THE JOURNAL OF FINANCE • VOL. LX, NO. 2 • APRIL 2005

The Stock Market’s Reaction to Unemployment News: Why Bad News Is Usually Good for Stocks

JOHN H. BOYD, JIAN HU, and RAVI JAGANNATHAN∗

ABSTRACT

We find that on average, an announcement of rising unemployment is good news for stocks during economic expansions and bad news during economic contractions. Un- employment news bundles three types of primitive information relevant for valuing stocks: information about future interest rates, the equity risk premium, and corpo- rate earnings and dividends. The nature of the information bundle, and hence the relative importance of the three effects, changes over time depending on the state of the economy. For stocks as a group, information about interest rates dominates dur- ing expansions and information about future corporate dividends dominates during contractions.

THIS STUDY INVESTIGATES THE SHORT-RUN response of stock prices to the arrival of macroeconomic news. The particular news event we consider is the Bureau of Labor Statistic’s (BLS) monthly announcement of the unemployment rate. We establish that the stock market’s response to unemployment news arrival depends on whether the economy is expanding or contracting. On average, the stock market responds positively to news of rising unemployment in expansions, and negatively in contractions. Since the economy is usually in an expansion phase, it follows that the stock market usually rises on the announcement of bad news from the labor market.1

∗Carlson School of Management, University of Minnesota, Moody’s Investors Service, Kellogg Graduate School of Management, Northwestern University and National Bureau of Economic Re- search. The authors benefited from comments by workshop participants at the June 2001 European Financial Management Meetings at Lugano, Federal Reserve Bank of New York, Federal Reserve Bank of Atlanta, McGill University, University of Akron, University of Vienna, and from comments by Olivier Blanchard, Jacob Boudoukh, Ross Levine, Roberto Rigobon, two anonymous referees, and the editor of the journal. We particularly benefited from discussions with Gordon Alexander and Sergio Rebelo. We are grateful to Bhaskaran Swaminathan for providing us with monthly data on the intrinsic value to market value ratios for the Dow 30 Index. Qianqiu Liu provided valuable research assistance. Any views expressed in the paper are those of the authors and are not necessarily those of the institutions they represent.

1 For example, on December 6, 1974, the Labor Department released substantial bad news: The unemployment rate had risen from 6.0% to 6.6%. Around the announcement, the S&P 500 index de- clined by about 3.6%. However, it is just as easy to find cases in which the stock market rose sharply in response to bad unemployment news. On August 3, 1984, the Labor Department announced that the unemployment rate had increased from 7.2% to 7.5%, and around that announcement the S&P 500 index gained 5.4%. It is no coincidence that the first case occurred during a contraction and the second during an expansion.

649

650 The Journal of Finance

We next provide an explanation for this seemingly odd pattern of stock price responses. Conceptually, three primitive factors determine stock prices: the risk-free rate of interest, the expected rate of growth of corporate earnings and dividends (hereafter, “growth expectations”), and the equity risk premium. Thus, if unemployment news has an effect on stock prices, that must be because it conveys information about one or more of these primitives.

We begin our explanation by determining whether the pattern of stock price responses can be explained solely by information about future interest rates. If this were the case, stock and bond prices would respond in the same way, except for differences that might arise due to differences in their durations. They do not. During contractions, stock prices react significantly and negatively to rising unemployment, but bond prices do not react in any significant way. Since bond prices do not respond significantly during contractions, it must be the case that unemployment news contains little information about future interest rates in that business cycle phase. Since stock prices do respond significantly during contractions, it must also be the case that the unemployment news contains information about growth expectations and/or the equity risk premium.

During expansions, both bond and stock prices rise significantly on the an- nouncement of rising unemployment. Given the bond response, it must be the case that during expansions, bad labor market news causes expected future interest rates to decline. This could also be what causes stock prices to rise during expansions, but it need not be, since growth expectations and the equity risk premium could be changing also. For example, suppose the real interest rate remains the same, but inf lation goes down when unemployment goes up. This would result in a decline in the nominal interest rate and would be good news for bonds. If higher unemployment also signals lower real earnings in the future on equities, stock prices need not go up.

The next step in understanding the pattern of stock price responses over the business cycle is to examine the effect of news arrival on the other two primitive factors: the equity risk premium and growth expectations. We must use proxy measures for both variables since neither is directly observable. In brief, we find evidence that an unanticipated increase in unemployment may lead to an increase in the risk premium during expansions, but we find no evidence of an effect during contractions.

There is also evidence that growth expectations change in response to the unemployment news. Specifically, we find that unemployment news is helpful in predicting the actual growth rate of the index of industrial production (IIP), one proxy measure for growth expectations. Rising unemployment is always followed by slower growth, but this relationship is much stronger during con- tractions than it is during expansions. Thus, if equity investors study the real sector data, they would be expected to revise their growth expectations more significantly during contractions than during expansions.

A. Related Literature

Blanchard (1981) shows that in equilibrium, the same news can sometimes be good and sometimes bad for financial assets, depending on the state of the

Stock Market’s Reaction to Unemployment News 651

economy. This study can be viewed as providing the necessary theoretical mo- tivation for our work. Orphanides (1992) gives empirical support for this view by showing that stock price responses to macroeconomic news may depend on the state of the economy. In particular, he shows that the stock price response to unemployment news depends on the average unemployment rate during the previous year.

McQueen and Roley (1993) also find a strong relationship between stock prices and macroeconomic news, such as news about inf lation, industrial pro- duction, and the unemployment rate, based on their own definition of business conditions. However, their purpose is to demonstrate the state dependence of stock price responses to all macroeconomic news. Krueger (1996) studies the market rationality of bond price responses to labor market news. His focus is on the market reaction to the availability of more reliable information, as the unemployment data are revised. His study found (as we do) that market prices were strongly affected by the unemployment announcements. Fleming and Remolona (1998) analyze the response of U.S. Treasury yields across the maturity spectrum to different types of macroeconomic announcements using high frequency data over four-and-one-quarter years. They find that the yields on intermediate term bonds are most sensitive to unemployment news.

Veronesi (1999), based on theoretical arguments, shows that bad news in good times and good news in bad times would generally be associated with in- creased uncertainty and hence with an increase in the equity risk premium. Jagannathan and Wang (1993) find that monthly stock returns are negatively correlated with the per-capita labor income growth rate. Jagannathan, Kubota, and Takehara (1998) report similar findings using Japanese data. Since most of the variation in per-capita labor income arises from variation in hours worked and not from the wage rate, these findings are consistent with the uncondi- tional positive correlation between the growth rate in unemployment and stock returns that we find in our data set.

B. The Rest of the Study

Brief ly, the rest of the study proceeds as follows. Section I describes the data set and the empirical methods for forecasting unemployment rates. Section II examines the effect of unemployment news on the S&P 500 stock index portfolio returns and on government bond returns. Section III examines how unemploy- ment news affects growth expectations and the equity risk premium. Finally, Section IV summarizes and concludes.

I. Data and Methodology

A. Unemployment Announcements

Although there are various macroeconomic information releases we could have considered, we chose the unemployment rate because it is viewed as news- worthy. It has frequently been the reference point of Federal Reserve policy and the target of wide speculation on Wall Street. In addition, and important for our purposes, this release has a long and accurately dated time series.

652 The Journal of Finance

The monthly unemployment announcements used in this paper cover the pe- riod from February 1957 to December 2000. The announcements were usually made at 8:30 a.m. on a Friday, although during earlier years some announce- ments were made on other days. All announcement dates, whether Fridays or not, are included in our study. On announcement days, the Department of Labor releases other information besides the most recent unemployment rate. This includes the total number of employed and its distribution across regions and industries. It also releases revisions of past unemployment announcements for the previous 3 months, after which the announcement is considered final. It also releases employment totals, weekly and hourly earnings, and weekly hours worked. This study focuses only on unemployment rate announcements.

A.1. Measuring Unemployment News

The focus of this paper is on examining how stocks respond to unemploy- ment news. That requires a model to measure the anticipated and the unan- ticipated (news) component of the unemployment figures that are announced every month.2 We use the following statistical model to forecast the unemploy- ment rate change on announcement dates:

DUMPt = b0 + b1 · IPGRATEt−1 + b2 · IPGRATEt−2 + b3 · IPGRATEt−4 + b4 · DUMPt−1 + b5 · DTB3t + b6 · DBAt + et , (1)

where DUMPt is the change in the unemployment rate, IPGRATEt is the growth rate of monthly industrial production, DTB3t is the change in the 3-month T-bill rate, and DBAt is the change in the default yield spread between Baa and Aaa corporate bonds, all for the periods t − 1 and t. The unemployment rate is very persistent, so our forecasts are based on the first differences.3 We used data for the period 1957 to 1962 to select the specification of the time series model used in constructing unemployment surprises. Appendix B describes the model selection procedure we used to choose the specification in equation (1).

Note that for these and most of the other regressions presented in this pa- per, both heteroskedasticity and autocorrelation are present in the residuals. We therefore compute hetereoskedasticity and autocorrelation consistent stan- dard errors and t-statistics with the Bartlett kernel. The bandwidth parameter is chosen to match the degree of autocorrelation in the residuals, where the length of autocorrelation is first estimated by the Yule-Walker method. For

2 McQueen and Roley (1993) and Krueger (1996) used forecasts made by Money Market Services International (MMS) to identify the surprise element of the unemployment rate announcement. We do not follow this procedure since MMS forecasts have only been available since November 1977, whereas our data set goes back to January 1962. Seeking to employ as much data as possible, we use our own time-series models to forecast the unemployment rate announcement and its unanticipated component.

3 Regression model (1) can be expanded to include Friday and day of the week dummy variables to account for the fact that announcements were not always made on Fridays. We do not report these results since inclusion of these variables did not affect our results in any substantial way.

Stock Market’s Reaction to Unemployment News 653

many regressions, especially daily stock returns, autocorrelation is not an im- portant factor, thus only White t-statistics are reported.

Forecasts for the change in the unemployment rate from month t − 1 to month t were constructed by first estimating equation (1) using monthly observations up to month t − 1. Adding back the unemployment rate at month t − 1 to this forecast gives us the predicted unemployment rate in month t.

Actually, equation (1) was estimated in three different ways. The first es- timation method (Method 1) is the best, in the sense of achieving the small- est out-of-sample forecast errors. In this case, we used final release numbers for unemployment and industrial production and we also included a dummy variable, which took on the value 1.0 during contractions and 0.0 during ex- pansions. This procedure could be criticized on two grounds. First, it takes into account the information conveyed by the state of the economy. However, it can be argued that agents do not necessarily know the state of the economy at the time a forecast is made, since the NBER’s announcement of an official turning point typically comes several months after the turning point date.4

To address this criticism, our second estimation method (Method 2) omits the business cycle dummy variable that allows the intercept for contractions to be different from that for expansions. This results in a small but significant bias in the forecasts: The average forecast error during expansions and con- tractions for the model is different from zero. Such a bias does not occur with Method 1.

A second criticism of our forecasting procedure relates to the use of final re- lease data both for the unemployment rate and the IIP. Since the final release numbers come out about 3 months after the initial release, forecasts made in this way could not have been made in real time. In view of this criticism, our third forecasting method (Method 3) uses final release figures for the unem- ployment rate and the IIP, but only employs data available up to 1 year before the estimation date. Then we employ the estimated parameters and the initial release numbers of the unemployment rate data and originally published and subsequently revised IIP to construct our estimate of the unemployment sur- prise. With this very conservative method, we can be sure we are only using information that was available to investors at the time the forecast was made. This method also has a small but significant bias in the forecasts.

All the three forecasting methods have the expected properties: Method 1 re- sults in smaller forecast errors than Method 2, and Method 2 results in smaller errors than Method 3. We feel that these three estimating methods span the space of reasonable real-time unemployment forecasts. That is, estimates made using Method 1 are undoubtedly better than market participants could actu- ally have made, and estimates made using Method 3 are clearly worse. What is most important for present purposes is that none of the results are particularly sensitive to the choice of estimation methods. We therefore present the results only for Method 3, for which the results are the weakest.

4 For example, the NBER Business Cycle Dating Committee, in its October 16, 2003 meeting, determined that a trough in the business activity occurred in the U.S. economy in November 2001.

654 The Journal of Finance

Table I Properties of the Forecasted Unemployment Rate

Forecasts are made using the following model described in equation (1) in the text:

DUMPt = b0 + b1 · IPGRATEt−1 + b2 · IPGRATEt−1 + b2 · IPGRATEt−1 + b4 · DUMPt−1 + b5 · DTBt + b6 · DBAt + et ,

where DUMPt denotes the change of unemployment rate from month t − 1 to t; IPGRATEt denotes the growth rate in industrial production; DTB3t denotes the change in the 3-month T-bill rate; and DBAt denotes the change in the yield spread between low and high grade bond yields. Appendix A describes the data in detail. Appendix B describes the forecasting methods used. We report the mean and the standard error for the mean (in parentheses) for the change in unemployment rate, DUMPt (in percent, annualized), its forecasted value, DUMPFt, and the forecast error, ERRUMPt = DUMPt − DUMPFt for the period June 1972 to December 2000 for forecasting Method 3 as described in Appendix B.

Unemployment Rate DUMP DUMPF ERRUMP

Whole sample 5.952 −0.0052 0.0220∗ −0.0270∗ (0.0713) (0.0100) (0.0066) (0.0090)

Contractions 6.819 0.222∗ 0.1612∗ 0.0605∗ (0.2397) (0.0306) (0.0291) (0.0293)

Expansions 5.832 −0.0399∗ 0.00046 −0.0405∗ (0.0723) (0.0090) (0.0051) (0.0091)

∗Indicates significance at the 5% level.

A.2. Properties of Unemployment News

To understand the properties of the forecasts and forecast errors during ex- pansions and contractions, we classified every sample month as an expansion or contraction month, using the NBER’s reference dating. The properties of the unemployment rate forecasts for Method 3 are in Table I.5 During the 343 monthly forecasts examined under Method 3, covering the period June 1972 to December 2000, the U.S. economy was in an expansion during 297 months and in a contraction during 46 months. There were four contractions and five expansions. The average duration of a contraction was 12 months and the av- erage duration of an expansion was 61 months. Unemployment was higher at 6.82% during contractions and lower at 5.83% during expansions. On average, the unemployment rate increased by 0.22% per month during contractions and declined by 0.04% per month during expansions. The forecasted changes in un- employment rates are smaller in expansions than in contractions. There is a small but statistically significant bias in the forecasts made using Model 3—the forecasts are biased downward during contractions and upward during expan- sions. The average forecast error was 6 basis points during contractions and −4 basis points during expansions (Table I).

5 Our data set actually begins in January 1957 (see Appendix A), but the first 5 years of data are used up in obtaining the initial forecasts. The initial release data begin in January 1972, thus we have a smaller sample size for Method 3 compared to Methods 1 and 2. Our findings are similar for all three methods, but we report the results for Method 3 only.