Nội dung

THE JOURNAL OF FINANCE • VOL. LXII, NO. 3 • JUNE 2007 Giving Content to Investor Sentiment: The Role of Media in the Stock Market PAUL C. TETLOCK∗ ABSTRACT I quantitatively measure the interactions between the media and the stock market using daily content from a popular Wall Street Journal column. I find that high media pessimism predicts downward pressure on market prices followed by a reversion to fundamentals, and unusually high or low pessimism predicts high market trading volume. These and similar results are consistent with theoretical models of noise and liquidity traders, and are inconsistent with theories of media content as a proxy for new information about fundamental asset values, as a proxy for market volatility, or as a sideshow with no relationship to asset markets. One of the more fascinating sections of the WSJ is on the inside of the back page under the standing headline “Abreast of the Market.” There you can read each day what the market did yesterday, whether it went up, down or sideways as measured by indexes like the Dow Jones Industrial Average . . . . In that column, you can also read selected post-mortems from brokerage houses, stock analysts and other professional track watchers explaining why the market yesterday did whatever it did, sometimes with predictive nuggets about what it will do today or tomorrow. This is where the fascination lies. For no matter what the market did—up, down or sideways—somebody will have a ready explanation. Vermont Royster (Wall Street Journal, “Thinking Things Over Abaft of the Market,” January 15, 1986) CASUAL OBSERVATION SUGGESTS THAT THE CONTENT OF NEWS about the stock market could be linked to investor psychology and sociology. However, it is unclear whether the financial news media induces, amplifies, or simply ref lects investors’ interpretations of stock market performance. This paper attempts to ∗ Tetlock is at the McCombs School of Business, University of Texas at Austin. I am indebted to Robert Stambaugh (the editor), an anonymous associate editor, and an anonymous referee for their suggestions. I am grateful to Aydogan Alti, John Campbell, Lorenzo Garlappi, Xavier Gabaix, Matthew Gentzkow, John Griffin, Seema Jayachandran, David Laibson, Terry Murray, Alvin Roth, Laura Starks, Jeremy Stein, Philip Tetlock, Sheridan Titman, and Roberto Wessels for their comments. I thank Philip Stone for providing the General Inquirer software and Nathan Tefft for his technical expertise. I appreciate Robert O’Brien’s help in providing information about the Wall Street Journal. I also acknowledge the National Science Foundation, Harvard University and the University of Texas at Austin for their financial support. All mistakes in this article are my own. 1139 1140 The Journal of Finance characterize the relationship between the content of media reports and daily stock market activity, focusing on the immediate inf luence of the Wall Street Journal’s (WSJ’s) “Abreast of the Market” column on U.S. stock market returns. To my knowledge, this paper is the first to find evidence that news media content can predict movements in broad indicators of stock market activity. Using principal components analysis, I construct a simple measure of media pessimism from the content of the WSJ column. I then estimate the intertemporal links between this measure of media pessimism and the stock market using basic vector autoregressions (VARs). First and foremost, I find that high levels of media pessimism robustly predict downward pressure on market prices, followed by a reversion to fundamentals. Second, unusually high or low values of media pessimism forecast high market trading volume. Third, low market returns lead to high media pessimism. These findings suggest that measures of media content serve as a proxy for investor sentiment or noninformational trading. By contrast, statistical tests reject the hypothesis that media content contains new information about fundamental asset values and the hypothesis that media content is a sideshow with no relation to asset markets. I use the General Inquirer (GI), a well-known quantitative content analysis program, to analyze daily variation in the WSJ “Abreast of the Market” column over the 16-year period 1984–1999. This column is a natural choice for a data source that both ref lects and inf luences investor sentiment for three reasons. First, the WSJ has by far the largest circulation—over two million readers—of any daily financial publication in the United States, and Dow Jones Newswires, the preferred medium for electronic WSJ distribution, reaches over 325,000 finance and investment professionals.1 Second, the WSJ and Dow Jones Newswires, founded in 1889 and 1897, respectively, are extremely well established and have strong reputations with investors. Third, electronic texts of the WSJ “Abreast of the Market” column are accessible over a longer time horizon than are the texts of any other column about the stock market.2 For each day in the sample, I gather newspaper data by counting the words in all 77 predetermined GI categories from the Harvard psychosocial dictionary. To mitigate measurement error and thereby enhance construct validity, I perform a principal components factor analysis of these categories. This process collapses the 77 categories into a single media factor that captures the maximum variance in the GI categories. Because this single media factor is strongly related to pessimistic words in the newspaper column, hereafter I refer to it as a pessimism factor. In standard return predictability regressions, changes in this pessimism factor predict statistically significant and economically meaningful changes in the distribution of daily U.S. stock returns and volume. I confirm the robustness of this relationship by looking at the sensitivity of the results to the timing 1 Sources: Circulation and subscription data from Dow Jones and Company’s filing with the Audit Bureau of Circulations, September 30, 2003; circulation rankings from 2004 Editor and Publisher International Yearbook. 2 This statement ref lects my knowledge of newspaper columns available electronically as of December 2001. Giving Content to Investor Sentiment 1141 of information and to the use of different measures of pessimism. The results remain the same when I allow for a significant time gap between the release of media pessimism and the event return window. I also replicate the results using alternative measures of media pessimism based on the original GI categories. Using the GI category for either Negative words or Weak words (the two GI categories most highly correlated with pessimism), I find similar relationships between the media and the market. This approach to modeling behavioral phenomena yields factors and corresponding regression coefficients that can be readily interpreted in terms of well-established psychological variables. Section I provides motivation for studying the impact of the media on the stock market. Section II describes the factor analysis of the content of the daily “Abreast of the Market” column in the Wall Street Journal. Section III reports myriad tests of whether the pessimism media factor correlates with future stock market activity. In Section IV, I interpret the pessimism factor in terms of investor sentiment and show that the risk premium explanation does not explain the results. I conclude in Section V with a brief discussion of the results and suggestions for future research on the inf luence of media in asset markets. Finally, because this study relies heavily on a technique unfamiliar to many economists, the Appendix introduces the method of quantitative content analysis as it is employed in this study—for more detailed information, see Riffe, Lacy, and Fico (1998). I. Theory and Background Since John Maynard Keynes coined the term “animal spirits” 70 years ago, economists have devoted substantial attention to trying to understand the determinants of wild movements in stock market prices that are seemingly unjustified by fundamentals (see Keynes (1936)). Cutler, Poterba, and Summers (1989) is one of the first empirical studies to explore the link between news coverage and stock prices. Surprisingly, the authors find that important qualitative news stories do not seem to help explain large market returns unaccompanied by quantitative macroeconomic events. Two recent studies identify interesting relationships between trading volume and measures of communication activity. Antweiler and Frank (2004) study messages in Internet chat rooms focused on stocks, characterizing the content of the messages as “buy,” “sell,” or “hold” recommendations. Although they do not find a statistically or economically significant effect of “bullish” messages on returns, Antweiler and Frank (2004) do find evidence of relationships between message activity and trading volume and message activity and return volatility. Similarly, Coval and Shumway (2001) establish that the ambient noise level in a futures pit is linked to volume, volatility, and depth—but not returns. Most theoretical models of the effect of investor sentiment on stock market pricing make two important assumptions (see, e.g., DeLong et al. (1990a)). First, these models posit the existence of two types of traders, noise traders who hold random beliefs about future dividends and rational arbitrageurs who 1142 The Journal of Finance hold Bayesian beliefs. In this paper, I refer to the level of noise traders’ beliefs relative to Bayesian beliefs as investor sentiment. For example, when noise traders have expectations of future dividends that are below the expectations of rational arbitrageurs, I call their beliefs pessimistic. Further, I assume that these misperceptions of dividends are stationary, implying that beliefs do not stray arbitrarily far from Bayesian expectations over time. Second, these models assume that both types of traders have downwardsloping demand for risky assets because they are risk averse, capital constrained, or otherwise impaired from freely buying and selling risky assets. These assumptions lead to an equilibrium in which noise traders’ random beliefs about future dividends inf luence prices. Specifically, when noise traders experience a negative belief shock, they sell stocks to arbitrageurs, increasing volume and temporarily depressing returns. However, because these shocks are stationary, on average returns rebound next period when there is a new belief shock. Models of investor sentiment such as DeLong et al. (1990a) therefore predict that low sentiment will generate downward price pressure and unusually high or low values of sentiment will generate high volume. More generally, models of trade for any noninformational reason, such as liquidity needs or sudden changes in risk aversion, make these same predictions. For example, Campbell, Grossman, and Wang (1993) model how changes in the level of risk aversion for a large subset of investors can affect short-term returns. The only way to distinguish noise trader and liquidity trader theories is to interpret the media pessimism variable as a proxy for either investor sentiment or risk aversion. Because this debate is more philosophical than economic, I defer to the reader to draw her own conclusions. The timing of media pessimism is important in each theory. This paper tests the specific hypothesis that high media pessimism is associated with low investor sentiment, resulting in downward pressure on prices. It is unclear whether media pessimism forecasts investor sentiment or ref lects past investor sentiment. If the former hypothesis is correct, then one would expect high media pessimism to predict low returns at short horizons and a reversion to fundamentals at longer horizons. If the latter theory is correct, then one would expect high media pessimism to follow low returns and predict high returns in the future. The most likely scenario, however, is that both theories have an element of truth. If media pessimism serves as a proxy for periods of low past and future investor sentiment, one would expect to find that high pessimism follows periods of low past returns, forecasts low future returns at short horizons, and predicts high future returns at longer horizons. Insofar as pessimism ref lects past investor sentiment, the high long-horizon returns will exceed the low short-horizon returns. These predictions concerning returns are summarized in Figure 1. One alternative hypothesis is that the media pessimism measure is a proxy for negative information about the fundamental values of equities that is not currently incorporated into prices. If pessimism ref lects negative news about past and future cash f lows rather than sentiment, then one would still observe Giving Content to Investor Sentiment 1143 Figure 1. Impact of a negative sentiment shock on prices. The graph depicts the theoretical impact of a one-time increase in negative investor sentiment on equity prices. If the media pessimism measure is a predictor of investor sentiment, it will predict low short-horizon returns followed by high long-horizon returns of approximately equal magnitude. If the media pessimism measure follows past investor sentiment, it will predict low short-horizon returns followed by high long-horizon returns of greater magnitude than the short-horizon returns. a negative relationship between media pessimism and short-horizon returns. However, the sentiment and information theories make different predictions about long-horizon returns and volume: The sentiment theory predicts shorthorizon returns will be reversed in the long run, whereas the information theory predicts they will persist indefinitely. Although this discussion focuses on extreme views of the newspaper column as either pure noise or pure information, it is possible that the column contains only some information, but that traders over- or underreact to this information. I explore these possibilities further in the empirical tests in Section III. Another theory of media pessimism is that it is a proxy for negative information about dividends that is already incorporated into prices. This theory predicts media pessimism should have no effect on future market activity. Similarly, if one believes that the media pessimism measure contains no information about past, present, and future dividends, then one would not expect to observe any impact of pessimism on market activity. Many economists who have read the “Abreast of the Market” column in the WSJ support some variant of this theory, believing the column’s goal is to entertain readers. Trading volume provides another measure of market behavior for assessing theories of media pessimism. If media pessimism either ref lects past or predicts future investor sentiment, unusually high or low levels of pessimism should be associated with increases in trading volume. More precisely, if pessimism has 1144 The Journal of Finance a mean of zero, then the absolute value of pessimism is high in times when irrational investors trade with rational investors.3 Although the sentiment theory makes a clear prediction about the relationship between volume and pessimism, the information theory makes no obvious prediction.4 Finally, the stale or no information theory predicts no effect of media pessimism on trading volume. II. Generating the Pessimism Media Factor As a completely automated program, the General Inquirer (GI) produces a systematic and easily replicable analysis of the WSJ column. The GI employs an extremely rudimentary measurement rule for converting the column into numeric values, namely, it counts the number of words in each day’s column that fall within various word categories. The word categories are neither mutually exclusive nor exhaustive—one word may fall into multiple categories and some words are not categorized at all. To reduce redundancy in categorization, I use only the most recent versions of categories in the General Inquirer’s Harvard IV-4 psychosocial dictionary. To minimize semantic and stylistic noise in the column, I recenter all GI categories so that their conditional means are equal across different days of the week. This ensures that I do not select media factors that capture the systematic variation in the WSJ column on different days of the week. I use day-of-theweek dummy variables in the regressions in the next section to control for the possibility that market behavior differs across different days of the week. I employ a principal components factor analysis to extract the most important semantic component from the (77 × 77) variance–covariance matrix of the categories in the Harvard dictionary. This process is designed to detect complex structure in the WSJ column and to eliminate the redundant categories in the dictionary. Factor analysis assumes the existence of an underlying media factor—a linear combination of GI categories that is not directly observable.5 Variation in this factor over time generates the observed daily correlations between the various GI categories. Operationally, factor analysis chooses the vector in the 77-dimensional GI category space with the greatest variance, where each GI category is given equal weight in the total variance calculation. I explore other factor analysis 3 Most models in finance focus on trades between groups of noise traders and rational traders. Traditional no-trade theorems suggest that within-group trades among rational traders should not occur. Furthermore, for noise traders to have an impact on prices, there must be a common component in the variation in their beliefs. This paper focuses on the common component of noise trader beliefs that could affect prices. 4 Of course, it is possible that new information produces divergence in opinion, which would lead to increases in volume. On the other hand, it seems equally likely that agents’ beliefs would converge when all agents observe the same piece of public information. 5 In an earlier version of this paper, I consider the top three factors, some of which have interesting interpretations. Adding additional factors to the regressions shown here does not substantially alter the results because all factors are mutually orthogonal by construction. Giving Content to Investor Sentiment 1145 techniques, such as principal factors analysis and maximum-likelihood factor analysis. The qualitative empirical conclusions are not sensitive to the methodology chosen, and the quantitative conclusions change only minimally. For the remainder of this paper, I present the results using the single factor identified by principal components analysis that captures the maximum variation in GI categories. Effectively, principal components analysis performs a singular value decomposition of the correlation matrix of GI categories measured over time. The single factor selected in this study is the eigenvector in the decomposition with the highest eigenvalue. Because this singular value decomposition uses only the GI category variables in the correlation matrix, completely disregarding all stock market variables, the resulting media factor may not correspond to any traditional measurements of past market performance. Also, because I do not subjectively eliminate any categories, the factor analysis generates a media factor that equally considers all sources of variation in the WSJ column, even though it is likely that some categories are more relevant for measuring investor sentiment.6 Theoretical imprecision is the cost of avoiding data mining. To facilitate the interpretation of the single factor chosen, I adopt a complementary approach to creating a summary media variable later in the paper. The principal components analysis exploits time variation in GI category word counts to identify the media factor. To avoid data mining and any lookahead bias in the regressions that follow, I construct the media factor using only information available to traders. I estimate the factor loadings in year t − 1 using principal components analysis. Then I use these loadings, along with the daily word counts in year t, to calculate the values of the factor throughout year t. Because this procedure does not guarantee any consistency in factor loadings across years, it is possible that changes in the structure of media content over time cause this procedure to generate a meaningless factor. For example, if the column writer focuses on different issues in different years, then this procedure will generate a single factor that covaries with different issues in different years. To examine whether time variation in the GI categories is stable and whether the above procedure is reasonable, I analyze the relationship between the loadings used in each yearly factor. For each year in the data sample, I use the loadings estimated from that year to calculate the value of the hypothetical value of the yearly factor in all years. Then I compare the correlations for these hypothetical yearly factors across the entire sample. Fortunately, the media factor estimated using the loadings from any given year looks very similar to the media factor estimated using another year’s loadings. In Table I, I report the correlation matrix of yearly factors. The average pairwise correlation between the yearly media factors is 0.96 and the average squared correlation is 0.91. The minimum pairwise correlation is 0.80, suggesting all pairs of media factors are very highly correlated. I conclude from this 6 For example, investors may not care how many religious words appear in the column each day. Nevertheless, the GI dictionary devotes an entire category to tracking these words. 1146 The Journal of Finance Table I Correlations of the Media Factors Constructed Yearly The table data come from the General Inquirer program. This table shows correlations between media factors constructed using factor analysis on the GI categories in different years. The mean pairwise correlation for a given yearly factor excludes the factor’s correlation with itself. Each yearly media factor is the linear combination of GI categories that captures the maximum variance in GI categories in that year. The factor analysis method is that of principal components analysis, which is equivalent to a singular value decomposition. Each yearly factor analysis is based on news columns from roughly 250 trading days per year. All GI category variables have been demeaned by day of the week using the prior year’s mean. Year ’84 ’85 ’86 ’87 ’88 ’89 ’90 ’91 ’92 ’93 ’94 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1.00 0.95 0.94 0.95 0.94 0.92 0.93 0.95 0.92 0.91 0.90 0.95 0.92 0.91 0.93 1.00 0.98 0.97 0.96 0.92 0.93 0.96 0.97 0.96 0.95 0.95 0.96 0.93 0.95 1.00 0.98 0.97 0.94 0.93 0.97 0.97 0.97 0.96 0.97 0.98 0.92 0.96 1.00 0.97 0.96 0.96 0.97 0.95 0.95 0.94 0.97 0.96 0.92 0.96 1.00 0.96 0.95 0.98 0.98 0.97 0.96 0.97 0.97 0.96 0.97 1.00 0.95 0.97 0.94 0.93 0.93 0.98 0.96 0.91 0.94 1.00 0.96 0.94 0.95 0.92 0.96 0.96 0.93 0.96 1.00 0.95 0.97 0.94 0.98 0.97 0.94 0.96 1.00 0.97 0.97 0.96 0.87 0.95 0.97 1.00 0.97 0.96 0.98 0.94 0.96 1.00 0.94 0.97 0.94 0.96 ’95 ’96 ’97 ’98 1.00 0.97 1.00 0.94 0.94 1.00 0.97 0.97 0.97 1.00 Mean 0.93 0.95 0.96 0.96 0.97 0.94 0.95 0.96 0.96 0.96 0.95 0.96 0.96 0.94 0.96 analysis that the loadings on the individual GI categories are quite stable over time. Each yearly factor analysis can be interpreted in terms of the underlying GI categories. The average of the first eigenvalue in each yearly factor analysis is 6.72, implying that the first factor contributes as much variance in media content as more than six of the original GI category variables. This first factor is approximately equal to a linear combination with positive weights on just 4 of the 77 GI categories: Negative, words associated with a negative outlook; Weak, words implying weakness; Fail, words indicating that goals have not been achieved; and Fall, words associated with falling movement. In fact, the Negative and Weak GI categories each can explain over 57% of the variance in the first factor. This factor also negatively weights categories such as Positive, words associated with a positive outlook;7 however, the negative relationship between the factor and Positive words is not as strong as the positive relationship between the factor and Negative words. 7 Intuitively, the number of Positive and Negative words in the column is strongly negatively correlated, holding constant the total number of words. Thus, it is natural that one media factor captures the variation in both Positive and Negative words. Giving Content to Investor Sentiment 1147 On days in which the WSJ column loads highly on this first factor, it probably contains negative interpretations of market events. These interpretations may or may not correspond to objectively negative news about corporate earnings and other relevant economic indicators. Hereafter, I use the term pessimism factor to refer to this first media factor. In thinking about the pessimism factor, it is important to remember that the typical “Abreast of the Market” column reads like a post-mortem of the market’s life on the prior day.8 The WSJ’s “Abreast of the Market” column is written immediately after the closing bell on Wall Street on the most recent day of market activity. In fact, the writing process usually begins even before the closing bell at 2:30–3:00 p.m. EST. It is unlikely that the column contains any information unknown to the specialists on the f loor of the exchange at the conclusion of trading. Nevertheless, in robustness tests shown below, I introduce a significant time gap between the release of the column in the afternoon and the beginning of the event return window. The same column produced and released by Dow Jones Newswires each afternoon is republished in the WSJ on the morning of the next day of market activity. III. Market Activity and Pessimism The market returns regressions, I present below control for many known sources of predictability found in daily return data. In a frictionless complete market in which traders have rational expectations, financial theory predicts daily prices will follow a random walk with drift (Samuelson (1965)). More generally, classic arbitrage arguments suggest that prices should approximate a random walk with drift. However, market microstructure phenomena, such as bid-ask bounce, nonsynchronous trading, and transactions costs, sully the purity of the theoretical prediction. Some of these mechanisms induce statistical artifacts that make observed returns appear autocorrelated; other mechanisms lead to genuine return autocorrelation.9 With these caveats in mind, this study attempts to control for the inf luence of the lagged returns of indices and the lagged returns of any leading, larger indices. Regressions also include lagged volume to try to capture liquidity effects. Finally, lagged return volatility acts as a proxy for the inf luence of several other market frictions. The return and volatility predictability regressions control for all lags up to 5 trading days, that is, at least 1 week of calendar time. 8 Journalists, not economic or financial experts, write the column. A typical writer has a B.A. degree in Journalism and has taken few, if any, courses in economics. The column writers sometimes leave their offices in Jersey City to go to the f loor of the stock exchange in Manhattan in search of opinion quotes from strategists and traders, but usually they ask market participants questions by telephone. On the whole, the writers consider their columns “more art than science.” 9 Bid-ask spreads can induce negative autocorrelation in the return time series when trades alternately occur at the bid and ask prices. Nonsynchronous trading causes spurious cross-correlations and autocorrelations in stock returns because quoted closing prices are not equally spaced at 24hour intervals (see Campbell, Lo, and MacKinlay (1995)). Transaction costs such as trading fees and short-run capital gains taxes preclude arbitrage strategies that would mitigate return predictability. 1148 The Journal of Finance This study runs two main sets of regressions to test whether the pessimism factor predicts returns and volume beyond known sources of predictability. In the first set of tests, I adopt a vector autoregressive (VAR) framework in which I simultaneously estimate the relationships among returns, volume, and the pessimism factor. In the second set of tests, I examine the robustness of these results. Finally, I assess the economic importance of the return predictability and whether it implies there are profitable sentiment-based trading strategies. A. VAR Estimates Because Dow Jones and Company produces the WSJ and the “Abreast of the Market” column focuses on the stocks in the Dow Jones index, I test whether pessimism forecasts daily returns on the Dow Jones Industrial Average (Dow). In addition, because previous authors (e.g., DeLong et al. (1990a)) suggest that investor sentiment plays a larger role in the pricing of small stocks, I consider whether pessimism forecasts the returns to the Fama and French (1993) daily small-minus-big (SMB) factor. As a measure of volume, I look at the detrended log of daily volume (Vlm) on the New York Stock Exchange (NYSE).10 I obtain a time series of daily returns from January 1, 1984 to September 17, 1999 from the Wharton Research Data Services’ access to the historical Dow Jones Industrial Averages. This sample period encompasses almost 4,000 observations of market returns; each year consists of only about 250 data points because the U.S. stock market is idle on weekends and national holidays. All VAR estimates include all lags up to 5 days prior to market activity. The endogenous variables in the first VAR are Dow, the pessimism media factor (BdNws), and Vlm.11 The exogenous variables include five lags of the detrended squared Dow residuals to proxy for past volatility,12 dummy variables for dayof-the-week and January to control for other potential return anomalies, and a dummy variable for the October 19, 1987 stock market crash to ensure the results are not driven by this single observation.13 Newey and West (1987) robust standard errors account for any heteroskedasticity and autocorrelation in the residuals up to five lags. It is convenient to define the variable Exog, which represents all of the exogenous variables. I also define a lag operator, L5, that transforms any 10 I focus on detrended log volume because the level of log volume is not stationary. I use a detrending methodology based on Campbell, Grossman, and Wang (1993). Specifically, I calculate the volume trend as a rolling average of the past 60 days of log volume. All results below are robust to using 30-day and 360-day averages. 11 The time series of daily NYSE volume comes from the NYSE database. 12 Specifically, I demean the Dow variable to obtain a residual, I square this residual, and then I subtract the past 60-day moving average of the squared residual. All results below are robust to using alternative detrended measures of past volatility in which I subtract the past 30-day or 360-day moving average of squared residuals from current squared residuals. The results for other volatility measures such as the CBOE’s volatility index (VIX) are qualitatively similar. 13 Interestingly, measures of pessimism seem to have been abnormally high prior to the 1987 crash, which is consistent with the findings below. I also verify that winsorizing all variables at the 0.5% upper and lower tails of their distributions does not affect the results.