001 −0.009 0.020351 552 (1.56) (−0.25) (−5.14) Panel B: Jul 1921 to Dec 1959Return 0.010 −0.00004 −0.001 −0.004318 462 (1.11) (−0.01) (−0.20) Panel C: Jul 1921 to Dec 2005Return 0.008 0.001 −0.004 0.000442 1014 (1.25) (0.18) (−1.96) Panel D: Jan 1960 to Dec 2005IP −0.037 0.031 0.005 0.412973 552 (−4.57) (9.44) (2.31) Panel E: Jul 1921 to Dec 1959IP −0.215 0.169 0.040 0.541544 462 (−8.76) (12.11) (4.02) Panel F: Jul 1921 to Dec 2005IP −0.156 0.112 0.028 0.441539 1014 (−8.42) (11.23) (4.36) Panel G: Jan 1960 to Dec 2005IP −0.003 0.003 −0.001 1.30 −0.373 0.946511 552 (−2.15) (4.04) (−1.10) (22.96) (−6.26) Panel H: Jul 1921 to Dec 1959IP −0.019 0.016 0.003 1.479 −0.590 0.952842 462 (−2.92) (3.15) (1.65) (28.44) (−10.97) Panel I: Jul 1921 to Dec 2005IP −0.008 0.006 0.001 1.482 −0.566 0.951180 1014 (−2.52) (3.00) (1.42) (31.53) (−11.35)


when the recession dummy variable BC is lagged by six months), real returns reverse and are almost 1% higher per month on average (presumably as the economy and the stock market begin to recover, on average, from the recession), and remain so for six more months. The findings of DeStefano (2004) foreshadow this result, since DeStefano (2004) demonstrates that post-World War II stock returns are higher during the second half of recessions than during all other recession stages; investors will, however, have extreme difficulty identifying ex-ante the stage of the reces-sion (especially during short recessions), so this paper’s ex-ante analysis extends this DeStefano (2004) result.

It can also be noted from Figure 1 that estimated ex-ante recession probability assessments tend to peak towards the end of a recession when the stock market might already be looking forward to a recovery, thus muting the observed overall empirical relationship between ex-ante recession fears and returns (see Tables 4 and 5). Stock market investors could therefore view high ex-ante reces-sion probability assessments as a subsequent recovery indicator towards the end of a recession (rather than as an investor ‘fear’). To informally examine this possibility, an ex-ante investment strategy can be considered whereby investors hold stocks when the NBER announces (with a delay) a peak until the NBER subsequently announces (also with a delay) a trough (see the NBER website for the announcement dates). Data exist for the announcement of five peaks (prior to 1979

Table 7. The lagged recession indicator, and real stock return differences. This table reports results for Ordinary Least Squares (OLS) regression for various lags of the NBER business conditions indicator BC using regression equation (8): r BCt t n t= + +−β β ε0 1 . The dependent variable is the monthly real return rt and the independent variable is the lagged business condition indicator BCt–n which equals 1 if there is a recession at month t and is zero otherwise, while n indicates the number of lags. The first column indicates the number of lags of the independent variable BC and the second column reports the estimated lagged business condition indicator (BC) regression coefficient. The second column coefficient estimate represents the estimated difference in average returns when a recession occurs versus when a recession does not occur at month t–n. The (OLS) regression constant is not reported.

Lag in months (n) BCt−n coefficient t-statistics Adjusted R2 Obs.

0 −0.0085 −3.039 0.005512 1788 1 −0.0038 −1.393 0.000680 1787 2 −0.0004 −0.142 −0.000546 1786 3 0.0028 0.940 0.000100 1785 4 0.0050 1.685 0.001526 1784 5 0.0079 2.688 0.004692 1783 6 0.0088 3.069 0.006039 1782 8 0.0072 2.607 0.003806 178010 0.0077 2.661 0.004590 177812 0.0084 2.970 0.005621 177614 0.0049 1.795 0.001565 177416 0.0054 1.966 0.001971 177218 0.0041 1.537 0.000905 177020 0.0056 2.075 0.002157 176822 0.0062 2.250 0.002755 176624 0.0019 0.678 −0.000268 176426 0.0020 0.753 −0.000213 176228 0.0020 0.807 −0.000202 176030 −0.0002 −0.094 −0.000564 1758


the NBER did not make formal announcements),20 and the strategy provides positive returns during four of the five ex-ante recession end sample periods, with the average return being just under 17% for an average holding period of just over one year.

While it is clear that ex-ante recession fears do not explain concurrent share returns, it is pos-sible that estimated ex-ante recession fears can be indirectly used to explain stock returns via their relationship with industrial production, especially since industrial production growth rates are an important macroeconomic asset pricing risk factor (see, e.g. Liu and Zhang, 2008). Campbell and Diebold (2005) demonstrate, for instance, that expected real GDP growth forecasts obtained from the Livingstone GDP growth forecast can be used to forecast aggregate stock returns (see also Goetzmann et al., 2007).

Table 8 therefore tests whether ex-ante recession fears might have an indirect rather than a direct effect on stock returns by examining whether the relationship between output and ex-ante reces-sion fears that is documented in Tables 4 and 5 could be indirectly used to explain or forecast stock returns (see also Chen et al., 1986). Panels A–C of Table 8 use the industrial production growth rate estimates (EIPt) implied by Panels D–F of Table 4 to explain concurrent real stock returns rt :

r EIPt t t= + +β β ε0 1 . (11)

The expected industrial production growth rate, EIPt, is estimated using regression Equation (4) once the Inverse Mill’s Ratio IMt estimate for each date and the estimated parameter values for Equation (4) are substituted in to regression Equation (4). The results reported in Table 8 are con-sistent with the Table 4 and 5 findings, since no relationship between real returns and recession fear-based forecasts of industrial production growth is found. A related, final check is to examine whether the term spread (St–12) explains real stock returns rt directly (see also Campbell, 1987; Chen et al., 1986) using the regression equation

r St t t= + +−β β ε0 1 12 . (12)

Once again, no indirect stock market return–ex-ante recession fear relationship is found (see Panels D–F of Table 8). Fama and French (1989) indicate, however, that the term spread forecasts stock market returns with a one-month lag, so Panels G–I of Table 8 employ a one-month (instead of a 12-month) lag of the term spread. Predictability is apparent in the most recent subsample only (see Panel G of Table 8), but once again no indirect stock market return–ex-ante recession fear relationship can be inferred, because ex-ante recession probability assessments are unrelated to the term spread when the term spread is lagged by one month only (results not reported).

It is important to note that the Table 8 results do not necessarily imply that expected industrial production growth is cross-sectionally irrelevant to the stock market (see also Liu and Zhang, 2008). Ex-ante recession fears are likely to be an important macroeconomic risk factor cross-sectionally, since firms that are more sensitive to recessions would be expected to do worse, all else being equal, when the risk of a recession is perceived ex-ante to be higher, a possibility that could be explored in future research.

3.3 Application to Australian markets

The paper’s approach can easily be applied to other important countries of the world such as Australia, a major financial centre. In Australia (and elsewhere), there is no equivalent to the National Bureau of Economic Statistics recession dating committee, so the standard procedure is


to define recessions as two consecutive quarters of contraction in GDP (see Bry and Boschan, 1971). Data for this purpose can be obtained from the Reserve Bank of Australia from the fourth quarter of 1959 onwards, thus allowing the recession dummy variable (BC) to be constructed quar-terly. Many of the important explanatory variables used to explain the ex-ante probability of a recession (see Equation (1)) are provided quarterly by the Australian Reserve Bank from 1976 onwards (or earlier), including the lagged-term spread (St–4), the lagged rate of change of retail

Table 8. Forecasting real stock returns using expected industrial production growth rates and the lagged term spread. This table reports regression results where the monthly real return rt is the dependent variable and either the expected industrial production growth rate EIP (see Equation (8)) or the lagged term spread Spread(−12) (see Equation (9)) is the independent variable. The expected industrial production growth rate, EIPt, is estimated using regression equation (4) once the Inverse Mill’s Ratio IMt estimate for each date and the equation (4) estimated parameter values are substituted in to regression equation (4). The t-statistics are adjusted using Newey–West HAC standard errors and covariance matrix set at lag 4. The adjusted R2 is also reported.

Constant EIP Adjusted R2 Obs.

Panel A: Jan 1960 to Dec 2005 0.007 −0.036 −0.001265 552 (1.95) (–0.48) Panel B: Jul 1921 to Dec 1959 0.010 −0.001 −0.002171 462 (2.63) (−0.03) Panel C: Jul 1921 to Dec 2005 0.008 0.001 −0.000987 1014 (3.26) (0.03)

Constant Spread(−12) Adjusted R2 Obs.

Panel D: Jan 1960 to Dec 2005 0.004 0.002 −0.000095 552 (1.03) (0.94) Panel E: Jul 1921 to Dec 1959 0.005 0.004 0.001206 462 (0.89) (0.92) Panel F: Jul 1921 to Dec 2005 0.005 0.002 0.001015 1014 (1.57) (1.25)

Constant Spread(−1) Adjusted R2 Obs.

Panel G: Jan 1960 to Dec 2005 –0.00002 0.004 0.009973 552 (–0.01) (2.56) Panel H: Jul 1921 to Dec 1959 0.011 −0.001 −0.001897 462 (2.32) (–0.29) Panel I: Jul 1921 to Dec 2005 0.005 0.002 0.000965 1014 (1.59) (1.24)


sales (Salet–1) and lagged GDP growth (GDPt–2), while lagged stock index return data (rt–1) for the Australian All Ordinaries Index can be obtained from Datastream.21 The full ex-ante recession probability assessment probit model, when applied to Australia, is

P BC y N S r GDP Salet t t t( ) ( ,== == ++ ++ ++ ++-- -- -- --1 1 4 2 1 3 2 4 2α β β β β (13)

where P denotes probability, BCt is a binary dependent variable equal to 1 if a recession is occur-ring in quarter t and 0 otherwise, N is the cumulative standard normal distribution function, a is a constant and the lagged business cycle explanatory variables are defined as: St–4 is the 10-year government bond yield minus the 90-day bank bill rate at the end of quarter t–4, GDPt–2 is the quarterly log change in real GDP (lagged by two quarters), rt–1 is the log change in the quarterly return on the Australian Stock Exchange All Ordinaries index (lagged by one quarter) and Salet–2 is the log rate of change of real retail sales (lagged by two quarters).

Results for ex-ante recession probability assessment probit model (13) are provided in Panel A of Table 9. The results are qualitatively similar to the corresponding results for the United States (compare Panel A of Table 9 with Table 3), with the overall goodness of fit also being roughly simi-lar (see, e.g. the McFadden adjusted R2 level in each table). It is interesting to note from Panel A of Table 9, however, that the term spread is nowhere near as important in Australia as in the United States for foreshadowing recessions using multivariate probit model (13) during the sample period. This result is important, because downward movements in the term spread precede recessions in many countries (see Marcellino, 2005), so it is interesting to speculate as to whether the Reserve Bank of Australia’s recent narrower focus on inflation (and not GDP growth) relative to the United States Federal Reserve could help to explain this interesting result.

Panels B–E of Table 9 reproduce the results testing the influence of ex-ante recession fears on business cycle variables in Australia (see Tables 4 and 5). The results are qualitatively and quanti-tatively similar in both countries (compare Panels B–E of Table 9 with Tables 4(a) and (b) and 5(a) and (b)). [It should be noted that change in GDP is used as the only economic growth dependent variable in Panels C and E of Table 9, not the industrial production growth rate, since industrial production data is not systematically compiled in Australia.] Panels C and E of Table 9 reveal that, as in the United States, recession fears have an important influence on output at all times, whereas Panel D of Table 9 reveals that recession fears have an especially severe effect on output during recessions in Australia (the estimated BC*IM coefficient, –0.008, is highly significant). Once again, recession fears are found to be unrelated to returns (see Panels B and D of Table 9). A final observation that emerges from the application of the paper’s approach to Australia is to note that it would be useful for Australian recession studies if monthly industrial production was to be com-piled and backdated, especially since industrial production growth rates have been found to be an important macroeconomic asset pricing risk factor (Campbell and Diebold, 2005; Goetzmann et al., 2007; Liu and Zhang, 2008).

4. Conclusion

This paper examines whether ex-ante recession fears influence aggregate real stock returns and output growth rates, and explores whether ex-ante recession fears have their greatest impact when the economy moves into a recession. The paper therefore contributes, both in its approach and its results, to an increasing interest in the finance literature as to whether recessions and ex-ante reces-sion fears have differential effects on financial and output markets during downturns versus upturns (see, e.g. Boyd et al., 2005; Henkel et al., 2007; Nieto and Rubio, 2008).


A probit model analysis of ex-ante recession probability assessments of the current state of the economy indicates that the lagged-term spread, lagged real stock returns and a lagged composite index of four business cycle coincident indicators play a consistent role in ex-ante recession prob-ability assessments (i.e. ex-ante recession fears). The study employs an extended 85-year sample, as well as two subsamples, in order to establish the consistency of the ex-ante recession probabil-ity assessment model. Interestingly, the term spread cannot be relied upon in isolation to explain ex-ante recession probability assessments of the current state of the economy in an extended sample, thus reinforcing the need to employ a multivariate probit model when estimating ex-ante recession fears.

Univariate and multivariate second-stage regression results indicate that estimated ex-ante recession fears appear to have an important effect on output whether or not a particular time period is classified ex-post as a recession. The analysis could therefore suggest a model of the economy whereby economic agents’ uncertainty about the current state of the economy is important enough that economic activity reacts to any increase in ex-ante recessionary fears, regardless of the current

Table 9. Australian results. Panel A of this table reports result for the multivariate probit regression model for the Australian data, specified as P BC y N S r GDP Salet t t t t t( | ) ( )= = + + + +− − − −1 1 4 2 1 3 2 4 2α β β β β , where P denotes probability, BCt is a binary dependent variable equal to 1 if a recession is occurring in quarter t and 0 otherwise, N is the cumulative standard normal distribution function, a is a constant and the lagged business cycle explanatory variables are defined as: St−4 is the 10-year government bond yield minus the 90-day bank bill rate at the end of quarter t−4, GDPt−2 is the quarterly log change in real GDP (lagged by two quarters), rt−1 is the log change in the quarterly return on the Australian Stock Exchange All Ordinaries index (lagged by one quarter) and Salet−2 is the log rate of change of real retail sales (lagged by two quarters). The z-statistics are reported in brackets. McFadden R2 is also reported. (-l) in the first row indicates the independent variable is lagged by l times. Panels B and C of this table report results for univariate regression analysis equation (3), while Panels D and E of this table report results for multivariate regression analysis equation (7). The independent variable in Panels B−E is the quarterly Inverse Mill’s ratio (IM) estimated using the multivariate probit regression model from Panel A. The t-statistics are adjusted using Newey−West HAC standard errors and covariance matrix set at lag 4. The adjusted R2 is also reported. The dependent variable for each regression is listed in the first column. The data set is constructed using Australian Reserve Bank data, except for the stock return which is constructed using Datastream.

Dependent variable Constant S(−4) r(−1) GDP(−2) Sale(−2) McFadden R2 Obs

Panel A −0.9532 −0.0921 −4.8674 −39.618 −19.84 0.28975 83BC (−3.93) (−1.01) (−2.47) (−1.77) (−1.80)

Dependent variable Constant IM Adjusted R2 Obs

Panel B 0.0671 −0.0216 0.006130 83Return (2.07) (−1.42) Panel C −0.0028 0.0056 0.130059 83GDP (−0.66) (2.75)

Dependent variable Constant IM BC*IM Adjusted R2 Obs

Panel D 0.0788 −0.0257 −0.0247 0.004968 83Return (2.41) (−1.68) (−1.03) Panel E 0.00102 0.0043 −0.0080 0.255176 83GDP (0.27) (2.30) (−8.62)


state of the economy (albeit from a lower trend line starting point during recessions). The results further suggest that ex-ante recession fears can become self-fulfilling due to their dampening effect on output leading up to recessions, since ex-ante recession fear effects are shown to be even more important prior to, versus within, a recession.

Ex-ante recession fears do not appear to be related to or capable of forecasting aggregate real stock returns, either directly or indirectly, thus implying that lower aggregate real stock returns during recessions are only evident, overall, ex-post. This unexpected result is subsequently explained by noting that, while stock returns are very low during recessions, lagged recession indi-cators are associated with very strong subsequent real stock returns, perhaps as the economy begins to recover, so the association between ex-ante recession fears and returns is considerably muted. The analysis therefore implies that the influences of recessions on stock returns cannot be directly explained or predicted using ex-ante recession fears, and would instead have to be modelled using a lag structure of the inter-relationship between recessions, ex-ante recession fears and subsequent economic recovery. Ex-ante recession fears could also be an important cross-sectional macroeco-nomic risk factor, thus suggesting an avenue for future research, since firms that are more sensitive to recessions would be expected to be affected to a greater extent, and might recover more slowly, as the economy potentially moves into and out of recession.

Funding

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

Acknowledgement

We would like to thank Heather Anderson, Campbell Harvey, Timothy Kam, Michael Martin, Jing Shi, and Tom Smith for helpful comments.

Notes

1 An alternative approach in the literature is to use leading indicators, such as the term spread, to directly forecast output or consumption growth rates (see, e.g., Harvey, 1988, 1989), whereas this paper’s approach is to directly explore the influence of ex-ante recession fears on business cycle outcomes, with recession fears being estimated from coincident indicators.

2 Occasionally in the literature, alternative approaches are also used. One alternative approach is to use Hamilton regime switching analysis to estimate ex-ante recession fears (see, e.g., Hamilton, 1994; Henkel et al., 2007). The estimated recession regimes do not necessarily correspond, however, to NBER dated recessions. Another alternative is to use stochastic discount factor estimates to infer recession fears (Nieto and Rubio, 2008); although the estimated stochastic discount factor is shown to be correlated with recessions, it is not, however, a direct estimate of recession fears.

3 In this study, the quarterly log GDP growth rate has (by necessity) the same recorded value each month of the quarter, thus inducing persistence. The results of the paper are insensitive, however, to various corrections for persistence, such as utilizing annual rather than monthly data, as indicated in Section 3.

4 The Livingstone growth rate forecast L has the same value each month of the half-year, thus inducing persistence, but the results of the study are insensitive to various corrections for persistence, as indicated in Section 3.

5 Fama and French (1989) find that the term spread lagged by one month only predicts stock returns, a find-ing that is discussed at the end of Section 3, but the term spread lagged by one month only is not related to ex-ante recession probability assessments of the current state of the economy (results not reported).

6 Following Heckman (1979), estimated probabilities are transformed into an Inverse Mill’s ratio regres-sion explanatory variable and are employed in a two-step estimation procedure, a standard approach to


sample selection truncation (e.g. the influence of recession fears would otherwise only be observed when a recession actually occurs; see Greene, 2003: 784). Sensitivity analysis indicates that the results are not, however, sensitive to the use of an Inverse Mill’s Ratio recession fear explanatory variable versus the use of the original ex-ante recession fear probability estimates as a regression model explanatory variable (results not reported).

7 Goyal and Welch (2008) use recursive bootstrapping to take account of explanatory variable persistence, whereas in this study dependent variable persistence could also be important, so recursive bootstrapping of both the dependent and independent variable series is utilized. Autocorrelated simulation could also be used (see Ferson et al., 2003; Granger et al., 2001). Simulated confidence intervals are therefore also constructed (but not reported) for regression Equations (3) and (4) as an extra sensitivity check.

8 Quarterly GDP data observations and the end of quarter month’s Inverse Mill’s ratio (IMt) observations are used to estimate regression equation (6).

9 Boyd et al. (2005) analyse differences in investors’ reactions to unemployment news surprises during recessions versus non-recessions, with recession states being identified ex-ante. Henkel et al. (2007) examine whether stock market return predictability disappears during upturns, with recession prob-abilities being estimated using Hamilton regime switching analysis. Nieto and Rubio (2008) examine whether economic fears, as represented by an estimated stochastic discount factor, are more pronounced during downturns as opposed to upturns.

10 To determine pre-recession effects, the dummy variable BCt is replaced in regression Equations (7) and (8) by a pre-recession dummy variable PRERt, which equals one during the 12 months prior to a reces-sion turning point and zero otherwise. A significantly positive Inverse Mill’s ratio explanatory variable coefficient b1 indicates that higher ex-ante recession probability assessments are associated with lower real returns or output growth rates. A significantly positive (negative) Inverse Mill’s ratio interaction variable coefficient b2 indicates that higher ex-ante recession probability assessments are associated with return or output growth rate effects prior to recessions that are even stronger (weaker) than at all other times.

11 Where possible, longer term samples are also examined as an extra sensitivity check of the paper’s main results, thus providing, for instance, the possibility to check even more extensively the consistency of the study’s ex-ante recession probability assessment results. The data sample for 10-year bond yields can be extended back to 1854 using Financial Review and Commercial and Financial Chronicle data made available by Global Financial Data. The short-term interest rate series is extended by following the Schwert (1989) procedure whereby pre-1920 short-term commercial paper yields, available at Global Financial Data, are adjusted to consistently match treasury bill yield characteristics (0.91% per annum is subtracted from the per annum commercial paper yield to make this adjustment). This enables the term spread variable S to be created from January 1857 onwards.

12 Schwert (1990) index total returns are obtained from Global Financial Data.13 Note that the variables are available with an information release delay, so their contemporaneous values

cannot be used to model ex-ante recession fears.14 Results for Table 3 are virtually identical when model (1) is estimated using logit regression analysis

(see Birchenhall et al., 1999; Stock and Watson, 1989), instead of probit analysis, so the indicator variables rather than the form of estimation appear to be responsible for the explanatory power of the model (results not reported).

15 To further emphasize this point, it can be noted that the coincident term spread (the term spread lagged by one month only) is unrelated to recessions (results not reported), whereas (as noted) the term spread lagged by 12 months foreshadows recessions.

16 The estimated ex-ante recession probabilities, when rounded, almost never misclassify a time period as a recession when (ex-post) a recession does not occur, but misclassify close to half of the recession months as being non-recessions. Overall concordance of the model can therefore be further optimized by adjusting downwards (below a half) the point at which a recession is classified ex-ante by the model.

17 Whenever possible, care has been taken when discussing the Table 3 results to indicate that ex-ante recession probability model (1) appears to display reasonable goodness of fit and tracks recessions well,


thus capturing investors’ recession fears, rather than focusing on whether its component explanatory variables are statistically significant (other than pointing out when explanatory variables are clearly insignificant).

18 The Table 4(a) results are also virtually identical when only one observation per year is used in regression Equation (4) in order to lessen dependent and independent variable persistence (results not reported).

19 It can be noted, however, that the profitability of the Resnick and Shoesmith (2002) share trading strat-egy is largely driven by the avoidance of only two bear markets in their sample.

20 The S&P500 returns are roughly 30%, –13%, 12%, 37% and 18% for the most recent ‘recession end’ sample periods, with the most recent recession end period not yet being complete (through December 2009).

21 Not all of the series used to construct the United States Conference Board Coincident Indicator Index are available in Australia for an extended sample period (specifically, industrial production is not available), so the rate of change of retail sales is used to represent the Coincident Indicator (recall that the compo-nents of the Coincident Indicator are highly correlated, as indicated in Table 1).

References

Anderson HM, Athanasopoulos G and Vahid F (2007) Nonlinear autoregressive leading indicator models of output in G-7 countries. Journal of Applied Econometrics 22: 63–87.

Bernanke BS (1983) Non-monetary effects of the financial crisis in the propagation of the Great Depression. American Economic Review 73: 257–276.

Birchenhall CR, Jessen H, Osborn DR and Simpson P (1999) Predicting U.S. business-cycle regimes. Journal of Business and Economic and Statistics 17: 313–323.

Boulier BL and Stekler HO (2000) The term spread as a monthly cyclical indicator: an evaluation. Economics Letters 66: 79–83.

Boyd JH, Hu J and Jagannathon R (2005) The stock market’s reaction to unemployment news: why bad news is usually good for stocks. Journal of Finance 60: 649–672.

Bry G and Boschan C (1971) Cyclical analysis of time series: procedures and computer programs. Technical Paper 20, NBER, Columbia University Press.

Campbell J (1987) Stock returns and the term structure. Journal of Financial Economics 18: 373–399.Campbell SD and Diebold FX (2005) Stock returns and expected business conditions: half a century of direct

evidence. University of Pennsylvania Working Paper.Chen N-F, Roll R and Ross SA (1986) Economic forces and the stock market. Journal of Business 59: 383–403.DeStefano M (2004) Stock Returns and the business cycle. Financial Review 39: 527–547.Efron B and Tibshirani R (1986) Bootstrap methods for standard errors, confidence intervals, and other

measures of statistical accuracy. Statistical Science 1: 54–77.Estrella A and Hardouvelis GA (1991) The term structure as a predictor of real economic activity. Journal of

Finance 46: 555–576.Estrella A and Mishkin FS (1998) Predicting U.S. recessions; financial variables as leading indicators. Review

of Economics and Statistics 80: 45–61.Fama E and French K (1989) Business conditions and expected returns on stocks and bonds. Journal of

Financial Economics 25: 23–49.Ferson WE, Sarkissian S and Simin TT (2003) Spurious regressions in financial economics? Journal of

Finance 58: 1393–1413.Goetzmann WN, Watanabe A and Watanabe M (2007) Investor expectations, business conditions, and the

pricing of beta-instability risk. Yale University Working Paper.Goyal A and Welch I (2008) A comprehensive look at the empirical performance of equity premium prediction.

The Review of Financial Studies 21: 1455–1508.Granger CWJ, Hyung N and Jeon Y (2001) Spurious regressions with stationary series. Applied Economics

33: 899–904.Greene WH (2003) Econometric Analysis. Englewood Cliffs, NJ: Prentice Hall.Hall RE (1991) Recession dating. Stanford University Working Paper.Hamilton JD (1994) Time-series Analysis. Princeton, NJ: Princeton University Press.


Harvey CR (1988) The real term structure and consumption growth. Journal of Financial Economics 22: 305–333.

Harvey CR (1989) Forecasts of economics growth from the bond and stock markets. Financial Analyst Journal 45: 38–45.

He J and Ng LK (1998) The foreign exchange exposure of Japanese multinational corporations. Journal of Finance 53: 733–753.

Heckman JJ (1979) Sample selection bias as a specification error. Econometrica 47: 153–160.Henkel SJ, Martin JS and Nardari F (2007) Short-horizon predictability and information erosion. Indiana

University Working Paper.Li H and Maddala GS (1997) Bootstrapping cointegrating regressions. Journal of Econometrics 80: 297–318.Liu LX and Lu Z (2008) Momentum profits, factor pricing, and macroeconomic risk. The Review of Financial

Studies 21: 2417–2448.Marcellino M (2005) Leading indicators: What have we learned? Innocenzo Gasparini Institute for Economic

Research Working Paper.Nieto B and Rubio G (2008) The volatility of consumption-based stochastic discount factors and economic

cycles. University of Alicante Working Paper.Öncü TS (2006) A parametric estimator for multi-state duration models with an application: the yield curve

slope as a predictor of the U.S. recessions. Bilkent University Working Paper.Politis DN and Romano JP (1994) The stationary bootstrap. Journal of the American Statistical Association

89: 1303–1313.Powell JG, Shi J, Smith T and Whaley R (2009) Political regimes, business cycles, seasonalities, and returns.

Journal of Banking and Finance 33: 1112–1128.Prabhala NR (1997) Conditional methods in event studies and an equilibrium justification for standard event-

study procedures. The Review of Financial Studies 10: 1–38.Resnick BG and Shoesmith GL (2002) Using the yield curve to time the stock market. Financial Analysts

Journal 58: 82–90.Rudebusch GD and Williams JC (2007) Forecasting recessions: the puzzle of the enduring power of the yield

curve. Federal Reserve Bank of San Francisco Working Paper.Schwert GW (1989) Why does stock market volatility change over time? Journal of Finance 44: 1115–1153.Schwert GW (1990) Indexes of United States stock prices from 1802 to 1987. Journal of Business 63:

399–426.Siegel JJ (1998) Stocks for the Long Run. New York: McGraw Hill.Stock JH and Watson M (1989) New indexes of coincident and leading economic indicators. In: Blanchard O

and Fischer S (eds) NBER Macroeconomics Annual. Cambridge, MA: MIT Press, 351–394.

Date of acceptance of final transcript: 24 August 2011.Accepted by Associate Editor, Garry Twite (Finance).

Appendix

The univariate regression recursive bootstrapping procedure works as follows. An AR1 regres-sion is estimated for the independent variable using the actual data (the natural logarithm of the Inverse Mill’s ratio is used to avoid negative bootstrapped values). The errors from this AR1 regression are stored, and the regression constant plus the regression AR1 coefficient estimate are also recorded. Similarly, an AR1 regression is estimated for each of the dependent variable series (real stock returns and real industrial production growth rates), with the errors from this regres-sion also being stored, along with the regression parameter estimates. A starting date for both series is randomly chosen, and the first bootstrapped observation for each series is constructed by randomly selecting a stored AR1 error that is added to the stored regression constant plus the stored AR1 coefficient times the prior period’s bootstrapped observation. The process is repeated


until a bootstrapped sample is created for the dependent and independent variables, and regres-sion Equation (3) or (4) is run on the bootstrapped sample. The explanatory variable t-statistic is recorded, the process is repeated 10,000 times and the t-statistics are ranked from highest to low-est to create a 5% bootstrapped cut-off (using the 250th and 9750th recorded t-statistics).

To accommodate the paper’s multivariate regression analysis, bootstrapped recession (BCt) series are also constructed. Recession and non-recession durations (in months) are recorded from the actual sample, and these stored durations are randomly selected to construct bootstrapped recession/ non-recession samples (see also Politis and Romano, 1994). Bootstrapped t-statistic cut-off values are then constructed once multivariate regressions on the bootstrapped dependent and independent variable series are estimated. Bootstrapping has the advantage, relative to simulation analysis, of preserving the distributional properties of the dependent variable series, even when they do not closely conform to the normal or lognormal distribution assumptions.

Recession fears as self-fulfilling prophecies? Influence on stock returns and output

Documents