When Does Investor Sentiment Predict Stock Returns?

San-Lin Chung , Chi-Hsiou Hung , and Chung-Ying Yeh

Professor of Finance, National Taiwan University, Department of Finance, 85, Section 4, Roosevelt Road, Taipei 106, Taiwan. Tel: 886-2-3366-1084. Email: [email protected]. Hung is with Durham Business School, Durham University, Mill Hill Lane, Durham, DH1 3LB. Tel:44-191-334-5498, E-mail: [email protected] Yeh is with Department of Finance, Tunghai University, 181, Section 3, Taichung-Kan Road, Taichung, Taiwan. Tel:886-4-2359-0121 ext.35819. Email: [email protected]

Electronic copy available at: http://ssrn.com/abstract=1570487

When Does Investor Sentiment Predict Stock Returns?

Abstract We examine the predictive eect of sentiment on the cross-section of stock returns across dierent economic states. The degree of mispricing and the subsequent price correction can be dierent between economic expansion and recession because of the limits of arbitrage and short sale constraints. The predictive ability of sentiment is asymmetric between dierent states of the economy. We implement a multivariate Markov-switching model to characterize the economic states. Conditional on the identied economic states, we use the lagged sentiment proxy to forecast the portfolio returns related to small stocks, non-earning stocks, growth stocks, and non-dividend-paying stocks. We nd that only in the expansion state does sentiment performs both in-sample and out-of-sample predictive power on these categories of stocks. When an expansion state has high sentiment, these categories of stocks earn relatively low subsequent returns. The predictive ability of sentiment can not be attributed to time-variation in the market beta driven by investor sentiment. Keywords: Investor Sentiment; Stock Returns; Return Predictability; Markov-Switching Vector Autoregressive Model; Bootstrap JEL classication code: E32; G11; G12; G14

Electronic copy available at: http://ssrn.com/abstract=1570487

Introduction

The behavioral theories posit that investors may form erroneous stochastic beliefs, either with excessive optimism or pessimism, and therefore incorrectly assess asset values. This eect can cause asset prices to deviate from their intrinsic values (De Long, Shleifer, Summers and Waldman (1990), Lee, Shleifer and Thaler (1991), Swaminathan (1996), and Kumar and Lee (2006)). The mispricing is eventually corrected as the economic fundamentals reveal and sentiment wanes and the subsequent pricing correction results in the negative correlation between sentiment and future stock returns. As a consequence, investor sentiment exhibits predictive power for stock returns. Lemmon and Portniaguina (2006), for example, provide evidence that sentiment forecasts the returns on small size stocks. The extent of overpricing caused by sentiment, however, can be stronger relative to underpricing because of the limits of arbitrage and short sale constraints. De Long, et al. (1990) show that arbitrage can become ineective because arbitragers do not only bear fundamental risk but also face the noise trader risk that in the near future investors optimism could become more extreme and stock prices increase even more signicantly. Shleifer and Vishny (1997) show that the capital for arbitrage may be withdrawn from institutional arbitrageors when it is most needed. More importantly, short sale constraints hold negative opinions o the market and thus allow substantial overpricing. The constraints for taking short positions include the risks, costs, legal and institutional restrictions and the need of sucient stock supply from investors who are willing to lend.1 Nagel (2005), for example, nds that most of sophisticated professional investors never sell short and therefore cannot trade against overpricing. In contrast, executing buying trading is straightforward when pessimistic investors depress stocks prices to be lower than the fundamental values. For example, the longonly institutional investors such as mutual funds can increase holdings of underpriced
See Jones and Lamont (2002), Hong and Stein (2003), Ofek, Richardson and Whitelaw (2004), and Chang, Cheng and Yu (2007) for evidence and detailed discussions.
1

Electronic copy available at: http://ssrn.com/abstract=1570487

stocks. Furthermore, during periods of drastic price declines in the market legislators often forbid short sales (Lamont (2005)) and therefore may prevent prices of stocks that have substantially lost their values from overshooting to the downside. The sentiment-driven mispricing is thus asymmetric and is greater in optimistic periods and less in pessimistic periods. The subsequent corrections for mispricing are therefore more signicant following periods of high sentiment than periods of low sentiment. Consequently, investor sentiment exhibits asymmetric patterns in the predictive ability of stock returns across dierent states of the economy. In this paper, we explore the asymmetric predictive ability of sentiment on the crosssection of stock returns in times of optimistic and pessimistic market conditions. Our goal is to elaborate on the asymmetry in the relation of sentiment and the subsequent returns of stocks during bull and bear periods. Recent research suggests that the relation between sentiment and the valuation in the aggregate market level is asymmetric (Brown and Cli (2005)) and that the pattern of sentiment in predicting returns varies with stock characteristics (Baker and Wurgler (2006, 2007)). The question has never been investigated in the literature, however, concerning the asymmetry in the predictive eects of investor sentiment on the cross-section of stocks returns during dierent general economic and market environments. The motivation of the paper stems from the empirical evidence and the prediction of the behavioral models that overpricing happens in expansion and underpricing in recession. Daniel, Hirshleifer, and Subrahmanyam (1998) model investors behavior and demonstrate that a string of good news leads to overpricing and a string of bad news leads to underpricing. Gervais and Odean (2001) demonstrate that agents evaluate their own abilities with a bias in that they attribute success too heavily to superior ability rather than luck. Their model implies that, since most investors are long in stocks, in times when the market gains, the aggregate overcondence will be higher, leading to more aggressive trading. Moreover, these eects may rise late in a bull market which in turn, increase investment capital and therefore push prices even higher. Brown and Cli (2005)

shows that sentiment increases with the market price and the build-up of optimism leads to extended periods of overvaluation of assets and that investors become pessimistic as the market price falls, leading to underpricing. Lamont and Thaler (2003) identify an example that violates the law of one price during the rapid rise of technology stocks prices. Ofek and Richardson (2003) provide evidence against market eciency during the booming market period in the late 1990s. Our analysis segregates the market condition into bull and bear regimes. We use a two-state Markov-switching model of the stock market returns with a time-varying state transition matrix to capture the unobservable dynamics in the change in the economy regime as in Perez-Quiros and Timmermann (2000) and Ozoguz (2009). The regimeswitching model has been widely used to characterize the evolution of regime shifts related to business cycles. In the process of identifying regimes we rst remove the potential impact on regime shifts caused by corresponding shifts in investor sentiment. Specically, we orthogonalize the monthly returns of the market, SMB, and HML portfolios from the variation in sentiment and then estimate the parameters of the Markov-switching model. We identify two economic regimes over the 40-year period from January 1966 to December 2005. Regime 1 is a high-volatility, recession state that captures most of the business cycle contractions and episodes of upheavals in the stock market. Regime 2 is a low-volatility, expansion state that covers most bull markets since the 1960s. The time series of the state probabilities estimated from the Markov-switching model corresponds to the evolution of investor sentiment. We nd that investor sentiment tends to increase when the economy is in the expansion regime, and decrease when the economy is in recession. Furthermore, when investor sentiment reached high levels, the economy usually has already shifted into regime 1; when sentiment remained low, the economy usually has already shifted into regime 2. Next, we control for the eect of regime shifts to examine the predictive eect of sentiment on the cross-section of stock returns. Specically, we estimate predictive regressions using the regime-sorted data.2 The predictive regressions regard the state of the regime
2

In our empirical analysis, to completely control the eect of regime shifts, we eliminate the observa-

as an exogenous input. We test whether the state of the regime aects the slope coefcient on sentiment in the model. We employ the lagged investor sentiment, which we proxy by the orthogonalized sentiment index of Baker and Wurgler (2006) to predict the equally-weighted monthly returns of the long-short portfolios that are long in stocks with high values of a characteristic and short stocks with low characteristic values.3 We use portfolios formed on size, book-to-market, dividend yield, and earnings-to-price ratio. We include control variables of the Fama-Frenchs (1993) size and value factors, the momentum factor and the liquidity factor of Pastor and Stambough (2003). In order to enhance the test power we use the bootstrapped p-values for the empirically computed t-statistic based on the procedure of Kosowski, Timmermann, Wermers and White (2006). The main result of the paper is that only during economic expansions sentiment does show a signicant and robust predictive power for the subsequent stock returns. Moreover, the asymmetry in the predictive eect of sentiment over economic states has cross-sectional impacts on stock returns and is strong for stocks whose valuations are highly subjective and are strongly aected by uctuations in the propensity of speculation. In particular, when the economy is in expansion, higher sentiment is associated with lower subsequent returns of small size, growth, non-earning and non-dividend-paying stocks because their prices tend to be more overvalued when sentiment is high. These stocks also tend to have higher idiosyncratic risk and lower liquidity, making them particularly risky and costly to arbitrage (Baker and Wurgler (2006, 2007)). In contrast, larger stocks who have valuable tangible assets and collaterals, high liquidity and low idiosyncratic volatility are less aected by sentiment. De Long, et al. (1990) also discuss that assets of long duration that promise fundamentally uncertain as opposed to immediate and certain risky cash payouts may thus be subject to an especially great amount of noise trader risk. We further use the consumer condence provided by the Survey Research Center of
tions at the turning points, where regime actually switches from one state to another. 3 Baker and Wurglers hypothesis predicts that large rms will be less aected by sentiment. Thus, value-weighted returns may obscure the predictive eect of sentiment.

the University of Michigan to proxy for investor sentiment in our empirical investigations. Qiu and Welch (2004) and Lemmon and Portniaguina (2006) document that consumer condence is an adequate proxy for investor sentiment and is able to predict the subsequent portfolio returns for small stocks. We examine the out-of-sample predictive eect of sentiment across the states of regimes using the test of Clark and West (2007). Overall results show that, only in the expansion state (regime 2), sentiment performs both in-sample and out-of-sample predictive power for the subsequent returns of small stocks, growth stocks, non-earning stocks and non-dividend-paying stocks. We also conduct the predictive regression with regime dummy variables. This predictive regression includes the regime dummies independently in addition to making sentiment loadings conditional on dummies. The results conrm our main ndings. Finally, in order to detect the source of the predictive power of sentiment, we apply a conditional beta model to examine whether the predictive eect of sentiment is due to time-variation in systematic risk or mispricing. We nd that the predictability patterns associated with sentiment reect mispricing and the subsequent correction, rather than time-variation in the market beta, even after conditioning on the state of the economy. The remainder of this paper is organized as follows. Section 2 describes the multivariate Markov-switching model and characterizes market regimes. Section 3 gives details of the data and the empirical tests. Section 4 conducts two robustness checks and reports results. Section 5 discusses the source of asymmetries and Section 6 concludes.

2
2.1

Characterizing Economic Regimes

A Multivariate Markov-Switching Model for Stock Returns

We estimate a Markov-switching model with time-varying regime transition probabilities to characterize economic regimes. The state of the regime switches at random times but is driven by a latent regime variable following a Markov chain that is assumed to change over time. Since recent research document that size and value premiums vary with the state of economic regime (Perez-Quiros and Timmermann (2000), Cooper, Gutierrez, and

Hameed (2004), and Gulen, Xing, and Zhang (2008), among others), we characterize economic regimes in the joint process of portfolio returns on the market, size and value factors of Fama and French (1993) as in Guidolin and Timmermann (2008). These factors are the monthly returns on the CRSP value weighted market index in excess of the Tbill rate and the monthly returns on the SMB and HML factors. The SMB factor is the dierence in average returns on the small-size stock portfolios and the big-size stock portfolios. The HML factor is the dierence between the average returns on the high book-to-market portfolios and the average returns on the low book-to-market portfolios. Including the size and value factors also has the advantage of utilizing more information to characterize regimes than that contained in the excess market return, which reduces the noise of the smoothed probabilities. We remove the sentiment variation from the factor portfolio returns prior to the estimation of the Markov-switching model. Specically, we regress the factor portfolio returns on the sentiment proxythe Baker and Wurglers (2006) orthogonalized sentiment index.4 The regressions residuals, labelled with a superscript , are the factor portfolio returns orthogonalized to the sentiment variation. The sample of orthogonalized factor portfolios covers the 480-month sample period from January 1966 to December 2005, which is dictated by the availability of the sentiment index. The mean returns of the unorthogonalized market, SMB, and HML portfolios are, respectively, around 0.4%, 0.3%, and 0.4% per month with volatility of 4.5%, 3.3%, and 3.0% per month. The portfolio returns are all skewed and leptokurtic.5 The orthogonalized factors retain the appealing properties of the unorthogonalized ones. We model the joint distribution of the vector of orthogonalized returns of the 3 factor portfolios, r t , as a multivariate Markov-switching process driven by a common discrete
Baker and Wurgler (2006) regress each of the six raw proxies on macroeconomic variables to remove the eects from business cycle variation prior to conducting the principal components analysis. As a result, this sentiment index is orthogonalized to business cycle uctuations. The descriptions of the sentiment data can refer to section 3.1. 5 We thank Kenneth French for making the data available at the web-page: http://mba.tuck.dartmouth.edu/pages/faculty/ken.french/.
4

regime variable st which takes two integer values {1, 2} as: r t = st + st Xt1 + t , (1)

where Xt1 is the vector of publicly available information used to predict stock returns, st is the 3 1 vector of the regime-dependent intercepts, st is the 3 3 matrix of the regime-dependent coecients. The vector of return innovations t N (0, st ) is assumed to follow a multivariate normal distribution with zero means and a regimedependent variance-covariance matrix st . The discrete regime variable st is assumed to follow a 2state rst-order Markov chain governed by a 2 2 transition probability matrix with time-varying elements in which p11,t = Pr(st = 1|st1 = 1, CLIt1 ) = N (a1 + b1 CLIt1 ), p22,t = Pr(st = 2|st1 = 2, CLIt1 ) = N (a2 + b2 CLIt1 ), where CLIt1 is the one-month lagged value of the change in log composite leading indicator6 and N () is the cumulative density function of a standard normal variable. The regime variable st , as a latent variable, can be statistically inferred by the realized observations. The possibilities of the regimes at each time point can be characterized by ltered probabilities Pr(st |Yt ) and smoothed probabilities Pr(st |YT ), where Yt is the information set at time t and YT is the complete information set. The vector of Xt1 comprises dividend yield, default premium, and the short-term interest rate, following Perez-Quiros and Timmermann (2000) and Ozoguz (2009). The dividend yield is dened as the dividends on the valued-weighted CRSP index over past 12 months. We construct the dividend payout series using the valued-weighted return including dividends, and the price index series associated with the value-weighted return excluding dividends. The dividend series is the sum of dividend payout over past 12 months. The default premium is the yield spread between Baa and Aaa corporate bonds. The short-term interest rate is dened as the 90-day bill rate. We obtain the bond yield data from the web-page of Federal Reserve at St. Louis.
6

(2)

The composite leading indicator data can be downloaded from the OECDs web-page.

2.2

Empirical Results of the Multivariate Markov-Switching Model

Denote the set of parameters by = (1 , 2 , 1 , 2 , 1 , 2 , a1 , a2 , b1 , b2 ). Because the size of the parameters will cause severe problems in numerical optimization, we are not able to estimate the Markov-switching model of (1) and (2) by maximum likelihood directly. We adopt a pragmatic approach instead. We divide the set of parameters into two subsets: 1 = (1 , 2 , 1 , 2 , 1 , 2 ) and 2 = (a1 , a2 , b1 , b2 ). Given an initial value of 1 , we use the maximum likelihood estimation to obtain the estimates of 2 . We then take these estimates as inputs and use the EM algorithm of Hamilton (1990) to estimate 1 . We repeat these procedures iteratively until achieving convergence. We compute the standard deviations of the parameter estimates following the standard convention of the maximum likelihood approach.7 Table 1 reports the results from the estimation of the two-regime Markov-switching model. The decision criterion for inferring the state of the regime at each time points is that the regime has smoothed probability above 0.7.8 We identify regime 1 as a highvolatility recession or bearish state and regime 2 as a low-volatility expansion or bullish state. Panel A reports the parameter estimates in (1). The coecient estimates of the excess market return on the default premium are highly signicant and positive in both regimes. The coecient estimate of SMB on the default premium is signicantly positive in regime 1. The magnitude of the coecient estimates on the default premium in the bearish state is larger than that in the bullish state, suggesting that the default premium is more important during the economic recessions or bearish markets and is particular relevant to the size premium. These ndings are consistent with Perez-Quiros and Timmermann (2000). The coecient estimates of the excess market return on the lagged interest rate are statistically signicant and negative in both regimes. For the SMB factor portfolio, the coecient on the lagged interest rate is signicantly negative in regime 2.
The complete estimation results are available upon request. According to this criterion, there are 114 points classied as regime 1 and 338 points classied as regime 2. In addition, there are 28 points classied as regime turning points dened by the beginning and end of each regime.
8 7

(0)

(1)

(1)

Panel B of Table 1 presents the estimates of volatilities and correlations in the diagonals and the o-diagonals of the correlation matrices, respectively, showing that in regime 1 the monthly volatility of the excess market return is 6.1%. The correlation between the market portfolio and the SMB portfolio is positive, while the HML portfolio is negatively correlated with both the market portfolio and the SMB portfolio, suggesting that HML may serve as a hedge against the market portfolio. Panel C displays the parameter estimates of a1 , a2 , b1 and b2 associated with the transition probabilities in (2) and shows that the coecient b1 on the change in the composite leading indicator is signicantly negative, indicating that an increase in the leading indicator decreases the probability of staying in regime 1. The upper panel of Figure 1 which plots the historical patterns of the smoothed probabilities of regime 1 shows that this regime captures most of the business cycle contractions (the shaded areas) and episodes of sharp declines in stock prices since the 1960s. These include the 1969 - 1970 recession, the oil crisis in 1973, the stock market crashes in 1973 1974 and 1987, the Gulf War in the early 1990s, the default of the Russian sovereign bonds and the near collapse of Long Term Capital Management in 1998, the Internet bubble burst and corporate malfeasance in the beginning of the 2000s. The rst column of Panel A of Table 2 shows that the correlation coecient between the NBER recession indicator and the smoothed probability of regime 1 is about 0.3. The rst column in Panel B of Table 2 shows that regressing the smoothed probability of regime 1 on the NBER recession indicator gives a positive coecient of 0.33 which is signicant at the 1% level. Regime 2 covers, as displayed in the lower panel of Figure 1, most of the bull markets with growing stock prices since the mid-1960s and the run-ups in the 1980s and 1990s. Panel B of Table 1 reports that the monthly volatility of the excess market return is 3.6% and that the volatility levels of the factor portfolios of SMB and HML in regime 2 are approximately a half of those in regime 1. Panel C of Table 1 shows that the estimated coecient b2 of the probability of staying in regime 2 on the change in the composite leading indicator is signicantly positive, suggesting that an increase in the

leading indicator brings up the probability of staying in regime 2. The second column of Panel A of Table 2 shows that the correlation coecient between the NBER recession indicator and the smoothed probability of regime 2 is negative. The second column in Panel B of Table 2 shows that the regression coecient of the smoothed probability of regime 2 on the NBER recession indicator is signicantly negative. Finally, we sort industrial production growth rates based on the identied regimes and nd that the average monthly growth rate of industrial production is 0.05% in regime 1 and 0.3% in regime 2. The volatility of the industrial production growth rate in regime 1 is higher than that in regime 2.9 These evidence, taken together, conrms the economic interpretation that the regimes characterized by the Markov-switching model are indeed associated with underlying economic fundamentals.

3
3.1

Predictive Regressions for Long-Short Portfolios

Data and Sample

To implement the predictability tests, we use the monthly equally weighted returns on portfolios formed on rm characteristics of: (i) size (ME), (ii) book-to-market (BE/ME), (iii) dividend yield (D/P), and (iv) earnings/price (E/P).10 All portfolios are constructed at the end of each June. In June of year t, all NYSE stocks are sorted by (i) ME, (ii) BE/ME, (iii) D/P, and (iv) E/P, respectively, to determine the decile breakpoints for each rm characteristic. ME is the June market equity of year t. BE/ME is book equity at the last scal year end of the prior calendar year t 1 divided by market equity at the end of December of the prior year t 1. D/P is the total dividends paid from July of the prior year t 1 to June of the present year t divided by market equity at June of the present year t. E/P is earnings before extraordinary at the last scal year end of the prior calendar year t 1 divided by market equity at the end of December of the prior year t 1. All NYSE, AMEX, and NASDAQ stocks are allocated based on the NYSEs breakpoints of each rm characteristic, forming the portfolios for July of year t to June
9 10

The results about the monthly growth rate of industrial production are available upon request. We obtained the data from Kenneth Frenchs data library.

10

of year t + 1. We use 10 deciles for size and book-to-market portfolios. For dividend yield and earnings/price, there are 11 portfolios in which portfolio returns for non-earning stocks or non-dividend-paying stocks are calculated separately (denoted as 0). Table 3 reports the summary statistics for the monthly portfolio returns in percentage for the period from January 1966 to December 2005. The results show that small stocks and value stocks have higher mean returns than large stocks and growth stocks. The size and value premiums in terms of the mean return are approximately 0.60% and 1.28%, respectively. In contrast, the mean return patterns for the portfolios formed on dividend yield and earnings/price do not reveal the cross-sectional eects. The skewness of the middle decile groups is negative and smaller than the skewness of the low- and high-decile groups. The excess kurtosis is positive in all cases. The summary results strongly reject the normality property for the cross-section of stock returns, which is empirically related to the regime-switching feature in the stock returns and this could invalidate the conventional t-statistic in the predictive regressions. We use the Baker and Wurglers (2006) orthogonalized, monthly sentiment index as the measure of investor sentiment which has been shown to exhibit predictive ability for the cross-section of stock returns.11 The top panel of Figure 2 plots the historical pattern of the orthogonal sentiment index over the period from January 1966 to December 2005. The sentiment index rapidly increased from October 1967 and reached its climax by May 1969, corresponding to the period of being in regime 2 as indicated by the smoothed probabilities displayed in the lower panel of Figure 1. The economy then changed into regime 1 and the sentiment index sharply declined until may 1971. Investor sentiment remained weak with episodes of small rises, following the shift into regime 2 for a short period from January 1971 to March 1973. The regime fell back into regime 1 again in July
11 The sentiment index of Baker and Wurgler (2006) is the rst principal component of six major proxies for investor sentiment including the closed-end fund discount, NYSE share turnover, the number of IPOs, the average rst-day returns on IPOs, the equity share in new issues, and the dividend premium. Both annual and monthly sentiment data are available from Jerey Wurglers website at: http://pages.stern.nyu.edu/ jwurgler/.

11

1973 until February 1976, during which period the sentiment index was mostly negative, reecting the outbreak of the oil crises and the impact of the stock market crash, and reached an all time low by August 1976. Most of the time from April 1976 to June 1998 the economy stayed in regime 2, while the sentiment index started to rise again from September 1976 and showed a spike in February 1981 and another in November 1983. It then turned into a decreasing pattern and came to a low level by September 1990 a month after the rst Gulf War started. After hovering for a year at low sentiment levels in 1998 surrounding the Russian Default investor sentiment reached another high by March 2000 and declined after the burst of the internet bubble, but climbed back up to an even higher level by February 2001 and became negative afterwards until 2004. Two observations are in order from a close scrutiny of the time-series data. First, investor sentiment tends to increase when the smoothed probabilities indicate the state being in regime 2, and decrease when the smoothed probabilities indicate the state being in regime 1. Second, the change in the sentiment index tends to lag behind the change in regimes by some months. Specically, the economy usually had already shifted into regime 1 when investor sentiment reached high levels, and the economy usually had already shifted into regime 2 while investor sentiment remained at low levels. Table 2 shows that the smoothed probability of regime 1 and the NBER recession indicator are positively correlated with the lagged sentiment index. The regression coecient of the smoothed probability of regime 1 on the lagged sentiment is signicantly positive.

3.2

The Regression Models

Following Baker and Wurgler (2006), we use the lagged sentiment proxy to predict the equally-weighted returns for long-short portfolios that are long on stocks with high values of a characteristic and short on stocks with low values. We run the predictive regressions using the regime-sorted data to control for the eects of regime shifts, r(i,k2 ),t r(i,k1 ),t = i + i,1 SENTIMENT t1 +
i,t ,

(3)

12

where r(i,k2 ),t r(i,k1 ),t is the monthly return on a long-short portfolio that are long on portfolio k2 and short on portfolio k1 with rm characteristic i at time t, and k1 , k2 { 0, 1, 2, . . . , 10}. 1, 2, . . . , and 10 indicate the portfolios in the 1st (the smallest), 2nd, . . . , and 10th deciles, respectively. 0 represents the portfolios of non-dividendpaying stocks or non-earnings stocks. SENTIMENT t1 is the Baker and Wurglers (2006) orthogonalized sentiment index at time t 1 which removes the eects of macroeconomic uctuations. Baker and Wurgler (2006) documented that, when the beginning-of-period sentiment is high, the subsequent portfolio returns of small stocks, growth stocks, non-dividend-paying stocks and non-earning stocks are lower than those of large size, value, dividend-paying and with-earnings stocks. Thus, sentiment forecasts the returns on a long-short portfolio formed on these characteristics and that the sign of the coecient on i,1 in the regression model is expected to be positive. We also distinguish predictability eects associated with rm characteristics using the multivariate regression, r(i,k2 ),t r(i,k1 ),t = i + i,1 SENTIMENT t1 + i, RMRFt +i,2 SMBt + i,3 HMLt + i,4 UMDt + i,5 LIQt +
i,t ,

(4)

where RMRFt is the return on the value-weighted market in excess of the risk-free rate. SMBt and HMLt are the Fama-French factors. UMDt is the return on high-momentum stocks minus the return on low-momentum stocks, where the measure for momentum is based on the cumulative raw return for the 11 months from 12 through 2 months prior to the return observation.12 LIQt is the Pastor and Stambaughs (2003) liquidity factor.13 We include the liquidity factor to mitigate the concern that market liquidity is related to sentiment (see, e.g., Baker and Stein (2004) and Deuskar (2004)). When we perform predictive regressions, we eliminate the observations at the turning points where regime actually switches from one state to another to completely control for
We downloaded this momentum factor from the web-page of Kenneth French. We obtain this liquidity factor from the Fama-French Research Portfolios in Wharton Research Data Services (WRDS). The sample period for the liquidity factor is from Feb. 1962 to Dec. 2006.
13 12

13

the eect of regime shifts, which helps clearly identify the predictive ability of sentiment on the stock returns. We perform a bootstrap testing procedure for drawing robust inferences, following Kosowski, Timmermann, Wermers, and White (2006). The Appendix contains details of the bootstrap testing procedure. This approach is useful for many reasons: First, Table 3 shows that return distributions of portfolios sorted by certain stock characteristics such as size, protability or dividend yield, may not be normally distributed. Second, the nite sample distribution of the standard t-statistic may be shifted to the right, leading to an over-rejection under the null of no-predictability (see, for example, Hodrick (1992), and Ang and Bekaert (2007)). Finally, the correlation between the endogenous regressors and return innovations may result in biased estimates as shown in Stambaugh (1999).

3.3

Results

For the characteristics of size and book-to-market ratio, we consider the long-short portfolios, 10-1, that are long on the decile with the top characteristic and short on the decile with the lowest characteristic; and, for more detailed examinations, the long-short portfolios, 5-1, that are long on the middle characteristic decile and short on the lowest characteristic decile. For the characteristics of dividend yield and earnings-to-price ratio, we examine two long-short portfolios, 10- 0 and 5- 0, that are, respectively, long on the top characteristic decile and short on the decile of non-earning stocks or non-dividend-paying stocks, and long on the middle characteristic decile and short on the decile of non-earning stocks or non-dividend-paying stocks. Table 4 reports the estimates of the coecient on sentiment, i,1 , of the predictive regressions for the period from January 1966 to December 2005. The column All corresponds to the results without sorting observations by regime. The columns under Regime j , j = 1, 2, report the results using regime-sorted observations with regime = j . The bootstrapped p-values are in parentheses. The numbers of observations for the sample period in the two regimes inferred by the smoothed probabilities are 114 and 338, respectively. 14

The column All shows that the coecient estimates, i,1 , are positive and signicant for the long-short portfolios associated with size, earnings and dividend payment. The results show that when sentiment is high, returns on small stocks, non-earning stocks, and non-dividend-paying stocks are relatively low over the coming year. These patterns are little aected by controlling for RMRF, SMB, HML, UMD, and LIQ. However, the i,1 estimates on the portfolios associated with growth stocks become insignicant after controlling for RMRF, SMB, HML, UMD, and LIQ. These results are consistent with those of Baker and Wurgler (2006). When the economic state is controlled, the results in the columns Regime j , j = 1, 2 show that the predictability patterns associated with sentiment are regime-dependent. In particular, sentiment has strong and signicant predictive power in regime 2, i.e. a high sentiment results in relatively low subsequent returns on small stocks, growth stocks, non-earning stocks, and non-dividend-paying stocks. Moreover, the signicance of the predictive ability of sentiment is little aected by controlling for RMRF, SMB, HML, UMD, and LIQ in most cases. Under the other regime, however, sentiment mostly dose not exhibit a signicant predictive power. For example, the predictability patterns for non-earning stocks and non-dividend-paying stocks disappear after controlling for RMRF, SMB, HML, UMD, and LIQ under regime 1. For most parts, the magnitude of the coecient estimate i,1 in regime 2 is larger than that in the case of All. It suggests that controlling for the economic state has signicant inuence on the magnitude of the coecient on sentiment, even after controlling for RMRF, SMB, HML, UMD, and LIQ. Overall results show that sentiment has signicant predictive power on the subsequent returns for small stocks, growth stocks, non-earning stocks, and non-dividend-paying stocks only under regime 2.14
In unreported results (available upon request), we run predictive regressions based on the regimes identied by ltered probabilities. The results are consistent with those in Table 4 and do not change our conclusions.
14

15

3.4

Out-of-Sample Test Using the MSPE-adjusted Statistic of Clark-West (2007)

We perform an out-of-sample test to examine the robustness of the regime-dependent predictability patterns associated with sentiment. We test whether the predictive power of an unrestricted model (with the predictor) is better than that of a restricted model (without the predictor) in terms of prediction errors. Clark and West (2007) develop a test statistic using mean squared prediction errors (MSPE) to measure the prediction performance of models. This test statistically assesses the dierence between the MSPEs of the restricted and the unrestricted models. The unrestricted model is deemed to have better forecasting performance if its MSPE is smaller than that of the restricted model. The unrestricted model in our test of the predictive power of sentiment is (4) (as re-written below for clarity), r(i,k2 ),t r(i,k1 ),t = i + i,1 SENTIMENT t1 + i, RMRFt + i,2 SMBt +i,3 HMLt + i,4 UMDt + i,5 LIQt +
i,t ,

and the restricted model with a constraint of i,1 = 0, i.e. no predictive power of sentiment, is r(i,k2 ),t r(i,k1 ),t = i + i, RMRFt + i,2 SMBt + i,3 HMLt + i,4 UMDt + i,5 LIQt +
i,t .

(5)

For a sample of T observations including R in-sample (t = 1, . . . , R) and P out-of-sample (t = R + 1, . . . , R + P ) observations, we perform the estimation and prediction procedures for calculating the model prediction errors as follows. First, for each of the long-short portfolios, we estimate models (4) and (5) using all the in-sample data to obtain the sets
c of the estimated coecients denoted as u i and i for the unrestricted and restricted

models, respectively. Next, for month t = R + 1, . . . , R + P we use the estimated coecients and incorporate the out-of-sample realizations of the explanatory variables xu i,t
c (including SENTIMENT t1 , RMRFt , SMBt , HMLt , UMDt and LIQt ) and xi,t (without u u c c SENTIMENT i,t = u i,t = c t1 ) to compute the return predictions y i xi,t and y i xi,t for c the unrestricted and restricted models, respectively. The prediction errors, u i,t and i,t ,

16

for the unrestricted and restricted models, respectively, are the dierences between the realized and the model predicted returns of the long-short portfolio. The MSPE-adjusted statistic of Clark and West (2007) is: MSPE-adjusted = = P 1 where f (f t f ). Under the null hypothesis of no dierence in the model prediction errors, the MSPEadjusted statistic is zero. Clark and West (2007) demonstrate that the asymptotic distribution for the MSPE-adjusted statistic can be approximated by a standard normal distribution. Note that since it is a one-sided test, the critical values for 5% and 1% signicance level are 1.645 and 1.96, respectively. Table 5 reports the results of the out-of-sample test over the period from January 1966 to December 2005. The rst column shows that in most of the cases sentiment doesnt exhibit out-of-sample predictive power for small stocks, growth stocks, non-dividend-paying stocks, and non-earnings stocks. A seemingly strong in-sample result almost breaks down out-of-sample. Butler, Grullon, and Weston (2005) argue that the discrepancy between in-sample and out-of-sample results may be attributed to the instability of parameter estimates. As shown in Table 4, the coecient estimate for i,1 changes with the state of regime, suggesting the existence of parameter instability. Importantly, after controlling for the economic state, the results in the column of Regime 2 show that the MSPE-adjusted statistics are signicant in all cases but the longshort position in the book-to-market deciles, suggesting that, out-of-sample, sentiment does have a signicant predictive power. The results of regimes 1, by contrast, show that sentiment does not have signicant predictive power. The out-of-sample performance of sentiment after controlling for the state of the market regime is consistent with the insample results in Table 4. We nd that when the economy is in the expansion state with high sentiment, subsequent returns for small stocks, growth stocks, non-earning stocks, and non-dividend-paying stocks tend to be low. 17
R +P t=R+1

Pf , Vf

(6)

c2 u2 c u 2 f is the sample variance of f yi,t y i,t ) ], and V t , ft = i,t [i,t (

4
4.1

Robustness Checks
Predictive Regressions Using Consumer Condence

Qiu and Welch (2004), and Lemmon and Portniaguina (2006) document that consumer condence is an adequate proxy for investor sentiment and is able to predict the returns for small stocks. For robustness checks, we use monthly consumer condence index provided by the University of Michigan as a proxy for investor sentiment to test the predictive power of sentiment over the period between January 1978 and December 2007.15 Following Baker and Wurgler (2006) and Lemmon and Portniaguina (2006), we orthogonalize consumer sentiment from macroeconomic variables by regressing the consumer sentiment on the growth rate of the industrial production, a dummy variable for NBER recessions, the growth rate of personal consumption expenditures in durables, nondurables and services.16 We denote the regression residuals as ConSENTIMENT t for the orthogonalized consumer condence and plot the time-series observations in the bottom panel of Figure 2. The orthogonalized sentiment dropped to a record low level in 1979 and then moved higher gradually before going back down again to a new low by 1992. It increased afterwards over time to reach a record high by 2000 which is then followed by dramatic decreases until 2003, corresponding to the boom and bust of the internet bubble. The overall pattern resembles that in the upper panel which depicts the Baker and Wurgler orthogonalized sentiment.
We next replace SENTIMENT t1 in (4) by ConSENTIMENTt1 and run the predic-

tive regression, r(i,k2 ),t r(i,k1 ),t = i + i,1 ConSENTIMENT t1 + i, RMRFt +i,2 SMBt + i,3 HMLt + i,4 UMDt + i,5 LIQt +
i,t .

(7)

Table 6 reports the coecient estimates on sentiment i,1 . The column under regime 2
The survey for consumer sentiment started in 1947 on a quarterly basis for Febuary, May, August and November.The index value on the rst quarter of 1966 is 100. After 1978 the sentiment index was published on a monthly basis. 16 We use the data of personal consumption expenditures in durables, nondurables and services from BEA National Income Accounts.
15

18

shows that the orthogonalized sentiment of the consumer condence index exhibits signicant ability in predicting stock returns, while the results in regime 1 show no evidence of predictability. These results are in line with those reported in Table 4. The main dierence between the results in Tables 4 and 6 is that, in the rst column for the whole sample period, the Baker and Wurglers orthogonalized sentiment is capable of predicting stock returns for all portfolios while the orthogonalized consumer condence is only able to predict stock returns of the size portfolios. Overall, the results suggest that the predictive power of sentiment is regime-dependent. When the economy is in the expansion state, high sentiment is associated with low subsequent returns for small stocks, growth stocks, non-earning stocks, and non-dividend-paying stocks.

4.2

Predictive Regressions with Regime Dummies

We also examine the regime-dependency of the predictive power of investor sentiment using the regime dummy variables in the regression model. Specically, we run the predictive regression: r(i,k2 ),t r(i,k1 ),t = i, + i,1 D1 + i,2 D2 + (i,1 D1 + i,2 D2 )SENTIMENT t1 +i, RMRFt + i,2 SMBt + i,3 HMLt + i,4 UMDt + i,5 LIQt + (8) i,t ,

where Dj is the dummy variable of regime j , which equals 1 for regime = j and 0 otherwise. Note that the specication in (8) treats the state of regime as an exogenous variable. Table 7 reports the estimates of the coecients on the two regime dummies. We include the observations at the turning points in this regression. In regime 2 investor sentiment displays positive and highly signicant predictive power as indicated by the bootstrapped p-values for all the returns on the long-short portfolios. Importantly for all cases, the coecient estimates on the dummy variable for regime 1 are statistically insignicant, suggesting that investor sentiment loses its predictive ability for the crosssection of stock returns in this economic regime. These results are consistent with those presented in Table 4 and provide further evidence that the predictive power of investor sentiment is regime-dependent. 19

5
5.1

Sources of the Predictive Ability of Sentiment

Time-Variation with Sentiment in Systematic Risk

We allow the market beta to vary with investor sentiment in the predictive regression below to examine whether the predictability in the cross-section of stock returns is due to time-varying sensitivity in the market risk factor driven by investor sentiment,
r(i,k2 ),t r(i,k1 ),t = i + i,1 SENTIMENT t1 + (i, + i,1 SENTIMENTt1 )RMRFt

+i,2 SMBt + i,3 HMLt + i,4 UMDt + i,5 LIQt

i,t .

(9)

We test the null that the return predictability of investor sentiment is attributable to the time-varying market beta associated with investor sentiment, that is, the slope coecient i,1 in the specication (i, +i,1 SENTIMENT t1 ) is signicantly dierent from zero. Table 8 reports the results of the coecient estimates of i,1 and i,1 . The results in the rst column show that over the period from January 1966 to December 2005 the estimates of i,1 are statistically insignicant in all cases as indicated by the bootstrapped p-values. The estimates of i,1 , by contrast, are positive and highly signicant in all cases but the long-short positions on the book-to-market portfolios. The results for regime 2 are very similar. For most of the cases in regimes 1, the estimates of i,1 are statistically insignicant. These results, taken together, show that the sentiment-driven time-variation in the market beta does not capture the predictive ability of sentiment in the cross-section of stock returns.

Conclusion

In this study we examine the predictive eect of investor sentiment on the cross-section of stock returns across dierent economic states. We rst implement a multivariate Markovswitching model to characterize two economic regimes, the expansion state and the recession state. We then use the lagged sentiment proxy to forecast the portfolio returns related to small stocks, non-earning stocks, growth stocks and non-dividend-paying stocks in the identied economic states. We nd that the predictive power of investor sentiment 20

is regime-dependent. Only under the bullish regime investor sentiment performs both in-sample and out-of-sample predictive power on the portfolio returns related to small stocks, growth stocks, non-earning stocks and non-dividend-paying stocks. When the expansion state has high investor sentiment, these categories of stocks earn relatively low subsequent returns. Our ndings highlight the asymmetry in the predictive ability of sentiment on stock returns over economic states. The predictive ability of sentiment can not be attributed to time-variation in the market beta driven by investor sentiment.

21

Appendix: A Simple Bootstrap Testing Procedure

We use the predictive regression model (4) to illustrate the testing procedure. The implementation is as follows: Step 1: We run the following regression model for returns of the k th portfolio with rm characteristic i:
r(i,k),t = ( i,k) + (i,k), RMRFt + (i,k),1 SENTIMENTt1 + ( i,k),2 SMBt + (i,k),3 HMLt + (i,k),4 UMDt + (i,k),5 LIQt + (i,k),t ,

(10)

and save all OLS-estimated risk loadings { ( ( ( ( ( i,k) ,(i,k), , i,k),1 , i,k),2 , i,k),3 , i,k),4 , ( i,k),5 }, residuals {(i,k),t , t = T0 , . . . , Tn } for all stock portfolios k = 1, . . . , N , where

T0 and Tn are the dates of the rst and last observations. i,t = ((i,1),t , . . . , (i,N ),t ) . Step 2: Denote the cross-section of residuals at time t by
b We resample a sequence of the time indices sb T0 , . . . , sTn that are drawn randomly

from [T0 , . . . , Tn ], where b is the index for the bootstrap sample (for example, b = 1 means the resample number one). The sequence of the resampled residual vectors
b b = (b b , t = sb , . . . , sb }, where are given by { i,t i,t T0 Tn (i,1),t , . . . , (i,N ),t ) .

Step 3: The pseudo portfolio returns are constructed under the null hypothesis (i,k),1 = 0 by
b ( ( ( ( r( i,k),t = i,k) + (i,k), RMRFt + i,k),2 SMBt + i,k),3 HMLt + i,k),4 UMDt b + ( i,k),5 LIQt + (i,k),t ,

(11)

b for k = 1, . . . , N and t = sb T0 , . . . , sTn . So, there are N resampled time series of b b b stock portfolio returns in a resampling, {r( i,k),t , t = sT0 , . . . , sTn }, k = 1, . . . , N .

Step 4: We run the predictive regression model (4) using the resampled data
b generated in Step 3, and compute the corresponding t-statistic, t( i, 1 ), for i,1 .

Repeating Steps 2 and 3 for M times (the largest b is M), a bootstrap distribution
b of t-statistic of i,1 , {t( i, 1 ), b = 1, . . . , M }, under the null hypothesis (i,1 = 0) is

available. 22

 Step 5: We compute the p-value associated with t-statistic by comparing T t( i,1 ) b i,1 )] to obtain the p-value. Note that T is the to the quantiles of T [t( i, 1 ) t(
b total sample size between T0 and Tn , t( i,1 ) and t( i, 1 ) are the t-statistics computed

by real data and the b-th resampled data, respectively. The bootstrapped p-value may be dened as the probability in favor of the null hypothesis Prob b T [t( i, i,1 )] > T t( i,1 ) . 1 ) t(

The readers can refer to Sullivan, Timmermann, and White (1999) and Wang (2005) for the details of implementing bootstrap methods.

23

References
Ang, A., and G., Bekaert, 2007, Stock Return Predictability: Is it There, Review of Financial Studies, 20, 651-707. Baker, M., and J., Stein, 2004, Market Liquidity as a Sentiment Indicator, Journal of Financial Markets, 7, 271-299. Baker, M., and J., Wurgler, 2006, Investor Sentiment and the Cross-Section of Stock Returns, Journal of Finance, 61, 1645-1680. Baker, M., and J., Wurgler, 2007, Investor Sentiment in the Stock Market, Journal of Economic Prospective, 21, 129-152. Brown, G., and M., Cli, 2004, Investor Sentiment and the Near-Term Stock Market, Journal of Empirical Finance, 11, 1-27. Brown, G., and M., Cli, 2005, Investor Sentiment and Asset Valuation, Journal of Business , 78, 405-440. Butler, A., G., Grullon, and J., Weston, 2005, Can Managers Forecast Aggregate Market Returns, Journal of Finance, 60, 963-986. Chang, E., C., Cheng, and Y. Yu, 2007, Short-Sale Constraint and Price Discovery: Evidence from the Hong Kong Market, Journal of Finance, 62, 2097-2121. Clark, T., and K., West, 2007, Approximately Normal Tests for Equal Predictive Accuracy in Nested Models, Journal of Econometrics, 138, 291-311. Cooper, M., R., Gutierrez, and A., Hameed, 2004, Market States and Momentum, Journal of Finance, 59, 1345-1365. Daniel, K., D., Hirshleifer, and A., Subrahmanyam, 1998, Investor Psychology and Security Market Under- and Overreactions, Journal of Finance 53, 183986. David, A, 1997, Fluctuating Condence in Stock Markets: Implications for Returns and Volatility, Journal of Financial and Quantitative Analysis, 32, 427-462.

24

De Long, J., A. Shleifer, L., Summers, and R., Waldmann, 1990, Noise Trader Risk in Financial Markets, Journal of Political Economy, 98, 703-738. Fama, E., and K. French, 1993, Common Risk Factors in the Returns on Stocks and Bonds, Journal of Financial Economics , 33, 3-56. Gervais, S., and T., Odean, 2001, Learning to Be Overcondent, Review of Financial Studies, 14,1-28. Guidolin, M., and A., Timmermann, 2008, Size and Value Anomalies under Regime Shifts, Journal of Financial Econometrics, 6, 1-48. Gulen, H., Y., Xing, and L., Zhang, 2008, Value versus Growth: Time-Varying Expected Returns, Working Paper, University of Michigan. Hamilton, J., 1990, Analysis of Time Series Subject to Changes in Regimes, Journal of Econometrics, 45, 39-70. Hodrick, R., 1992, Dividend Yields and Expected Stock Returns: Alternative Procedures for Inference and Measurement, Review of Financial Studies, 5, 357-386. Hong, H., and J., Stein, 2003, Dierence of Opinion, Short-Sales Constraints, and Market Crashes, Review of Financial Studies, 16, 487-525. Jones, C., and O. Lamont, 2002, Short-Sale Constraints and Stock Returns, Journal of Financial Economics, 66, 207239. Kosowski, R., A., Timmermann, R., Wermers, and H., White, 2006, Can Mutual Fund Stars Really Pick Stocks? New Evidence from a Bootstrap Analysis, Journal of Finance, 61, 2551-2595. Kumar, A., and C., Lee, 2006, Retail Investor Sentiment and Return Comovements, Journal of Finance, 61, 2451-2486. Lamont, A., 2005, Short-Sale Constraints and Overpricing, in The Theory and Practice of Short Selling, edited by Frank J. Fabozzi.

25

Lamont, A., and R., Thaler, 2003, Can the Market Add and Subtract? Mispricing in Tech Stock Carve-Outs, Journal of Political Economy, 111, 227-268. Lee, C., A., Shleifer, and R., Thaler, 1991, Investor Sentiment and the closed-end fund puzzle, Journal of Finance, 46, 75-109. Lemmon, M., and E., Portniaguina, 2006, Consumer Condence and Asset Prices: Some Empirical Evidence, Review of Financial Studies, 19, 1499-1529. Nagel, S., 2005, Short Sales, Institutional Investors and the Cross-Section of Stock Returns, Journal of Financial Economics, 78, 277309. Ofek, E., and M., Richardson, 2003, DotCom Mania: The Rise and Fall of Internet Stock Prices, Journal of Finance, 58, 1113-1137. Ofek, E., M., Richardson, and R., Whitelaw, 2004, Limited Arbitrage and Short Sale Restrictions: Evidence from The Options Market, Journal of Financial Economics, 74, 305-342. Ozoguz, A., 2009, Good Times or Bad Times? Investors Uncertainty and Stock Returns, Review of Financial Studies, 22, 4377-4422. Pastor, L., and R., Stambaugh, 2003, Liquidity Risk and Expected Stock Returns, Journal of Political Economy, 111, 642-685. Pastor, L., and P., Veronesi, 2006, Was there a Nasdaq bubble in the late 1990s?, Journal of Financial Economics, 81, 61-100. Perez-Quiros, G., and A., Timmermann, 2000, Firm Size and Cyclical Variation in Stock Returns, Journal of Finance, 55, 1229-1262. Qiu, L., and I., Welch, 2004, Investor Sentiment Measures, NBER Working Paper 10794. Shleifer, A., and R., Vishny, 1997, The Limits of Arbitrage, Journal of Finance, 52, 35-55.

26

Stambaugh, R., 1999, Predictive regressions, Journal of Financial Economics, 54, 375421. Sullivan, R., A., Timmermann, and H., White, 1999, Data Snooping, Technical Trading Rule Performance and the Bootstrap, Journal of Finance, 54, 1647-1692. Summers, L., 1986, Does the Stock Market Rationally Reect Fundamental Values? Journal of Finance, 41,591-601. Swaminathan, B., 1996, Time-Varying Expected Small Firm Returns and Closed-End Fund Discounts, Review of Financial Studies, 9, 845-897. Timmermann, A., 2000, Moments of Markov Switching Model, Journal of Econometrics 96, 75-111. Veronesi, P., 1999, Stock Market Overreaction to Bad News in Good Times: A Rational Expectations Equilibrium Model,Review of Financial Studies, 12, 975-1007. Wang, K., 2005,Multifactor Evaluation of Style Rotation, Journal of Financial and Quantitative Analysis, 40, 349-372.

27

Figure 1: The Smoothed Probabilities of the Two-Sate Markov-Switching Model for the Orthogonalized Market, SMB, and HML Returns
This gure plots the smoothed probabilities for the two-state Markov-Switching model comprising monthly excess returns on the value-weighted market portfolio and return series on Fama and Frenchs (1993) SMB and HML portfolios that have been orthogonalized to sentiment variation. The sample period is from Jan. 1966 to Dec. 2005. Parameters estimates underlying these plots are reported in Table 1. Regime 1 is a high-volatility recession state that captures episodes of sharp declines in stock prices since 1960, such as the two oil shocks in the 1970s, the Gulf War in the beginning of the 1990s, the default of the Russian sovereign bonds and the collapse of Long Term Capital Management surrounding the crash in 1998, the internet bubble burst and corporate malfeasance in the beginning of the 2000s. Most of the periods classied as regime 1 occur in the NBER recessions (the shaded areas). Regime 2 is a low-volatility expansion state that covers most of the bull markets with growing stock prices since the mid-1960s, including the run-ups in the 1980s and 1990s.

0.5

0
1966-01 1967-12 1969-11 1971-10 1973-09 1975-08 1977-07 1979-06 1981-05 1983-04 1985-03 1987-02 1989-01 1990-12 1992-11 1994-10 1996-09 1998-08 2000-07 2002-06 2002-06 2004-05 2004-05

0.5

0
1966-01 1967-12 1969-11 1971-10 1973-09 1975-08 1977-07 1979-06 1981-05 1983-04 1985-03 1987-02 1989-01 1990-12 1992-11 1994-10 1996-09 1998-08 2000-07

28

Figure 2: The Historical Patterns of the Sentiment Proxies

The upper panel plots the Baker and Wurglers monthly orthogonalized sentiment index for the period from January 1966 to December 2005 and the lower panel displays an orthogonalized consumer sentiment measure for the period from Jan. 1978 to December 2005.
4 3 2 1 0 -1 -2 -3
1966-01 1968-02 1970-03 1972-04 1974-05 1976-06 1978-07 1980-08 1982-09 1984-10 1986-11 1988-12 1991-01 1993-02 1995-03 1997-04 1999-05 2001-06 2003-07 2003-07 2005-08 2005-08

SENTIMENT

35 25 15 5 -5 -15 -25 -35 -45

1966-01 1968-02 1970-03 1972-04 1974-05 1976-06 1978-07 1980-08 1982-09

ConSENTIMENT

1984-10

1986-11

1988-12

1991-01

1993-02

1995-03

1997-04

1999-05

29

2001-06

Table 1: Parameter Estimates of the Markov-Switching Model for the Orthogonalized Market, SMB, and HML Returns Parameters Market SMB HML Panel A: Mean Parameters Constant, Regime 1 -0.036 0.020 0.011 Constant, Regime 2 -0.006 -0.015 0.001 Default Premium(Deft1 ), Regime 1 42.371** 20.638** -6.245 Default Premium(Deft1 ), Regime 2 18.963** 9.431 -0.656 Interest rate(It1 ), Regime 1 -6.150** -3.822 -1.760 Interest rate(It1 ), Regime 2 -3.170** -2.771** 0.726 Dividend yield(Divt1 ), Regime 1 0.703 -0.270 0.250 Dividend yield(Divt1 ), Regime 2 0.073 0.664** -0.153 Panel B: Correlations/Volatilities Regime 1 Market 0.067** SMB 0.295** 0.055** HML -0.441** -0.345** 0.049** Regime 2 Market 0.036** SMB 0.297** 0.023** HML -0.406** -0.211* 0.022** Panel C: Transition Matrix Parameters Regime 1 Regime 2 Constant 0.461 1.457** Leading indicator(CLIt1 ) -69.413** 18.026**
Note: This table reports the results of the parameter estimates for the Markov-switching model: r t = st + st Xt1 + t , where st is the 3 1 intercept vector in regime st , st is the 3 3 regime-dependent coecients, Xt1 is the vector of dividend yield (Divt1 ), default premium (Deft1 ) and interest rate (It1 ), and t N (0, st ) is the 3 1 innovation vector of returns. st is an unobserved state variable driven by a two-state rst-order Markov chain governed by a 2 2 transition probability matrix with time-varying elements pii,t = Pr(st = i|st1 = i, CLIt1 ) = N (ai + bi CLIt1 ), i = 1, 2, where CLIt1 is the one-month lagged value of the change in log composite leading indicator and N () is the cumulative density function of a standard normal variable. The three series are excess returns on the values-weighted market portfolio and returns on Fama and Frenchs (1993) SMB and HML portfolios that have been orthogonalized to sentiment. The sample period is from Jan. 1966 to Dec. 2005. Values reported on the diagonals of the correlation matrices are volatilities. All estimates are monthly. and denote signicance at 5% and 1% level, respectively.

30

Table 2: Correlations and Regressions of Smoothed Probabilities with the NBER Indicator and Sentiment Panel A: Correlations Pr(st = 1|YT ) Pr(st = 2|YT ) SENTIMENT t1 SENTIMENTt1 0.171 -0.171 1 NBERt 0.289 -0.289 0.129 Panel B: Regressions with NBER Recession Index Pr(st = 1|YT ) Pr(st = 2|YT ) Constant 0.226** 0.272** 0.774** 0.728** (0.000) (0.000) (0.000) (0.000) NBERt 0.330** -0.330** (0.000) (0.000) SENTIMENTt1 0.067** -0.067** (0.000) (0.000)
Note: This table reports (i)the correlations of smoothed probabilities with the NBER recession index NBERt and lagged SENTIMENT ; (ii) regressions of smoothed probabilities on the NBER recession index Pr(st = j |YT ) = + 1 NBERt + t , j = 1, 2, and regressions of smoothed probabilities on lagged SENTIMENT Pr(st = j |YT ) = + 1 SENTIMENT t1 + t , j = 1, 2. The p-values are in parentheses. and denote signicance at 5% and 1% level, respectively.

31

Table 3: Summary Statistics for the Portfolio Returns Decile 0 1 2 3 4 5 6 7 8 9 10 Panel A: Portfolios Formed on Size Mean 1.49 1.13 1.16 1.12 1.14 1.08 1.14 1.05 1.02 0.89 Standard Deviation 6.84 6.56 6.33 6.11 5.89 5.58 5.40 5.27 4.85 4.67 Skewness 0.31 -0.07 -0.19 -0.33 -0.35 -0.38 -0.30 -0.30 -0.22 -0.22 Excess Kurtosis 2.70 2.74 2.30 2.17 2.43 1.85 2.13 1.58 1.43 1.74 Maximum 32.88 32.52 29.90 26.41 27.54 23.01 24.81 21.13 19.06 20.92 Minimum -27.68 -30.31 -28.87 -29.40 -28.12 -25.93 -25.90 -24.07 -22.09 -20.38 Panel B: Portfolios Formed on Book-to-Market Mean 0.66 1.02 1.14 1.26 1.32 1.45 1.54 1.57 1.71 1.94 Standard Deviation 7.73 6.59 6.19 5.86 5.53 5.36 5.19 5.25 5.58 6.38 Skewness 0.02 -0.22 -0.30 -0.32 -0.27 -0.18 -0.10 0.12 0.06 0.60 Excess Kurtosis 2.64 1.76 2.41 2.90 3.42 3.55 3.81 4.16 4.30 4.81 Maximum 41.56 25.59 27.50 29.18 29.87 29.70 29.71 32.02 33.08 40.43 Minimum -32.58 -29.45 -30.12 -29.09 -27.94 -26.11 -25.12 -24.30 -26.83 -25.59 Panel C: Portfolios Formed on Dividend Yield Mean 1.34 1.23 1.31 1.28 1.35 1.32 1.41 1.39 1.44 1.34 1.21 Standard Deviation 7.72 5.90 5.44 5.21 5.03 4.79 4.63 4.39 4.26 3.90 4.13 Skewness 0.19 -0.54 -0.53 -0.37 -0.51 -0.40 -0.49 -0.43 -0.24 0.09 1.08 Excess Kurtosis 2.18 2.23 3.11 3.59 3.65 3.70 3.77 4.38 4.48 4.31 8.28 Maximum 35.35 19.92 25.70 27.40 24.63 24.83 23.30 24.24 25.53 25.06 33.20 Minimum -30.29 -28.64 -26.95 -25.74 -26.24 -25.14 -24.80 -23.98 -22.09 -17.13 -12.40 Panel D: Portfolios Formed on Earnings/Price Mean 1.41 0.96 1.17 1.17 1.24 1.29 1.35 1.41 1.48 1.62 1.75 Standard Deviation 8.76 7.02 6.05 5.67 5.33 5.18 5.03 4.88 4.84 5.04 5.81 Skewness 0.59 -0.15 -0.35 -0.36 -0.38 -0.42 -0.33 -0.29 0.00 0.00 0.13 Excess Kurtosis 3.37 1.78 2.21 2.85 4.09 4.04 4.33 4.20 4.77 4.67 4.42 Maximum 46.37 27.30 25.97 27.78 29.62 27.93 28.82 26.99 30.74 30.74 36.32 Minimum -31.34 -31.58 -29.48 -28.05 -28.66 -28.05 -26.97 -25.58 -22.95 -23.61 -25.43
Note: This table reports the summary statistics of the monthly equal-weighted portfolio returns formed on size, book-to-market, dividend yield, and earnings/price. The data sample is downloaded from the web-page of Kenneth French. All portfolios are constructed at the end of each June. In June of year t, all NYSE stocks are sorted by (i) ME, (ii) BE/ME, (iii) D/P, and (iv) E/P, respectively, to determine the decile breakpoints for each rm characteristic. ME is the June market equity of year t. BE/ME is book equity at the last scal year end of the prior calendar year t 1 divided by market equity at the end of December of the prior year t 1. D/P is the total dividends paid from July of the prior year t 1 to June of the present year t divided by market equity at June of the present year t. E/P is earnings before extraordinary at the last scal year end of the prior calendar year t 1 divided by market equity at the end of December of the prior year t 1. All NYSE, AMEX, and NASDAQ stocks are allocated based on the NYSEs breakpoints of each rm characteristic, forming the portfolios for July of year t to June of year t + 1. The monthly equal-weighted returns are calculated for portfolios. For size and book-to-market, there are 10 portfolios corresponding to each decile. There are 11 portfolios for dividend yield and earnings/price in which 0 represents the portfolios for non-dividend-paying stocks and non-earning stocks. The values of portfolio returns are in terms of percentage. The sample period is from Jan. 1966 to Dec. 2005.

32

Table 4: Predictive Regressions for Long-Short Portfolio Returns Long-Short All Regime 1 Regime 2 Panel A: Portfolios Formed on Size Only Sentiment 10 - 1 0.75** (0.00) 0.38 (0.16) 1.07** (0.00) 5-1 0.36** (0.00) 0.21 (0.18) 0.47** (0.00) Controlling for RMRF, SMB, HML, UMD, and LIQ 10 - 1 0.27** (0.01) 0.04 (0.51) 0.30** (0.00) 5-1 0.24** (0.00) 0.20 (0.18) 0.19** (0.01) Panel B: Portfolios Formed on Book-to-Market Only Sentiment 10 - 1 0.24* (0.03) -0.12 (0.62) 0.50** (0.00) 5-1 0.47** (0.00) 0.38 (0.18) 0.57** (0.00) Controlling for RMRF, SMB, HML, UMD, and LIQ 10 - 1 -0.04 (0.77) -0.61 (1.00) 0.21** (0.00) 5-1 0.16 (0.07) -0.13 (0.78) 0.25** (0.00) Panel C: Portfolios Formed on Dividend Yield Only Sentiment 10 - 0 1.02** (0.00) 1.02** (0.00) 1.13** (0.00) 5 - 0 0.83** (0.00) 0.78* (0.03) 0.90** (0.00) Controlling for RMRF, SMB, HML, UMD, and LIQ 10 - 0 0.38** (0.00) 0.27 (0.19) 0.33** (0.00) 5 - 0 0.38** (0.00) 0.19 (0.31) 0.35** (0.00) (Continued)

33

Long-Short

10 510 5-

Table 4-Continued All Regime 1 Regime 2 Panel D: Portfolios Formed on Earnings/Price Only Sentiment 0 0.77** (0.00) 0.95** (0.00) 0.75** (0.00) 0 0.85** (0.00) 0.90* (0.02) 0.82** (0.00) Controlling for RMRF, SMB, HML, UMD, and LIQ 0 0.43** (0.00) 0.43 (0.10) 0.31** (0.00) 0 0.44** (0.00) 0.25 (0.29) 0.32** (0.00)

Note: This table contains the results about (i) regressions of long-short portfolio returns on lagged SENTIMENT , r(i,k2 ),t r(i,k1 ),t = i + i,1 SENTIMENT t1 +
i,t ,

and (ii) regressions of long-short portfolio returns on lagged SENTIMENT , the market factor (RMRF), the Fama-French factors (HML and SMB), the momentum factor (UMD), and the liquidity factor (LIQ), r(i,k2 ),t r(i,k1 ),t = i + i,1 SENTIMENT t1 + i, RMRFt + i,2 SMBt +i,3 HMLt + i,4 UMDt + i,5 LIQt + i,t , where r(i,k2 ),t r(i,k1 ),t is the long-short portfolio return that longs portfolio k2 and shorts portfolio k1 with rm characteristic i (including size, book-to-market, dividend yield, and earnings/price) at time t, and k1 , k2 { 0, 1, 2, . . . , 10}. 1, 2, . . . , and 10 indicate the portfolios in the 1st (the smallest), 2nd, . . . , and 10th deciles, respectively. 0 represents the portfolios for non-dividend-paying stocks or non-earning stocks. This table only reports the parameter estimates of i,1 . The sample period is from Jan. 1966 to Dec. 2005. The values of portfolio returns are in terms of percentage. SENTIMENT is the Baker and Wurglers orthogonalized sentiment proxy. The column All reports the results without regime sorting. The other columns Regime j , j = 1, 2, show the results based on regime-sorted observations as regime= j . The bootstrapped p-values are in parentheses. and denote signicance at 5% and 1% level, respectively.

34

Table 5: Out-of-Sample Predictability Test Results Using the Clark and Wests (2007) MSPE-adjusted Statistic Long-Short All Regime 1 Regime 2 Panel A: Portfolios Formed on Size 10 - 1 0.50 -4.50 2.25** 5-1 0.78 -2.39 2.17** Panel B: Portfolios Formed on Book-to-Market 10 - 1 -0.69 2.65** 0.98 5-1 -0.12 -2.85 1.65* Panel C: Portfolios Formed on Dividend Yield 10 - 0 1.58 -3.68 2.24** 5 - 0 1.42 -4.12 2.65** Panel D: Portfolios Formed on Earnings/Price 10 - 0 1.82* -3.30 2.25** 5 - 0 0.98 -4.41 2.34**
Note: This table reports the results of the out-of-sample tests over the period from Jan. 1966 to Dec. 2005. The Clark and Wests (2007) MSPE-adjusted statistic is computed using the prediction errors of the unrestricted and restricted models of (3) and (4) for the returns of the long-short portfolios of size, book-to-market, dividend yield, and earnings/price. The unrestricted model is the regressions of long-short portfolio returns on the lagged Baker and Wurglers orthogonalized sentiment proxy, the market factor (RMRF), the Fama-French factors (HML and SMB), the momentum factor (UMD), the liquidity factor (LIQ). The restricted model is the predictive regression model without the lagged Baker and Wurglers orthogonalized sentiment proxy. The long-short portfolio return that longs portfolio k2 and shorts portfolio k1 with rm characteristic i (including ) the portfolios for non-dividend-paying stocks or non-earning stocks. Portfolio returns are in percentage. The column All reports the results without regime sorting. The other columns Regime j , j = 1, 2, show the results based on regime-sorted observations as regime= j . The critical value for the one-sided test at the 5% and 1% signicance levels are 1.645 and 1.96, respectively. and denote signicance at 5% and 1% level, respectively.

35

Table 6: Predictive Regressions for Long-Short Portfolio Returns Using Consumer Condence Long-Short All Regime 1 Regime 2 Panel A: Portfolios Formed on Size 10 - 1 0.07** (0.00) -0.01 (0.61) 0.01* (0.04) 5-1 0.02* (0.05) 0.01 (0.33) 0.00 (0.37) Panel B: Portfolios Formed on Book-to-Market 10 - 1 0.00 (0.70) -0.05 (1.00) 0.03** (0.00) 5-1 0.01 (0.19) -0.01 (0.69) 0.02** (0.00) Panel C: Portfolios Formed on Dividend Yield 10 - 0 0.02 (0.10) -0.01 (0.74) 0.03* (0.02) 5 - 0 0.00 (0.48) -0.03 (0.89) 0.02** (0.01) Panel D: Portfolios Formed on Earnings/Price 10 - 0 0.01 (0.21) -0.01 (0.65) 0.03** (0.00) 5 - 0 0.01 (0.20) -0.01 (0.64) 0.03** (0.00)
Note: This table represents the results about regressions of long-short portfolio returns on lagged ConSENTIMENT , the market factor (RMRF), the Fama-French factors (HML and SMB), the momentum factor (UMD), and the liquidity factor (LIQ), r(i,k2 ),t r(i,k1 ),t = i + i,1 ConSENTIMENT t1 + i, RMRFt + i,2 SMBt +i,3 HMLt + i,4 UMDt + i,5 LIQt + i,t , where r(i,k2 ),t r(i,k1 ),t is the long-short portfolio return that longs portfolio k2 and shorts portfolio k1 with rm characteristic i (including size, book-to-market, dividend yield, and earnings/price) at time t, and k1 , k2 { 0, 1, 2, . . . , 10}. 1, 2, . . . , and 10 indicate the portfolios in the 1st (the smallest), 2nd, . . . , and 10th deciles, respectively. 0 represents the portfolios for nondividend-paying stocks or non-earning stocks. This table only reports the parameter estimates of i,1 . The sample period is from Jan. 1978 to Dec. 2005. The values of portfolio returns are in terms of percentage. ConSENTIMENT is the orthogonalized consumer sentiment proxy. The column All reports the results without regime sorting. The other columns Regime j , j = 1, 2, show the results based on regime-sorted observations as regime= j . The bootstrapped p-values are in parentheses. and denote signicance at 5% and 1% level, respectively.

36

Table 7: Predictive Regressions with Regime Dummies Long-Short i,1 i,2 Panel A: Portfolios Formed on Size 10 - 1 -0.09 (0.70) 0.53** (0.00) 5-1 0.14 (0.25) 0.30** (0.00) Panel B: Portfolios Formed on Book-to-Market 10 - 1 -0.63 (1.00) 0.28** (0.00) 5-1 -0.11 (0.75) 0.30** (0.00) Panel C: Portfolios Formed on Dividend Yield 10 - 0 0.19 (0.31) 0.53** (0.00) 5 - 0 0.18 (0.33) 0.48** (0.00) Panel D: Portfolios Formed on Earnings/Price 10 - 0 0.46 (0.12) 0.43** (0.00) 5 - 0 0.33 (0.26) 0.43** (0.00)
Note: This table reports the results about regressions of long-short portfolio returns on the regime dummies, the interactions of regime dummy variables and lagged SENTIMENT , the market factor (RMRF), the Fama-French factors (HML and SMB), the momentum factor (UMD), and the liquidity factor (LIQ), r(i,k2 ),t r(i,k1 ),t = i, + i,1 D1 + i,2 D2 + (i,1 D1 + i,2 D2 )SENTIMENT t1 +i, RMRFt + i,2 SMBt + i,3 HMLt + i,4 UMDt + i,5 LIQt +

i,t ,

where r(i,k2 ),t r(i,k1 ),t is the long-short portfolio return that longs portfolio k2 and shorts portfolio k1 with rm characteristic i (including size, book-to-market, dividend yield, and earnings/price) at time t, k1 , k2 { 0, 1, 2, . . . , 10}, and Dj is the dummy variable of regime j , which equals to 1 when regime= j and equals 0 otherwise. 1, 2, . . . , and 10 indicate the portfolios in the 1st (the smallest), 2nd, . . . , and 10th deciles, respectively. 0 represents the portfolios for non-dividend-paying stocks or non-earning stocks. The sample period is from Jan. 1966 to Dec. 2005. The values of portfolio returns are in terms of percentage. SENTIMENT is the Baker and Wurglers orthogonalized sentiment proxy. The bootstrapped p-values are in parentheses. and denote signicance at 5% and 1% level, respectively.

37

Long-Short 10 5 10 5 10 5 10 5 10 5 10 5 1 1 1 1 1 1 1 1

0 0 0 0

10 - 0 5 - 0 10 - 0 5 - 0

Table 8: Conditional Market Betas All Regime 1 Regime 2 Panel A: Portfolios Formed on Size i,1 0.24* (0.03) -0.07 (0.67) 0.28** (0.00) i,1 0.24** (0.00) 0.24 (0.17) 0.17** (0.01) i,1 -0.06 (1.00) -0.09 (0.98) -0.05 (1.00) i,1 0.00 (0.58) 0.04 (0.14) -0.05 (0.99) Panel B: Portfolios Formed on Book-to-Market i,1 -0.08 (0.90) -0.82 (1.00) 0.21** (0.00) i,1 0.13 (0.12) -0.31 (0.95) 0.25** (0.00) i,1 -0.09 (1.00) -0.18 (1.00) 0.01 (0.25) i,1 -0.07 (1.00) -0.16 (1.00) 0.01 (0.20) Panel C: Portfolios Formed on Dividend Yield i,1 0.34** (0.00) 0.04 (0.52) 0.34** (0.00) i,1 0.35** (0.00) -0.04 (0.64) 0.35** (0.00) i,1 -0.10 (1.00) -0.20 (1.00) -0.03 (0.86) i,1 -0.08 (1.00) -0.20 (1.00) 0.00 (0.27) Panel D: Portfolios Formed on Earnings/Price i,1 0.39** (0.00) 0.21 (0.32) 0.30** (0.00) i,1 0.41** (0.01) 0.03 (0.56) 0.33** (0.00) i,1 -0.10 (1.00) -0.19 (1.00) -0.04 (0.98) i,1 -0.07 (1.00) -0.20 (1.00) 0.02 (0.23)

Note: This table reports the results about regressions of long-short portfolio returns on lagged SENTIMENT , the market factor (RMRF), the Fama-French factors (HML and SMB), the momentum factor (UMD), the liquidity factor (LIQ), and interaction of RMRF and SENTIMENT ,
r(i,k2 ),t r(i,k1 ),t = i + i,1 SENTIMENT t1 + (i, + i,1 SENTIMENTt1 )RMRFt +i,2 SMBt + i,3 HMLt + i,4 UMDt + i,5 LIQt i,t ,

where r(i,k2 ),t r(i,k1 ),t is the long-short portfolio return that longs portfolio k2 and shorts portfolio k1 with rm characteristic i (including size, book-to-market, dividend yield, and earnings/price) at time t, and k1 , k2 { 0, 1, 2, . . . , 10}. 1, 2, . . . , and 10 indicate the portfolios in the 1st (the smallest), 2nd, . . . , and 10th deciles, respectively. 0 represents the portfolios for non-dividendpaying stocks or non-earning stocks. The sample period is from Jan. 1966 to Dec. 2005. The values of portfolio returns are in terms of percentage. SENTIMENT is the Baker and Wurglers orthogonalized sentiment proxy. The column All reports the results without regime sorting. The other columns Regime j , j = 1, 2, show the results based on regime-sorted observations as regime= j . The bootstrapped p-values are in parentheses. and denote signicance at 5% and 1% level, respectively.

38