Media Makes Momentum - European Retail Investment Conference

0 downloads 164 Views 603KB Size Report
The effect is more pronounced in states with higher investor individu- .... We follow this line of reasoning by using th
Media Makes Momentum Alexander Hillert, Heiko Jacobs and Sebastian M¨ uller∗

February 2013 Appendix attached

Abstract Relying on 2.2 million articles from 45 national and local U.S. newspapers between 1989 and 2010, we find that firms particularly covered (neglected) by the media exhibit ceteris paribus significantly stronger (weaker) momentum. The effect is more pronounced in states with higher investor individualism and among stocks predominantly held by overconfident fund managers. Findings suggest that media coverage can exacerbate investor biases, leading return predictability to be strongest for firms in the spotlight of public attention. In line with prominent models such as Daniel et al. (1998), our results collectively support an overreaction-based explanation of the momentum effect. Keywords: Momentum, media, overreaction, attention effects, investor biases JEL Classification Codes: G12, G14

∗ Alexander

Hillert, Chair of International Finance and CDSB, University of Mannheim, L9, 1-2, 68131 Mannheim. E-

Mail: [email protected]. Heiko Jacobs, Chair of Banking and Finance, University of Mannheim, L 5, 2, 68131 Mannheim, Germany. E-Mail: [email protected]. Sebastian M¨ uller, Chair of Banking and Finance, University of Mannheim, L 5, 2, 68131 Mannheim, Germany. E-Mail: [email protected]. We thank seminar participants at the Second Helsinki Finance Summit on Investor Behavior, the University of Mannheim, the LMU Munich, the Campus for Finance conference 2013, as well as the University of California, Berkeley for valuable comments.

1

1

Introduction

The momentum effect is among the strongest and most pervasive return anomalies. While its existence has been convincingly documented in different time periods, countries, indices, and asset classes1 , a central issue is still far from being resolved: What are the underlying causes of momentum? In other words: Why exactly do winner stocks of the recent past tend to outperform loser stocks of the recent past? The goal of this paper is to employ an extensive media data set to shed new light on this long-standing debate. Recent research demonstrates that media coverage directly affects the way in which investors collect, process, and interpret information (e.g. Barber and Odean (2008), Engelberg and Parsons (2011), Tetlock (2011), Engelberg et al. (2012)). Collectively, findings suggest “a potentially important role for the media in shaping the behavior of the stock market” (Hong and Stein (2007), p. 118). The interesting link to the momentum literature lies in the fact that investors’ attention and information processing also play a crucial role in prominent behavioral theories of momentum.2 Whether momentum will ceteris paribus be weaker or stronger among high coverage stocks, is, however, not clear ex ante. On the one hand, media coverage may lead to faster information diffusion and thus weaken momentum. On the other hand, media coverage may contribute to investor biases, thereby strengthening momentum. As we outline in the literature review, theoretical considerations offer support for both views, making exploring the role of the media an important empirical task. To this end, we rely on a novel and carefully constructed data set of newspaper articles. It comprises approximately 2.2 million news stories in four leading national as well as 41 local U.S. newspapers from 1989 to 2010, which we match to 7,815 firms. The essence of our findings is captured in figure 1.

Insert figure 1 here

The cumulative profits to a winner minus loser long-short portfolio are displayed separately for stocks in the highest media coverage quintile (red solid line) and lowest quintile (blue dashed line). Media coverage 1 The

momentum effect was first described in Jegadeesh and Titman (1993). For U.S. stock momentum in earlier and

later timer periods see e.g. Chabot et al. (2009) and Jegadeesh and Titman (2001) as well as Fama and French (2008), respectively. For international evidence, see for instance Asness et al. (2012), Chui et al. (2010), Fama and French (2012), Liew and Vassalou (2000), Rouwenhorst (1998) and Rouwenhorst (1999). For momentum at the stock index level, see Bhojraj and Swaminathan (2006), Chan et al. (2000), Connolly and Stivers (2003), or Moskowitz et al. (2012). Momentum effects in other asset classes such as bonds, commodities, or currencies have been identified in e.g. Asness et al. (2012), Erb and Harvey (2006), or Moskowitz et al. (2012). 2 The

profitability of momentum is difficult to reconcile with rational asset pricing models. For instance, Chui et al.

(2010) note: “Given the magnitude of momentum profits, about 12% per year in the United States and Europe, they are unlikely to be explained by risk-based theories”. Consistent with this view, most theoretical and empirical work focuses on behavioral approaches to explain momentum. For risk-based momentum explanations, see e.g. Chordia and Shivakumar (2002), Avramov et al. (2007), Sagi and Seasholes (2007), and Stivers and Sun (2010).

2

is based on the number of firm-specific articles in four leading national newspapers (New York Times, USA Today, Wall Street Journal, and Washington Post), published in the skipped month between the momentum formation and evaluation period. Importantly, media coverage is defined as a residual from monthly cross-sectional regressions which control for a set of firm variables known to drive the probability of being covered (most notably firm size). While we describe the methodology in detail below, the basic advantage is that we can better isolate the true impact of excess media attention. The plot reveals a clear and economically strong positive relation between media coverage and the size of initial momentum profits and subsequent reversal. The winner minus loser spread amounts to approximately 5.3% after six months in the high media coverage portfolio, but only to 2.5% in the low media coverage portfolio. The resulting 2.8% return difference closely corresponds to a monthly return spread of 44 basis points (bp) that we obtain using the Jegadeesh and Titman (1993) portfolio construction procedure. The difference in momentum profits mostly comes from the winner side, it is economically large (the unconditional momentum return over the sample period is just 65 bp per month), and highly statistically significant with a t-statistic of 4. It is, however, not a permanent effect. Ten months after portfolio formation, the reversal begins for both momentum portfolios, but it is substantially stronger in the high media coverage portfolio. As a consequence, the cumulative return difference between both portfolios shrinks towards zero after 24 months. The overall pattern in figure 1 is confirmed in more than 20 sensitivity tests. Given the endogeneity of press articles (authors do not randomly choose about which companies they write), we are particular careful in making sure that we do not simply pick up a spurious correlation between media coverage and momentum. For instance, newspapers may be more attentive towards firms that did particularly well or poorly in the recent past. If the extremity of past returns was also systematically related to momentum profitability, then not controlling for this effect would lead to the incorrect conclusion that media coverage per se, and not its interaction with formation period returns, causes stronger momentum. As recently uncovered by Bandarchuk and Hilscher (2013), this line of reasoning indeed affects the role of several stock-level variables such as size, analyst coverage, or turnover, all of which have previously been argued to enhance momentum profits. Once one adequately controls for formation period returns, the significance of these characteristics largely vanishes. Applying the same methodology, we are able to confirm the findings of Bandarchuk and Hilscher (2013). Importantly, however, excess media coverage is among the very few characteristics which are effectively immune to their critique and thus qualify to convincingly support behavioral momentum explanations. The number of press articles may also rise for companies announcing important corporate news. This again raises concerns that higher momentum profits might be caused by the event itself and not by increased media coverage. We therefore exclude or control for media articles around earnings announcement dates, but find no impact on our baseline results. As a more comprehensive approach, we collect all 8-K filings, which are accessible via the EDGAR database from 1995 on. In total, we are able to gather more

3

than 315,000 of these documents for our sample firms. http://www.sec.gov/answers/form8k.htm defines an 8-K filing as a “‘current report’ companies must file with the SEC to announce major events that shareholders should know about”. For instance, potential events include changes in executive management, changes in ownership, entry into a material definitive agreement, material impairments, matters related to accountants and financial statements, or asset redeployments. However, inferences again do not change if we control for these events. Firms with high (low) residual press coverage exhibit significantly stronger (weaker) return momentum and reversal. These baseline findings are thus suggestive of an overreaction explanation in the spirit of Daniel et al. (1998). The authors advocate a model in which momentum arises as a result of two central investor biases: Overconfidence and self-attribution. Mistaken beliefs lead investors to overweight (underweight) public signals which confirm (contradict) their private information. Confirming news will be considered as evidence of one’s own skills, whereas disconfirming news will largely be neglected. As a consequence, overconfidence increases even further and prices temporarily overshoot, before the mispricing is gradually collected. In the context of the Daniel et al. (1998) model, media coverage might catch investors’ attention (see also Barber and Odean (2008)), which is a necessary ingredient for the overreaction-driven price momentum mechanism as pointed out by Hou et al. (2009). In addition, it should be easier for investors to identify confirming news for highly covered firms than for firms generally neglected by the press. In sum, this has two testable implications: First, and as verified above, momentum and reversal effects should all else equal be more pronounced for firms with high media coverage. Second, cross-sectional differences in biases underlying investor overreaction should translate into cross-sectional differences in media-based momentum and subsequent reversal. As a consequence, these effects should be most pronounced in stocks traded by a particularly overconfident investor base. To explore this second hypothesis, we run two conceptionally quite different tests. Our first analysis is motivated by Chui et al. (2010). The authors document that momentum profits are higher in countries with stronger individualism, a personal trait which they argue to be correlated with overconfidence and biased self-attribution. We follow this line of reasoning by using the ranking system of Vandello and Cohen (1999), which accounts for variation in individualism across the 50 U.S. states. While their index has been employed in a large variety of different psychological studies, it has largely been neglected in finance research so far. With reference to investors’ well-known bias for local investments (see e.g., Grinblatt and Keloharju (2001) and Seasholes and Zhu (2010)), we hypothesize that the media’s influence on momentum profits is substantially larger for firms incorporated in states with high individualism tendencies. Considering the influence of local media reports on the trading behavior of local investors (e.g. Engelberg and Parsons (2011)), we enrich our media data set with articles from 41 local newspapers. A number of portfolio approaches as well as multivariate Fama/MacBeth regressions provide support for our prediction.

4

The difference in the media-based momentum effect between individualistic and collectivistic areas is estimated to be about 30 to 40 bp per month. The second test is more direct in nature in that we rely on explicit holdings data of mutual fund managers. Our analysis is inspired by Fang et al. (2011), who uncover a negative relationship between a manager’s propensity to buy stocks covered in the media and subsequent fund performance. We link this finding with recent work uncovering a systematic negative relation between overconfidence or self-attribution bias and subsequent performance (e.g. Nikolic (2011), Choi and Lou (2010)). To this end, we follow Nikolic (2011) in constructing a fund manger overconfidence index comprising twelve characteristics such as e.g. managerial gender, tenure, or portfolio concentration. We then reveal that overconfident managers appear to rely particularly heavily on media coverage in their buying decisions. Switching from the fund manager level to the stock level, we finally verify that both media-based momentum and the subsequent reversal are more pronounced among stocks predominantly held by more overconfident fund managers. Taken together, these findings suggest that, contrary to the wide held belief, it is not always the “dark corners” where the violation of the weak form of market efficiency appears particularly pronounced. Our analysis instead lends support to the idea that media coverage might exacerbate investor biases so that momentum and reversal effects can be strongest for firms in the spotlight of public attention.

2

Related literature

Several behavioral models can give rise to the momentum effect (e.g. Long et al. (1990), Barberis et al. (1998) or Grinblatt and Han (2005)). The gradual information diffusion model of Hong and Stein (1999) is among the most popular. It argues that momentum is primarily caused by investors’ underreaction to fundamentally relevant information. Hong and Stein (1999) focus on heterogeneity among investors, who are assumed to acquire different pieces of value-relevant news at different points in time without being able to extract information from observed prices. This causes valuable information being incorporated only slowly into prices. Media coverage could potentially speed up this process (e.g. Huberman and Regev (2001), Klibanoff et al. (1998), Peress (2008)). Coverage has been argued to disseminate information to a broad audience (Fang and Peress (2009)), thereby potentially reducing information asymmetries (Tetlock (2010)). Given these insights, one might conjecture: If slow information diffusion is the major underlying source of stock price momentum, momentum should ceteris paribus be less pronounced for firms with high media coverage. The intuition behind Hong and Stein (1999) has received empirical support in several studies. For instance, Hong et al. (2000) show that price drifts, particularly among loser stocks, are stronger among firms with comparatively little analyst coverage. Da et al. (2012) show that, during their sample period from

5

2000 to 2007, return continuation is ceteris paribus more pronounced among stocks without large price movements and without Dow Jones Newswire coverage shocks. The authors interpret these findings as investors’ underreaction to a continuous flow of small amounts of information. Similarly, relying on firms covered by Thomson Reuters NewsScope news services between 2003 and 2008, Sinha (2011) argues that the market is slow in incorporating the qualitative content of the news into prices. Using the Dow Jones Interactive Publications Library as database, Chan (2003) analyzes differences in stock price reactions subsequent to large absolute returns in the previous month depending on whether the underlying company had a news headline story in the given month or not. He finds evidence of a strong drift without a reversal after bad news, which is in line with slow information diffusion. While his paper is probably the one closest to our study, our approach differs in several crucial ways. The appendix discusses differences in detail, but the major factors can be summarized as follows. First, his findings appear to be considerably driven by articles from newswires, which are likely to proxy for the arrival of valuable news. In contrast, we focus on newspaper articles, which often rather seem to proxy for (not necessarily information-driven) investor attention. Second, the reported return drift in Chan (2003) is largely restricted to smaller and more illiquid stocks. In contrast, by dropping about 50% of CRSP firm months, we aim at concentrating on larger, easily investable firms. Third, there are methodological differences. Chan relies on a binary analysis quantifying whether a stock was mentioned in the headline or lead paragraph in a given month. In contrast, we compute an excess coverage measure which quantifies the unexpectedly high or low weight the media attaches to a given firm, holding important characteristics such as firm size or analyst coverage fixed. Together, these aspects might help to explain why we arrive at partly different conclusions, especially with regard to long-term reversals and the role of formation period loser stocks. A second prominent model is the one of Daniel et al. (1998), which we already sketched in the introduction. As outlined, its intuition suggests that momentum might be actually more (instead of less) pronounced for firms particularly covered by the press. Credibility for this idea also comes from recent work arguing that the way the media reflects the news can induce temporary stock price distortions (e.g. Gurun and Butler (2011), Engelberg et al. (2012)) or even predict market movements (e.g. Tetlock (2007), Dougal et al. (2011)). Moreover, this line of reasoning also suggests a more subtle hypothesis: As buying attentiongrabbing stocks should generally be easier than short-selling attention-grabbing stocks (e.g. Barber and Odean (2008), D’Avolio (2002)), we would expect the impact of the media to be stronger on the long side of the momentum portfolio. The model of Daniel et al. (1998) has also received considerable empirical support. For instance, evidence from a recent field experiment shows that investors indeed seek information from stock message boards that confirms their prior beliefs (Park et al. (2013)). On the market level, Cooper et al. (2004) as well as Hou et al. (2009) show that momentum is much more pronounced in up markets, during which aggregate

6

investor overconfidence is assumed to be higher (Gervais and Odean (2001)). Antoniou et al. (2011) and Stambaugh et al. (2012) further document an incremental impact of investor sentiment indicators. Da et al. (2011) documents attention-induced price pressure over two weeks for those firms which are abnormally often searched for via Google. In sum, behavioral momentum theories and previous work lend credibility to quite different roles of the media in the price formation process. Assessing whether, and, if so, how media coverage affects the magnitude of momentum and reversal effects is thus an empirical task.

3

Data and empirical setup

Our initial sample consists of all common stocks traded on NYSE, AMEX or NASDAQ that appear on both CRSP and Compustat at any time during 1989 and 2010. To ensure that our findings are not driven by small illiquid stocks or bid-ask bounces, we follow Jegadeesh and Titman (2001) and exclude stocks with a market capitalization below the 10th NYSE size percentile and stocks with prices below $5 at the end of the formation period. For all companies we then collect newspaper articles from LexisNexis. Depending on the specific focus of the different tests in this study, we either rely on major national newspapers only or add articles from the local press. For the baseline examination, we follow Fang and Peress (2009) and focus on the four major U.S. newspapers with weekday circulation: New York Times (NYT), USA Today (USAT), Wall Street Journal (WSJ), and Washington Post (WP). In order to restrict ourselves to articles which truly address a specific company, we use the “relevance score” measure of LexisNexis. In our baseline tests, we retain all articles with a relevance score from 80% to 99%. To gather all company-related articles, we rely on an algorithm that automatically searches the LexisNexis database using company names from Compustat as keywords in the company search function.3 We additionally collect articles from the maximum set of local newspapers available via LexisNexis. A complete list of the resulting 41 local and four national newspapers as well as summary statistics are given in table 1. 3 These

are current names which are appropriate in most cases since LexisNexis takes into account name changes and

stores older articles also under the new company name. However, in few instances articles before a name change were obviously missing. To obtain these articles, we therefore also use historical names from CRSP (manually corrected for abbreviations) in the search engine and afterwards delete any duplicates. For all firms with more than 200 articles over the entire period we conduct manual plausibility checks to ensure that the articles really address the specific company and make corrections where necessary. Moreover, by manually screening a randomly selected sample of 1,000 newspaper articles out of the total sample at the end of the downloading process, we find that the algorithm was able to correctly assign articles to companies in most cases (hit rate of 95.5%).

7

Insert table 1 here

In total we are able to assign approximately 2.2 million newspaper articles to more than 7,800 different firms in our data set. Of these, 562,103 articles on more than 6,500 different firms are released in one of the four major national newspapers. The appendix shows summary statistics of media coverage for these newspapers. We find that around 45% of the 78,986 firm-year observations in our sample receive at least one newspaper article. The coverage of NYSE stocks is about double as high (64%) as the coverage of NASDAQ stocks (33%). Media coverage is highly skewed, which is consistent with the findings of Fang and Peress (2009). Among firms that are covered by the media in a given year, the mean number of articles is around 16, but the median is only 3 and the 25th percentile is 1. In figure 2 we additionally plot the time-series evolution of the percentage of firms with press coverage. As can be seen, the decrease in percentage of covered firms is mainly restricted to the national audiences, while local media coverage remain rather constant in the last years of the sample period. Note that the sharp increase during the early 90s for local newspapers is largely a consequence of the limited availability in LexisNexis at the beginning of the sample period. Coverage by all newspapers is also relatively stable and fluctuates between 60% and 70% for most years.

Insert figure 2 here

Panel A of table 2 shows time-series averages of monthly cross-sectional correlations between media coverage and a set of other firm characteristics. Given the skewness of the media data, we define media coverage as ln(1+number of articles) rather than using the raw number of articles. The construction of the remaining firm variables follows the standard in the literature and is explained in the appendix. It turns out that firm size (0.49), S&P 500 membership (0.43), NASDAQ membership (-0.20) and analyst coverage (0.33) in particular are strongly related to media coverage. These results raise concerns that when sorting stocks solely by media coverage we might also capture some hidden effects.

Insert table 2 here

In order to isolate the impact of media coverage, we closely follow the approach of Hong et al. (2000). That is, we estimate excess media coverage, defined as residuals of cross-sectional regressions with media coverage as dependent variable. While we have experimented with a large set of different estimation models (the results are reported in the robustness section 4.2), our primary specification is based on month-to-month rolling OLS regressions with size, analyst coverage, and dummies for S&P 500 and NASDAQ membership as explanatory variables: ln(1+no. articles) = α+β1 ·f irm size+β2 ·analyst coverage+β3 ·S&P 500+β4 ·N ASDAQ+εmedia (1) 8

Our model closely follows the approach of Hong et al. (2000). The only difference is that the model includes S&P 500 membership as a dummy variable in addition to NASDAQ membership and analyst coverage, which is the primary variable of interest in their paper. In our view, this estimation procedure has the advantage of being simple and easy to follow, while at the same time still removing the most serious dependencies between media coverage and other variables. The model achieves an average adjusted R2 of 0.26. As the appendix shows, adding a host of additional firm characteristics, including industry dummies, only leads to a modest further increase in the adjusted R2 to 0.30. Panel B of table 2 provides correlation results for residual media coverage. While the residuals are still highly correlated with raw media coverage (0.86), the correlations between the residuals and other firm characteristics shrink toward zero. This is not only the case for the variables included in the regression, but also for the book-to-market ratio, stock turnover, idiosyncratic volatility, the stock’s absolute value of its return over the past six months, stock price, the Amihud (2002) illiquidity ratio, and firm age. We are thus confident that our procedure provides a parsimonious and intuitive approach to isolate the effect of media coverage to the greatest extent.

4 4.1

Empirical results Baseline analysis: The effect of media on momentum returns

Following the standard in the literature, we construct momentum portfolios using a formation and a holding period of six months each, and skipping one month in between. To interact media exposure with momentum profits, we first create five equally sized stock portfolios based on residual media coverage. Within each of these quintiles, the winner (loser) portfolio then consists of all stocks with a return above the 70th percentile (below the 30th percentile) across their formation period. These return cut-offs are used throughout all tests in the paper. Residual media coverage is based on the number of articles published during the skipped month in one of the four national newspapers. A benefit of this approach is that there are no overlaps between the formation period and the period measuring excess media coverage. Table 3 shows our results. Returns are given in % per month (as in the remainder of the paper) and are based on overlapping equally-weighted portfolios as in the original work of Jegadeesh and Titman (1993). T-statistics are adjusted for serial autocorrelation using West and Newey (1987) standard errors with a lag of five months. Insert table 3 here As shown in panel A, there is a clear relation between media portfolios and raw momentum profits. While the winner minus loser spread amounts to 41 bp per month (t-stat 1.61) for the lowest coverage 9

portfolio, it is more than twice as large for the highest coverage portfolio (85 bp, t-stat 3.13). The difference of 44 bp per month (or approximately 5.3% per year) is not only economically meaningful, but also statistically highly significant with a t-value of 4. In the following, we refer to this finding as the media-based momentum effect. The appendix displays its full return distribution. The impact of the media is not confined to the extreme portfolios. The recently proposed monotonicity test of Patton and Timmermann (2010) verifies that momentum profits are indeed strictly increasing with excess media coverage. Notably, and in contrast to Hong et al. (2000), rising momentum profits are not primarily driven by the short leg of the portfolio. In fact, and in line with the reasoning in e.g. Barber and Odean (2008), the winner portfolio makes a stronger contribution than the loser portfolio (28 vs. -16 bp, t-stat 2.22 vs. -1.42). We next control for standard risk factors. Specifically, we employ a CAPM one factor model, a Fama and French (1993) three factor model, a Carhart (1997) four factor model as well as a six factor model augmented with the short-term and long-term reversal factor. Alphas reported in panel B of table 3 show that risk adjustment has only little impact on the size of the return difference. The highly significant intercepts range from 41 to 46 bp and are thus virtually unchanged.

4.2

Baseline analysis: Robustness tests

In this section, we document a variety of modifications to our primary specification. Given the endogeneity of media coverage, to a large extent these tests might be considered as first steps to control for additional factors that might be correlated with media coverage and which might also have an impact on momentum returns. Our findings are presented in table 4.

Insert table 4 here

The first set of robustness checks is intended to control for changes in the model for the estimation of media coverage (see panel A). We start by using a rolling average of residual media coverage based on articles not only from the skipped month, but also from the previous three or six months. As shown in specifications (1) and (2), extending the measurement period leads to even stronger results. To address the possibility that our results are driven by stock with zero media coverage, we exclude in (3) all stocks that were not covered in press during the period of interest. Reported are findings from specification (2), but relying on the baseline model yields similar results. It turns out that media-based momentum is about as strong as before, with both the statistical and economical significance monotonically increasing from portfolio 1 to 5. In specifications (4) and (5), we scale residual media coverage before the portfolio sorting takes place. In specification (4), we divide residual coverage by 1+ln(1+number of articles). In specification (5), we 10

instead rely on the standard deviation of the residual in the previous six months. Both procedures are intended to control for the impact firms with extreme coverage could have on our findings. It turns out that results again even become slightly stronger. In specification (6), we restrict the analysis to articles with a “relevance score” of 90% or above. Inferences do not change. We also include a number of additional factors in the OLS regressions used to estimate residual media coverage, namely dummies for size deciles (7), industry dummies (8), a firm’s book-to-market ratio (9), turnover (10), or the maximum daily return within the skipped month (11, see Bali et al. (2011) for a motivation). Size deciles dummies aim to control for potential non-linearities between media coverage and firm size that our baseline model does not capture. In fact, the analysis of the data for the baseline model shows a slight U-shaped relation: Firms in the lowest residual media coverage portfolio are on average second-largest, and firms in the highest coverage portfolio are on average largest (see the appendix for further details). Hence, while there are slight data imperfections, the difference in momentum returns in the baseline specification does not pick up a simple size effect. Adding size dummies to control for size more rigorously confirms this judgment. Regarding the other modifications, we also observe only modest changes compared to our baseline results. With industry dummies the return difference is slightly reduced (30 bp difference as compared to 44 bp before), but still significantly different from zero at the 1% threshold (t-stat of 3.04).4 As an alternative econometric way to control for heterogeneity at the firm level, we run a panel regression with firm-fixed and month-fixed effects in (12). Findings again are slightly weaker, but remain economically and statistically significant. Panel B displays the findings obtained when adding local newspaper articles to our database before computing residual media coverage. However, if local media cover primarily local stocks, certain geographic regions might receive systematically high or low residual media coverage values. To mitigate this issue, we report a second specification, which includes a total of 35 region dummies. Their construction will be explained in detail in section section 5.3. As specifications (13) and (14) show, the inclusion of local media articles has no substantial impact on the results either. Inferences also remain largely unchanged in specification (15), where we only consider local newspaper articles.5 4 We

have also explored cross-sectional differences of media-based momentum within (the 12 Fama/French) industries.

Findings are mostly in line with the idea that the effect should be stronger among hard to value, sentiment-prone industries. For instance, media-based momentum is stronger than on average within the “Business Equipment” (computers, software, electronic equipment) sector. On the other hand, it is much less pronounced for utility firms. The inclusion of industry dummies eliminates these effects, which might explain why findings are somewhat weaker than in our baseline analysis. Another way to control for industry effects is to run the baseline model without industry dummies, but then forming the five media portfolios within each industry separately. In the final step, low coverage and high coverage portfolios are pooled across industries. Here, media-based momentum ranges from 39 to 42 bp, and is thus closer to our baseline estimate. 5 Media-based

momentum turns out to be slightly weaker (though still highly significant) than in most other robustness

tests. Digging deeper, we find that this finding appears to be mostly rooted in the noise induced when measuring local (as

11

A third set of control specifications deals with changes in the sorting procedures and stock sample. We find that differences in results are rather minor and that the discrepancy in momentum returns is still highly significant with t-statistics above 3 if we rely on media deciles as breakpoints instead of quintiles (16), if we reverse the dependent sorting procedure for media coverage and formation returns (17), or if we conduct an independent sort to construct our portfolios (18). In addition, specification (19) verifies that our findings are not driven by stocks with the most extreme past returns. More specifically, repeating the exercise with winners defined as previous return deciles 8 and 9 and as with losers defined as deciles 2 and 3 leads to similar inferences. Finally, specifications (20) and (21) show results for the two subperiods from January 1989 to December 1999 as well as January 2000 to January 2010. The media-based momentum effect appears persistent, though statistically weaker in the more recent period with t-statistic of 2.02 considering the raw difference in momentum returns. However, momentum itself was much less successful in the 2000s with an average monthly spread between the winner and loser portfolio of just 23 bp (t-stat: 0.48). Hence, the difference in momentum returns of 30 bp that we obtain by conditioning on media coverage is economically still quite substantial. To sum up, the effect of media coverage on momentum returns remains remarkably robust to a large variety of methodological changes. However, we have not yet addressed the recent critique brought forward by Bandarchuk and Hilscher (2013). The authors revisit the existing literature on “enhanced momentum”, which typically sorts stocks first by certain stock-level characteristics (deemed to be proxy for behavioral or rational momentum drivers) and then by past returns to consequently document higher momentum profits. The analysis in Bandarchuk and Hilscher (2013) reveals that, in most cases, it is not the characteristic per se that matters, but its interaction pattern with extreme past returns. Adequately controlling for this relationship with a suitable regression approach causes the enhanced momentum profits seemingly stemming from size, R2 , turnover, analyst coverage and forecast dispersion, market-to-book ratio and illiquidity to vanish almost entirely. To address these findings, we imitate the Fama/MacBeth regression-based framework advocated in their study (see also Fama and French (2008)). Momentum profits are defined as a stock’s forward return above (below) the median, multiplied by a winner/loser dummy, which takes on a value of 1 (-1) if the stock was a winner (loser) in the formation period. This procedure ensures that both past winners with future positive returns and past losers with future negative returns will enter the regression with a positive sign. opposed to national) newspaper coverage only in the skipped month. If we rely on local coverage measured over the skipping month and the previous six months (see also specification (2)), media-based momentum increases to 40 bp per month. Inferences remain unchanged if we only consider the period starting from 1995, when most local newspapers are available. Findings also hold when we use other approaches than region dummies to control for systematic regional differences.

12

To infer to which extent extreme returns over the formation period cause high momentum profits, Bandarchuk and Hilscher (2013) calculate a return control variable, denoted as momentum strength. It is based on the absolute value of the difference between the stock’s (log) formation period return and the median formation period return of all stocks in the sample:

M omentum strengthi,t = exp(| ri,t−6;t−1 − rmedian,t−6;t−1 |) − 1.

(2)

Bandarchuk and Hilscher (2013) also acknowledge that many characteristics are related to idiosyncratic volatility and that higher levels of idiosyncratic volatility induce higher momentum profits. To control for this effect, they sort stocks in 25 portfolios based on idiosyncratic volatility, and use the assigned portfolio rank as a control in the Fama/MacBeth regressions (denoted as “IVOL rank”). Portfolio ranks are assigned to reduce noise in the data. We follow their set-up. In addition, we run multivariate regressions with a number of other firm characteristics (size, book-tomarket, analyst coverage, turnover, NASDAQ dummy, the illiquidity ratio of Amihud (2002), price, and industry dummies). The dependent variable is either the momentum profit in t + 1, as in Bandarchuk and Hilscher (2013), or the average momentum return over a six months period, which better corresponds to our baseline approach.6 Table 5 shows the results.

Insert table 5 here

Using the methodology of Bandarchuk and Hilscher (2013), we can univariately confirm the influence of residual media coverage on momentum returns (see specification I). We also re-confirm their finding that momentum strength and IVOL rank have a positive impact on momentum (specification II). Importantly, when using all three variables as predictors in specification III, we find that the coefficient of residual media coverage decreases only slightly and its level of statistical significance is not affected. This holds true in specifications IV and V where additional firm characteristics are included. Findings also remain economically meaningful. A one standard deviation change in excess media coverage is roughly associated with a 50 bp increase in momentum in the next month. 6 In

order to compute these momentum profits, we use the average of the returns in each of the following six months

minus the average median return as dependent variable. We do so for two reasons. First, average monthly returns are much more in line with the normality assumption than buy-and-hold returns. Second, taking averages better corresponds to the overlapping portfolio construction approach of Jegadeesh and Titman (1993), which we have used earlier. However, using buy-and-hold momentum profits as dependent variable results in very similar findings.

13

5

Overreaction to public information or dissemination of new information?

This section is devoted to the exploration of possible drivers of the media-based momentum effect in greater depth. The analysis proceeds in three steps. First, we test whether comprehensively controlling for important corporate events weakens the results. This should be the case if the publication of valuable information by the media, which then diffuses gradually among investors, was be the primary mechanism behind our findings. In this case, return differences between high and low coverage portfolios should also be permanent or even increase in the long run. We therefore test in a second step whether media-based momentum is followed by a reversal effect, which instead would be suggestive of investor overreaction. As the data lends credibility to this second hypothesis, our third step consists of studying the impact of cross-sectional differences in relevant investor characteristics on media-based momentum and subsequent return reversals.

5.1

The impact of corporate news

We start by gathering all quarterly earnings announcements for sample firms from Compustat. We then augment the data set with all 8-K forms provided by the SEC from 1995 on. Panel A of table 6 shows findings obtained from our baseline model of residual coverage over 1995 to 2010, which serves as a benchmark.

Insert table 6 here

Newspapers might be more likely to report on companies that recently had an earnings announcement, which might influence our residual coverage measures. Our findings could then be driven by the postearnings-announcement drift (e.g. Peress (2008)). A similar logic applies to other major corporate events, which should comprehensively be captured by the more than 315,000 8-K filings we gather from the EDGAR database. Across all firm months, there are 0.58 filings on average. The standard deviation is slightly more than 1, and the distribution is right-skewed. As for media articles, we therefore measure the monthly number of 8-K filings for a given firm as ln(1+number of documents). We control for the impact of corporate news in two ways. In panel B, we augment our baseline model for the estimation of residual media coverage with an earnings announcement dummy, with the monthly number of 8-K filings, or both. In panel C, we exclude all articles published in the seven day period centered around the event. Table 6 reveals that media-based momentum is hardly affected, no matter in which way we control for major corporate events. Findings are thus less suggestive of an underreactiondriven effect and instead point to an overreaction-driven phenomenon. 14

5.2

Long-run reversal

A central prediction of overreaction theories in the spirit of Daniel et al. (1998) is that medium-term momentum profits will reverse in the long run. We thus follow the previous literature in analyzing momentum profits for up to 36 months after portfolio formation. The following table displays the main results.

Insert table 7 here

The table shows average monthly momentum portfolio returns over months t+1 to t+12 (panel A), over t+13 to t+36 (panel B), and finally over the whole three year period from t+1 to t+36. A clear pattern emerges, which is also evident from figure 1 discussed in the introduction. Judging from a period of up to twelve months after formation, the high residual coverage momentum portfolio (still) outperforms the low coverage portfolio at the 5% significance level. This is primarily driven by the first six months, as can be seen from table 3. In the first year, the media-based momentum effect is about 20 bp per month on average. After slightly less than a year, however, the effect slowly starts to reverse. In years 2 and 3 after portfolio formation, there is a statistically significant reversal. The high coverage momentum portfolio now underperforms its low coverage counterpart. As can be seen from panel B, this reversal effect almost completely offsets the media-based momentum effect. Based on the whole three year period after portfolio formation, the difference is virtually zero. As can be seen from panel C, this finding holds for both raw momentum returns and after controlling for various well-known risk factors. In all specifications, monthly return differences between high and low coverage portfolios are estimated to be 2 bp or less when considering a holding period of three years. In sum, the analysis provides strong evidence that the media effect we document is primarily driven by investor overreaction.

5.3

Cross-sectional variation in individualism

A recent study by Chui et al. (2010) proposes an elegant way to distinguish between competing explanations for the momentum effect. The authors rely on the individualism index for 50 countries developed by Hofstede (2001) to explain differences in momentum profits around the world. While it is well documented that momentum strategies work very well in North America and Europe, they are unprofitable in many Asian countries, most notably in Japan. Building on a literature review of psychological research, Chui et al. (2010) first establish a link between the degree of individualism on the one hand and the two biases at the heart of the model in Daniel et al. 15

(1998) on the other: Overconfidence and self-attribution bias. For instance, the psychological literature suggests that people in individualistic cultures (e.g. the U.S. or the UK) tend much more to believe that their abilities are above average than do people in collectivistic cultures (such as Japan). This behavior, however, is a central characteristic of overconfidence. Given these insights, Chui et al. (2010) hypothesize and verify that individualism is strongly positively related to the profitability of countryspecific momentum strategies around the world. The authors end by highlighting the need of future research on the link between cultural differences and patterns in stock returns. We aim to progress on this front by transferring the idea from the cross-country to the within-country perspective. We thereby build on insights of two streams of research, which, to our knowledge, have not been brought together so far: Variation in collectivism/individualism within U.S. states, which is typically studied in the psychological literature, and the phenomenon of investors’ local bias, which is a recent prominent theme in the finance literature. While patterns related to individualism appear to be mostly examined in the U.S. versus Asian context, there is also measurable variation within the U.S. Vandello and Cohen (1999) have developed a state-level individualism/collectivism index based on eight indicators related to personal, social, economical, religious and political views and practices. Specifically, the index comprises the following items (partly reversed): Percentage of people living alone, percentage of elderly people living alone, percentage of households with grandchildren, divorce to marriage ratio, percentage of people without religious affiliation, ratio of people carpooling to work to driving alone, percentage of Libertarian votes over the last four presidential elections, percentage of self-employed people. It is generally understood that absolute differences in cultural norms only within the U.S., which is an individualistic country as a whole, might be less pronounced than in the international context. On the other hand, as Vandello and Cohen (1999) stress, “many of the problems of extraneous variation that make isolating cross-cultural differences challenging” (p. 290) can be mitigated, so that the information contained in the index should provide valuable insights and allow for useful individualism rankings. Indeed, the index, which contains collectivism scores for 50 U.S. states, is widely used in the psychological literature. However, it has been neglected in the finance literature so far. Notable exceptions are two recent study by Chen et al. (2012) and Truong et al. (2011). Chen et al. (2012) rely on the index to study geographic variation in CEO overconfidence and takeover gains. They show that, in more individualistic states, overconfidence is not only more pronounced, but also has a greater (negative) impact on takeover decisions. Truong et al. (2011) establish a significantly negative relationship between individualism and cash holdings of firms. In sum, these findings strongly confirm the usefulness of the collectivism index in a finance context. To our knowledge, however, we are the first to exploit the index in order to explain (geographic)

16

differences in return anomalies.7 The appendix displays state rankings of the U.S. collectivism index. It shows the intuitively appealing result that Southern states (particularly the Deep South) have greater collectivistic tendencies, whereas the Mountain West and the Great Plains tend to be rather individualistic. We follow the consensus in the literature and employ a firm’s headquarter as a proxy for its location. Based on this classification, the table also reveals that the number of firms and corresponding articles in both local and national newspapers varies greatly across states. We later address this issue by running a number of tests at different aggregation levels. In a different line of research, many recent finance papers have demonstrated that investors, both retail and professional, exhibit a strong preference for local stocks (e.g. Grinblatt and Keloharju (2001), Huberman and Regev (2001), Seasholes and Zhu (2010), Coval and Moskowitz (1999), Coval and Moskowitz (2001)). For our analysis, two aspects of this so-called local bias are especially important. First, recent findings highlight the role of (local) media in this context.8 Second, the impact of a local bias is not limited to individual decision making. The trading action of local investors have recently been shown to also affect economic aggregates such as the turnover of local stocks (e.g. Loughran and Schultz (2005), Jacobs and Weber (2012)), their prices (e.g. Hong et al. (2008)), and their returns (e.g. Shive (2012), Kumar et al. (2011)). Combining insights from the literature on individualism and local bias suggests the following: Provided that (differences in) local cultural norms indeed affect investment decisions, their impact on return patterns should be most easily identified in the behavior (differences) of local firms’ stocks. Consequently, a testable hypothesis for our setting is: We expect the media-based momentum effect to be particularly strong (weak) in more individualistic (collectivistic) areas. A distinctive feature of our study is that we can base our analysis on about 1.65 million news stories from a geographically widely dispersed cross-section of 41 local newspapers. It is crucial to take the role of local newspapers into account in a test based on cross-sectional differences in investor behavior along a geographical line. In the following sections, we therefore rely on the whole universe of local and 7 Given

the findings of Chui et al. (2010), one might hypothesize that “traditional” momentum strategies should work

better in individualistic US states. In unreported multivariate firm-level Fama/MacBeth regressions similar as in table 5, we find some supporting, though statistically insignificant, support for this idea over our sample period. However, replicating the exercise over an extended sample period starting in 1973, when the NASDAQ firms enter the sample, leads to considerably stronger findings. In most specifications, the collectivism score is then negatively significant at the 5% level. 8 Engelberg

and Parsons (2011) demonstrate that local media coverage triggers local trading activity. The presence or

absence of local newspaper articles about earnings releases of S&P 500 Index firms is strongly related to the magnitude of local trading. Gurun and Butler (2011) show that the local media uses fewer negative words in articles about local firms than they do in articles about non-local companies. The authors conjecture that this local media slant might influence local investors and create local bias.

17

national newspaper articles. In the light of the findings of Bandarchuk and Hilscher (2013), another benefit of our approach is that we do not rely on standard firm characteristics used in previous work to explain momentum profits. Instead, we exploit cross-sectional differences in investor characteristics to draw inferences. To explore the hypothesis outlined above, we first conduct portfolio-level tests as in our baseline analysis. The main idea here is to compute the media-based momentum effect separately for portfolios of firms located in individualistic and collectivistic areas and to compare the difference. Effectively, this type of analysis is based on triple sorts: Firms are sorted based on their residual media coverage, their formation period return, and their location. Implementing this idea leaves some degrees of freedom. For the sake of robustness, we thus run three specifications which differ in the way firm-specific residual media coverage is estimated and in the way the sorting procedures are carried out. Table 8 shows the main findings.

Insert table 8 here

In panel A, the model of residual coverage is the baseline model (see table 3) augmented with 35 region dummies included to capture unobservable effects at the regional level.9 They are deemed to control for potential cross-sectional differences in firm characteristics, industry composition as well as the number and availability of (local) newspapers. We consider the whole sample stock universe to run the regression. We then sort firms into three residual coverage portfolios: Those 30% of the firms with the highest coverage, 40% of firms in the middle, and those 30% of the firms with the lowest coverage. We consider only three instead of five media coverage portfolios, as we follow this procedure separately for collectivistic and individualistic areas. Both areas are defined in a way that they each comprise about 30% of sample firms. This amounts to an average of about 810 to 840 firms in both portfolios. In panel B, we rely on the baseline model (now without region dummies), which we again run for all sample firms. To nevertheless account for cross-sectional differences along a geographical line, the subsequent sorting procedure into three residual media coverage portfolios is carried out separately for each of the 35 regions. Equal-sized portfolios of firms from collectivistic and individualistic areas are then constructed 9 These

regions either represent single states or a collection of neighboring states with similar collectivism scores. Some

states such as Alaska, Hawaii, or Montana sometimes feature hardly any firms so that controlling for state-fixed effects would not be meaningful. More specifically, we switch to the regional level when two conditions are met: First, the state contains an average of less than 15 sample firms or less than 9 sample firms in any month. These cut-offs are a heuristic. The procedure aims at constructing a sufficient cross-section of geographic areas with similar collectivism scores and, at the same time, it aims at assuring that each area contains a meaningful number of firms. We have verified that a change in cut-offs does not change the quality of our findings. Second, there are neighboring states with a similar collectivism score, which do not meet our minimum requirements on sample firms either. The appendix gives more information about these regions, including descriptive statistics about the number of sample firms and the number of news stories.

18

from the pooled region portfolios. In panel C, we again rely on the baseline model (without region dummies). However, we now run both the regression and sorting procedure separately for firms in collectivistic and individualistic regions, which again account for 30% of all sample firms each. Table 8 shows a pervasive picture. In all specifications, the media-based momentum effect in collectivistic regions is comparatively low. The return difference between the high and the low residual coverage portfolios in these areas ranges from 19 to 26 bp per month, depending on the model employed. These values are well below our estimate in the baseline analysis (about 44 bp). Moreover, we can hardly reject the hypothesis that they are different from zero. In contrast, the media-based momentum effect in individualistic countries is exceptionally high. Estimates here range from 55 to 66 bp per month. These values are highly significant with a t-statistic of about 4. The difference in the media-based momentum effect between individualistic and collectivistic states consequently amounts to about 30 to 40 bp per month. The same holds true if one compares risk-adjusted returns obtained as intercepts from multifactor models. Depending on the exact specification, geographic differences in cultural norms appear to lead to an annualized media-based momentum difference of 3.5% to more than 5% per year. To see whether these findings are robust to additional methodological changes, we run a test similar in spirit to the analysis in Chui et al. (2010). We construct 15 portfolios of firms representing areas with increasing collectivism score. This number is chosen in order to assure that we have at least roughly 120 firms in each portfolio. This allows us to compute the media-based momentum effect within each of these areas, on which the appendix gives more information. More specifically, we now use a median split to sort firms on residual media coverage in each area. Momentum breakpoints are still the 30th and the 70th percentile. This means that we have an average of at least about 20 firms in each portfolio sorted on residual media coverage, formation period and collectivism area. If investor biases related to individualism cause our results, we would expect higher media-based momentum in individualistic areas. Indeed, this is exactly what figure 3 and table 9 show.

Insert figure 3 here

Insert table 9 here

Figure 3 displays average monthly media-based momentum profits in each of the 15 areas. While there is a considerable amount of noise due to the sometimes rather low number of firms in the portfolios, the pattern is obvious: Stronger individualism tends to go along with greater differences between high coverage momentum and low coverage momentum. Table 9 verifies this statement more formally. Here,

19

we run monthly Fama/MacBeth regressions of the media-based, area-specific momentum profits on the area-specific collectivism score and a number of control variables. In line with our expectations, the collectivism score persistently attains an (almost exclusively) significant coefficient. As a final test, we run a multi-variate firm-level regression analysis analogously to the one in table 5. We include the collectivism score and its interaction effect with residual media coverage in the regression. The appendix shows the results. In line with our hypothesis, we find that the interaction effect is significantly negative. This finding again confirms our insight from all tests in this analysis: As predicted by overconfidence models, the media-based momentum effect is positively related to investor individualism.

5.4

Cross-sectional variation mutual fund manager overconfidence

Our final test of the overreaction story relies on direct ownership data. Specifically, we use mutual fund holdings information for a sample of actively managed US-domestic equity funds. We obtain our fund sample by merging the CRSP Survivorship Bias Free Mutual Fund Database and the Thomson Reuters Mutual Fund Holdings database using MFLinks from Wharton Research Data Services (WRDS). Our final sample consists of 2,719 different active funds over the period from 1989 to 2010. A detailed description of the mutual fund data selection process is provided in the appendix. Our line of reasoning suggests that the media-based momentum effect should be stronger among those stocks that have a more overconfident investor base. To test this hypothesis, we rely on the mutual fund manager overconfidence index (denoted as OCI) recently developed by Nikolic (2011). The index is based on a set of twelve different characteristics that have been found to be related to overconfidence in prior literature (e.g. managerial gender, tenure, past performance, and portfolio concentration). Nikolic (2011) codes indicator variables (like gender) as zero-one. For all non-indicator variables (like past performance), she ranks managers into percentiles and assigns values from 0.1 to 1. Overall, the index can take values between one and twelve and a higher index score suggests that the manager is more overconfident. While it is a quarterly fund measure, Nikolic (2011) calculates a two-year moving average of OCI and we follow her approach.10 10 More

precisely, Nikolic (2011) uses the following overconfidence proxies: turnover, residual turnover (orthogonalized

with respect to year and fund size), Herfindahl concentration index of the fund’s portfolio positions, active share with respect to the market portfolio (see Cremers and Petajisto (2009)), prior 36-month four factor alpha, fund idiosyncratic risk, average market capitalization decile rank of the fund’s stock holdings, average book-to-market decile rank of the fund’s stock holdings, average idiosyncratic volatility decile rank of the fund’s stock holdings, manager gender, manager tenure, and an indicator variable indicating whether the fund is single- or team-managed. For details on the underlying motivation of using these variables as well as the construction, we refer to her paper. All our proxies are computed using data from the CRSP and Thomson Reuters mutual fund database except the gender and fund team variable. Gender is gratefully obtained from Niessen-Ruenzi and Ruenzi (2011), the team dummy from Baer et al. (2011). Whenever information for gender is missing, we assume that the manager is male, and whenever information for the team variable is missing, we enrich our data using information from Morningstar Direct. Morningstar Direct collects information on the management

20

We consider Nikolic’s (2011) overconfidence index as well suited for the analysis because it is a parsimonious and easy to calculate measure which also acknowledges the fact that every single characteristic is only a noisy proxy for overconfidence. In untabulated tests, we verify that our results are not materially influenced if we drop any one of the twelve characteristics from the index calculation. Panel A of table 10 provides descriptive statistics of the overconfidence index.

Insert table 10 here

Our analysis proceeds in three steps. First, we investigate whether overconfident fund managers particularly heavily buy stocks excessively covered by the media. Second, we try to assess whether overconfident managers interpret press articles more in line with their prior beliefs (and invest accordingly) than their less overconfident colleges do. Third, switching from the manager to the firm perspective, we test whether media-based momentum is particularly strong for firms predominantly held by overconfident fund managers. In doing so, our three-step procedure also extends and brings together recent work. These studies uncover a negative relationship between a manager’s propensity to buy stocks covered in the media and subsequent fund performance (Fang et al. (2011)) and fund manager overconfidence or self-attribution bias and subsequent performance (Choi and Lou (2010), Nikolic (2011), Puetz and Ruenzi (2011)), respectively. Fang et al. (2011) find that in aggregate, stocks are more likely to be bought by mutual funds if they have more press articles. However, there is also a substantial across-fund variation in this relation and funds whose buys are less affected by media coverage have higher returns. We augment the analysis of Fang et al. (2011) by studying the moderating influence of overconfidence in this context. Our hypothesis is that more overconfident managers have a stronger propensity to buy stocks with excessive media coverage. Support for this conjecture comes from the combined insights of the recent work sketched above. Empirically, we start by following Fang et al. (2011), in calculating the dollar value of a fund’s f buys for each stock i in a given quarter t:

$buyf,i,t = pricei,t · (nsharesf,i,t − nsharesf,i,t−1 ) if nsharesf,i,t >= nsharesf,i,t−1 .

(3)

Fang et al. (2011) argue and empirically verify that, due to funds managers’ inability to sell short and the attention-grabbing nature of mass-media coverage, only buying decisions, but not selling decisions of mutual fund managers might be affected by stock media coverage. For each fund-quarter with at least eight available observations we then regress (ln(1+buy)) on residual structure, but this information is not available in the time-series. We merge the Morningstar data to our fund data set using fund CUSIPs.

21

media coverage measured over the same quarter. Robustness checks verify that using lagged media coverage yields similar findings. More precisely, based on the monthly average of residual media coverage measured over the particular quarter, each stock is assigned a portfolio rank of one to five. Portfolio five here represents the 20% stocks with the highest residual press coverage in the four nation-wide newspapers. This first regression results in a fund-quarter-specific media beta. In the second step, we pool these values in each quarter and regress the cross-section of fund media betas on the fund manager overconfidence index OCI. This Fama/MacBeth approach results in a time-series of betas quantifying the relation between propensity to buy stocks excessively covered by the media and manager overconfidence. As the first row of panel B in table 10 verifies, overconfident managers indeed appear to more heavily rely on the media when it comes to the decision which stock to buy. A one standard increase in the overconfidence index goes along with a 0.03 standard deviation increase in the media beta. This finding is highly statistically significant with a t-value of more than 4. In sum, there is overall a positive relation between overconfidence and the likelihood to buy stocks excessively covered by the media. However, with reference to biased self-attribution, one would also expect that the impact of the media depends on the manager’s prior signal. To assess this empirically, we aim at inferring the managers prior beliefs from his trading decisions in the previous quarter. Specifically, if for a given fund manager f , stock i, and quarter t, the manager has effectively sold the stock in that quarter (i.e. if nsharesf,i,t < nsharesf,i,t−1 ), we assume that he has a negative prior about this stock during the following quarter. Otherwise, he is assumed to have a positive prior. We then perform a similar analysis as before, but now separately for firms with positive and negative prior. Findings are in line with our expectations: In the case of a positive (negative) prior, a one standard increase in the overconfidence index goes along with a 0.034 standard deviation increase (0.056 standard deviation decrease) in the media beta. This is consistent with the idea that particularly overconfident fund managers use the media to identify those articles that confirming their view and to make investment decisions accordingly. Those managers who excessively rely on the media thus not only seem to be particularly overconfident, but also to exhibit a particularly pronounced self-attribution bias. As these two characteristics are central in the model of Daniel et al. (1998), one might expect media-based momentum to be stronger among the stocks particularly held by these managers. We thus switch from the fund manager to the stock level by running Fama/MacBeth regressions analogously to table 5. More specifically, we run monthly cross-sectional regressions of average monthly stock-level momentum profits on the firm’s media coverage, on the value-weighted overconfidence index for managers holding that stock (see Nikolic (2011)), on the interaction effect, and on a number of controls. These control variables as well as the methodology is identical to the approach outlined in table 5. An analysis of the coefficients displayed in panel D of 10 shows that media-based momentum is indeed

22

stronger for stocks primarily held by overconfident fund managers. Consider for instance two stocks disproportionately present in the media (residual coverage portfolio rank 5), which differ only in their investor base. More precisely, imagine that, at the end of a given quarter (month t), one stock would be held by very overconfident fund managers (index value 9.13, 90th percentile), whereas the other stock would be held by much less overconfident managers (index value 5.6, 10th percentile). Findings suggest that the stock held predominantly by overconfident fund managers will ceteris paribus have a media-based momentum effect which is on average 14 bp per month higher during t+1 to t+6. Our analysis further suggests that the difference in month t+1 amounts to even 30 bp. These findings, however, should not be interpreted in the sense that overconfident managers on average make money by exploiting media-based momentum. As the appendix shows, the stocks heavily bought by overconfident fund managers also show stronger reversal effects. In the above mentioned example, the average monthly reversal is about 15 bp per months stronger between month t+13 and t+36. Moreover, in untabulated tests, we find that overconfident managers continue to heavily buy winner stocks with excess media coverage immediately before and during this return reversal period, which is likely to (at least) offset potential previous gains.

6

Conclusion

While the momentum effect is among the most prominent return anomalies, its underlying drivers are far from being well understood. We believe to provide some novel evidence to the ongoing debate by exploiting on a unique data set comprising about 2.2 million firm-specific articles from 45 newspapers. Our findings support the idea of a systematic link between the extent of a firm’s excess media coverage and the magnitude of the momentum effect in its stock. Results are broadly in line with the intuition of the overconfidence-driven momentum model in Daniel et al. (1998): Firms highly covered (neglected) by the media exhibit ceteris paribus a particularly strong (weak) momentum effect, followed by a long-term mean reversion. Our analysis thereby adds to the literature highlighting the importance of media for financial market outcomes. Motivated by psychological research on cross-sectional differences in cultural norms, we further test whether the effect also has a geographic dimension. Indeed, we find that media-based momentum is particularly pronounced (weak) in individualistic (collectivistic) regions within the United States. We thereby contribute to the rapidly emerging strand of literature arguing that cultural values can explain financial outcomes. Furthermore, our analysis suggests that overconfident mutual fund managers use media coverage to identify news that confirms their prior. Consequently, stock media coverage appears to heavily influence 23

their buying patterns. Also, media-based momentum is more pronounced among stocks disproportionately invested in by overconfident mutual fund managers. Our analysis thereby adds to the emerging debate on the potential links between manager overconfidence and self-attribution bias, fund performance, and the managers propensity to buy highly covered stocks. In sum, our analysis lends support to the idea that momentum is primarily an overreaction-driven phenomenon. In this way, return predictability can be strongest for those firms which stand particularly in the public spotlight. In a broader sense, our findings suggest that a joint analysis of cross-sectional differences in information transmission and investor biases might help to develop a better understanding of return anomalies.

24

References Amihud, Y., 2002, “Illiquidity and stock return: Cross-section and time-series effects,” Journal of Financial Markets, 5, 31–56. Antoniou, C., J. A. Doukas, and A. Subrahmanyam, 2011, “Sentiment and momentum,” Unpublished working paper, University of Exeter, University of Cambridge, University of California. Asness, C. S., T. J. Moskowitz, and L. H. Pedersen, 2012, “Value and momentum everywhere,” Journal of Finance, forthcoming. Avramov, D., T. Chorida, G. Jostova, and A. Philipov, 2007, “Momentum and credit rating,” Journal of Finance, 62, 2503–2520. Baer, M., A. Kempf, and S. Ruenzi, 2011, “Is a team different from the sum of its parts? Evidence from mutual fund managers,” Review of Finance, 15, 359–396. Bali, T. G., N. Cakici, and R. F. Whitelaw, 2011, “Maxing out: Stocks as lotteries and the cross-section of expected returns,” Journal of Financial Economics, 99, 427–446. Bandarchuk, P., and J. Hilscher, 2013, “Sources of momentum profits: Evidence on the irrelevance of characteristics,” Review of Finance, 17, 809–845. Barber, B. M., and T. Odean, 2008, “All that glitters: The effect of attention and news on the buying behavior of individual and institutional investors,” Review of Financial Studies, 21, 785–818. Barberis, N., A. Shleifer, and R. Vishny, 1998, “A model of investor sentiment,” Journal of Financial Economics, 49. Bhojraj, S., and B. Swaminathan, 2006, “Macromomentum: Return predictability in international equity indices,” Journal of Business, 79, 429–451. Carhart, M. M., 1997, “On persistence in mutual fund performance,” Journal of Finance, 52, 57–82. Chabot, B., E. Ghysels, and R. Jagannathan, 2009, “Price momentum in stocks: Insights from Victorian age data,” NBER Working Paper. Chan, K., A. Hameed, and W. Tong, 2000, “Profitability of momentum strategies in the international equity markets,” Journal of Financial and Quantitative Analysis, 35, 153–172. Chan, W. S., 2003, “Stock price reaction to news and no-news: Drift and reversal after headlines,” Journal of Financial Economics, 70, 223–260. Chen, Y., E. J. Podolski, G. S. Rhee, C. Truong, and M. Veeraraghavanc, 2012, “State individualism, CEO overconfidence and corporate takeovers,” Unpublished working paper, Monash University, University of Hawaii. 25

Choi, D., and D. Lou, 2010, “A test of the self-seving attribution bias: Evidence from mutual funds,” Unpublished working paper, Hong Kong University, London School of Economics. Chordia, T., and L. Shivakumar, 2002, “Momentum, business cycle and time-varying expected returns,” Journal of Finance, 57, 985–1019. Chui, A. C., S. Titman, and K. J. Wei, 2010, “Individualism and momentum around the world,” Journal of Finance, 65, 361–392. Connolly, R., and C. Stivers, 2003, “Momentum reversals in equity-index returns during periods of abnormal turnover and return dispersion,” Journal of Finance, 58, 1521–1555. Cooper, M. J., R. C. Gutierrez, and A. Hameed, 2004, “Market states and momentum,” Journal of Finance, 59, 1345–1365. Coval, J. D., and T. J. Moskowitz, 1999, “Home bias at home: Local equity preference in domestic portfolios,” Journal of Finance, 54, 2045–2073. , 2001, “The geography of investment: Informed trading and asset prices,” Journal of Political Economy, 109, 811–841. Cremers, K. M., and A. Petajisto, 2009, “How Active Is Your Fund Manager? A New Measure That Predicts Performance,” Review of Financial Studies, 22, 3329–3365. Da, Z., J. Engelberg, and P. Gao, 2011, “In search of attention,” Journal of Finance, 66, 1461–1499. Da, Z., U. G. Gurun, and M. Warachka, 2012, “Frog in the pan: Continuous information and momentum,” Unpublished working paper, University of Notre Dame, University of Texas, Singapore Management University. Daniel, K., D. Hirshleifer, and A. Subrahmanyam, 1998, “Investor psychology and security market underand overreactions,” Journal of Finance, 53, 1839–1885. D’Avolio, G., 2002, “The market for borrowing stock,” Journal of Financial Economics, 66, 271–306. Dougal, C., J. Engelberg, D. Garcia, and C. A. Parsons, 2011, “Journalists and the stock market,” Review of Financial Studies, 25, 639–679. Engelberg, J., and C. A. Parsons, 2011, “The causal impact of media in financial markets,” Journal of Finance, 66, 67–97. Engelberg, J., C. Sasseville, and J. Williams, 2012, “Market madness? The case of mad money,” Management Science, 58, 351–364. Erb, C. B., and C. R. Harvey, 2006, “The strategic and tactical value of commodity futures,” Financial Analysts Journal, 62, 69–97.

26

Fama, E., and K. R. French, 2008, “Dissecting Anomalies,” Journal of Finance, 63, 1653–1678. , 2012, “Size, value, and momentum in international stock returns,” Journal of Financial Economics, 105, 457–472. Fama, E. F., and K. R. French, 1993, “Common risk factors in the returns on stocks and bonds,” Journal of Financial Economics, 33, 3–56. Fama, E. F., and J. D. MacBeth, 1973, “Risk, return, and equilibrium: Empirical tests,” Journal of Political Economy, 81, 607–636. Fang, L., and J. Peress, 2009, “Media coverage and the cross-section of stock returns,” Journal of Finance, 64, 2023–2052. Fang, L. H., J. Peress, and L. Zheng, 2011, “Does Media Coverage of Stocks Affect Mutual Funds’ Trading and Performance?,” Unpublished working paper, Insead. Gervais, S., and T. Odean, 2001, “Learning to be overconfident,” Review of Financial Studies, 14, 1–27. Grinblatt, M., and B. Han, 2005, “Prospect theory, mental accounting, and momentum,” Journal of Financial Economics, 78, 311–339. Grinblatt, M., and M. Keloharju, 2001, “How distance, language, and culture influence stockholdings and trades,” Journal of Finance, 56, 1053–1073. Gurun, U. G., and A. W. Butler, 2011, “Don’t believe the hype: Local media slant, local advertising, and firm value,” Journal of Finance, 67, 561–598. Hofstede, G., 2001, “Cultures consequences: Comparing values, behaviors, institutions, and organizations across nations,” Sage Publication, Beverly Hills, CA, 2nd edition. Hong, H., J. D. Kubik, and J. C. Stein, 2008, “The only game in town: Stock-price consequences of local bias,” Journal of Financial Economics, 90, 20–37. Hong, H. G., T. Lim, and J. C. Stein, 2000, “Bad news travels slowly: Size, analyst coverage, and the profitability of momentum strategies,” Journal of Finance, 55, 265–295. Hong, H. G., and J. C. Stein, 1999, “A unified theory of underreaction, momentum trading, and overreaction in asset markets,” Journal of Finance, 54, 2143–2184. , 2007, “Disagreement and the stock market,” Journal of Political Economy, 36, 3–27. Hou, K., L. Peng, and W. Xiong, 2009, “A tale of two anomalies: The implications of investor attention for price and earnings momentum,” Unpublished working paper, Ohio State University, City University of New York, Princeton University.

27

Huberman, G., and T. Regev, 2001, “Contagious speculation and a cure for cancer: A non-event that made stock prices soar,” Journal of Finance, 56, 387–396. Jacobs, H., and M. Weber, 2012, “The trading volume impact of local bias: Evidence from a natural experiment,” Review of Finance, 16, 867–901. Jegadeesh, N., and S. Titman, 1993, “Returns to buying winners and selling losers: Implications for stock market efficiency,” Journal of Finance, 48, 65–91. , 2001, “Profitability of momentum strategies: An evaluation of alternative explanations,” Journal of Finance, 56, 699–720. Klibanoff, P., O. Lamont, and T. A. Wizman, 1998, “Investor reaction to salient news in closed-end country funds,” Journal of Finance, 53, 673–699. Kumar, A., J. Page, and O. Spalt, 2011, “Religious beliefs, gambling attitudes, and financial market outcomes,” Journal of Financial Economics, 102, 67–708. Liew, J., and M. Vassalou, 2000, “Can book-to-market, size and momentum be risk factors that predict economic growth?,” Journal of Financial Economics, 57, 221–245. Long, J. B. D., A. Shleifer, L. H. Summers, and R. J. Waldmann, 1990, “Noise trader risk in financial markets,” Journal of Political Economy, 98, 703–738. Loughran, T., and P. Schultz, 2005, “Liquidity: Urban versus rural firms,” Journal of Financial Economics, 78, 341–374. Moskowitz, T., Y. H. Ooi, and L. H. Pedersen, 2012, “Time series momentum,” Journal of Financial Economics, 104, 228–250. Niessen-Ruenzi, A., and S. Ruenzi, 2011, “Sex matters: Gender and prejudice in the mutual fund industry,” Unpublished working paper, University of Mannheim. Nikolic, B., 2011, “Momentum, Reversal, and Investor Overconfidence: An Empirical Investigation of Mutual Fund Managers,” Unpublished working paper, Robert J. Trulaske Sr. College of Business, University of Missouri. Park, J., A. Kumar, and R. Raghunathan, 2013, “Confirmation bias, overconfidence, and investment performance: Evidence from stock message boards,” Information Systems Research, forthcoming. Patton, A., and A. Timmermann, 2010, “Monotonicity in asset returns: New tests with applications to the term structure, the CAPM and portfolios sorts,” Journal of Financial Economics, 98, 605–625. Peress, J., 2008, “Media coverage and investors’ attention to earnings announcements,” Unpublished working paper, Insead.

28

Puetz, A., and S. Ruenzi, 2011, “Overconfidence among professional investors: Evidence from mutual fund managers,” Journal of Business Finance and Accounting, 38, 684–712. Rouwenhorst, G. K., 1998, “International momentum strategies,” Journal of Finance, 53, 267–284. , 1999, “Local return factors and turnover in emerging stock markets,” Journal of Finance, 54, 1439–1464. Sagi, J. S., and M. S. Seasholes, 2007, “Firm-specific attributes and the cross-section of momentum,” Journal of Financial Economics, 84, 389–434. Seasholes, M. S., and N. Zhu, 2010, “Individual investors and local bias,” Journal of Finance, 65, 1987– 2011. Shive, S. A., 2012, “Local investors, price discovery and market efficiency,” Journal of Financial Economics, 101, 145–161. Sinha, N. R., 2011, “Underreaction to news in the US stock market,” Unpublished working paper, University of Illinois. Stambaugh, R. F., J. Yu, and Y. Yuan, 2012, “The short of it: Investor sentiment and anomalies,” Journal of Financial Economics, 104, 288–302. Stivers, C., and L. Sun, 2010, “Cross-sectional return dispersion and time variation in value and momentum premiums,” Journal of Financial and Quantitative Analysis, 45, 987–1014. Tetlock, P. C., 2007, “Giving content to investor sentiment: The role of media in the stock market,” Journal of Finance, 62, 1139–1168. , 2010, “Does public financial news resolve asymmetric information?,” Review of Financial Studies, 23, 3520–3557. , 2011, “All the news that’s fit to reprint: Do investors react to stale information?,” Review of Financial Studies, 24, 1481–1512. Truong, C., Y. Chen, Y. Dou, S. G. Rhee, and M. Veeraraghavan, 2011, “National culture and cash holdings around the world,” Unpublished working paper, Monash University, Macquarie University, University of Hawaii. Vandello, J. A., and D. Cohen, 1999, “Patterns of individualism and collectivism across the United States,” Journal of Personality and Social Psychology, 77, 279–292. West, K. D., and W. K. Newey, 1987, “A simple, positive semi-definite, heteroskedasticity and autocorrelation consistent covariance matrix,” Econometrica, 55, 703–708.

29

Figure 1: Buy-and-hold long-term momentum profits for firms with high and low residual media coverage This figure shows the cumulative buy-and-hold returns for a winner minus loser long-short strategy separately for two different stock categories. The first category (red solid line) compiles stocks for the long and short portfolio from the universe of stocks in the highest residual media coverage quintile, and the second category (blue dashed line) consists of stocks falling in the lowest residual media coverage quintile. Residual media coverage is obtained from a month-to-month rolling OLS regression with ln (1+number of media articles) as dependent variable and size, analyst coverage, and dummies for S&P 500 and NASDAQ membership as explanatory variables. For the portfolio construction, stocks are first divided into residual media coverage quintiles and then further separated into a loser portfolio (30% of stocks with the lowest return over the previous six-month formation period), a neutral portfolio, and a winner portfolio (30% of stocks with the highest return over the formation period).

Cumulative return in %

8.00% 6.00% 4.00% 2.00% 0.00% 0

6

12

18

24

-2.00%

-4.00% -6.00%

Number of months after formation period Momentum Low Residual Media Coverage Quintile Momentum High Residual Media Coverage Quintile

30

30

36

Figure 2: Percentage of firms with a least one article in national, local, and all newspapers This figure shows the time-series evolution of the percentage of firms with press coverage separated by national, local, and all newspapers in a given year. The total media data set consists of 45 different newspapers with weekday circulation, of which the four major U.S. newspapers New York Times, USA Today, Wall Street Journal, and Washington Post make up the national newspaper data set. For details about local newspapers, see table 1. 80.00% 70.00% 60.00% 50.00% 40.00% 30.00% 20.00% 10.00% 0.00% 1989

1991

1993

1995

All newspapers

1997

1999

Local newspapers

31

2001

2003

2005

National newspapers

2007

2009

Figure 3: Individualism and media-based momentum This graph compares the media-based momentum effect across 15 areas sorted by the collectivism score of Vandello and Cohen (1999) in ascending order. Details about the computation are given in the text. The 15 areas contain the following states: Area ID

States

1

Washington, North Dakota, Colorado, South Dakota, Wyoming, Nebraska, Oregon, Montana

2

Oklahoma, Idaho, Vermont, Minnesota, Iowa, Kansas

3

Ohio, Maine, New Hampshire

4

Massachusetts, Missouri

5

Michigan, Wisconsin

6

New Mexico, Connecticut, Arizona, Alaska, Rhode Island, West Virginia

7

Nevada, Illinois

8

Kentucky, Pennsylvania

9

New York

10

Delaware, Arkansas, Florida

11

Alabama, Indiana, North Carolina, Tennessee

12

New Jersey, Texas

13

California

14

Virginia, Georgia

15

Hawaii, Louisiana, South Carolina, Mississippi, Maryland, Utah

32

Table 1: Summary statistics for the newspaper data set

This table reports data on availability, total number of articles, average number of articles per year, the average percentage of firms covered in a given year, as well as average weekday circulation for each of the local newspapers and national newspapers (in bold) that we use in the analysis. Articles are obtained from LexisNexis using the company search function and a “relevance score” of at least 80%. Availability is the maximum of the start of our sample period (1989) and the first year of complete journal coverage. To calculate the average percentage of covered firms in a given year (last column), we first calculate the percentage of coverage separately for each year in which the newspaper is available. This number is based on all firms that appear in our sample at any point in time during that specific year. We then take the times-series (i.e. yearly) average. Average weekday circulation is based on circulation data for the years 2004, 2006 and 2009 which was compiled by BurrellesLuce using the Audit Bureau of Circulations as data source. See http://www.burrellesluce.com. For Augusta Chronicle and Santa Fe New Mexican we obtain circulation data directly from the Audit Bureau of Circulations.

Newspaper

Availability

Total no.

Mean no.

% of firms

Average

articles

articles per

covered in

weekday

year

given year

circulation

New York Times

1989-2010

201,035

9,138

26.59%

1,100,174

Washington Post

1989-2010

93,548

4,252

12.45%

716,553

Wall Street Journal

1989-2010

214,617

9,755

37.51%

2,077,664

USA Today

1989-2010

52,903

2,405

7.16%

2,192,879

Arkansas Democrat Gazette

1989-2010

33,353

1,667

6.48%

182,859

Atlanta Journal and Constitution

1991-2010

73,899

3,695

10.99%

345,571

Augusta Chronicle

1993-2010

46,443

2,576

8.08%

59,243

Austin American Statesman

1994-2010

46,472

2,734

9.46%

173,850

Birmingham News

1993-2010

19,067

1,059

6.18%

146,605

Boston Herald

1992-2010

35,869

1,863

8.12%

205,057

Buffalo News

1993-2010

43,726

2,394

9.12%

188,113

Chicago Sun Times

1992-2010

89,083

4,689

13.12%

377,987

Daily News New York

1995-2010

20,269

1,267

4.54%

674,668

Dallas Morning News

1993-2010

89,696

4,913

12.87%

446,923

Dayton Daily News

1994-2010

25,452

1,497

6.55%

125,376

Denver Post

1994-2010

30,763

1,809

7.79%

322,449

Fresno Bee

1994-2010

8,260

486

3.39%

157,422

Houston Chronicle

1992-2010

75,115

3,941

11.71%

495,942

Las Vegas Review Journal

1997-2010

9,462

667

3.93%

178,722

Minneapolis Star Tribune

1992-2010

27,353

1,420

6.67%

353,366

[continued overleaf ]

33

Table 1: Summary statistics for the newspaper data set (continued) Newspaper

Availability

Total no.

Mean no.

% of firms

Average

articles

articles per

covered in

weekday

year

given year

circulation

New Orleans Times Picayune

1993-2010

34,125

1,895

7.56%

214,053

New York Post

1998-2010

28,309

2,159

7.28%

624,788

Palm Beach Post

1989-2010

26,618

1,210

5.84%

168,693

Pittsburgh Post Gazette

1993-2010

45,740

3,042

11.59%

224,973

Richmond Times-Dispatch

1996-2010

26,612

1,759

8.41%

178,920

Sacramento Bee

2002-2010

10,399

1,155

6.88%

281,962

Salt Lake Tribune

1995-2010

21,471

1,263

6.94%

128,631

San Antonio Express News

1996-2010

32,017

2,134

8.83%

225,712

San Francisco Chronicle

1990-2010

48,554

2,279

8.45%

403,124

San Jose Mercury News

1994-2010

74,565

4,386

10.81%

249,841

Santa Fe New Mexican

1995-2010

3,724

232

2.30%

21,245

Seattle Post Intelligencer

1990-2008

53,094

2,774

8.81%

139,818

St. Louis Post Dispatch

1989-2010

95,879

4,358

13.30%

265,813

St. Petersburg Times

1989-2010

55,964

2,544

9.58%

318,209

Star Ledger (Newark, NJ)

1996-2010

35,546

2,734

10.48%

360,137

Tampa Tribune

1998-2010

16,512

1,269

7.34%

223,510

The Oklahoman

1992-2010

29,057

1,529

7.00%

189,827

The Oregonian

1989-2010

49,032

2,229

7.32%

310,233

The Philadelphia Inquirer

1994-2010

52,082

3,064

11.68%

338,394

The Plain Dealer

1993-2010

60,984

3,353

10.17%

334,107

The Providence Journal

1994-2010

31,861

1,874

8.78%

146,798

The Record (Bergen)

1996-2009

51,037

3,594

12.98%

176,701

The Virginian Pilot

1994-2010

17,654

1,038

6.37%

186,931

Tulsa World

1996-2010

55,496

3,695

13.04%

134,668

Wisconsin State Journal

1992-2010

19,633

1,033

6.15%

96,927

Total number of articles Total number of articles in WSJ, WP, NYT, USAT

34

2,212,350 562,103

Table 2: Correlations between media coverage and firm Variables

This table presents time-series averages of monthly cross-sectional correlations between media coverage and a list of firm variables. Media coverage is based on articles written on a firm in one of the following four U.S. newspapers: New York Times, USA Today, Wall Street Journal, and Washington Post. Articles are obtained from LexisNexis using the company search function and a “relevance score” of at least 80%. Panel A shows the results for raw media coverage which is - due to the skewness of the media data - calculated as the natural log of (1+ number of articles). Panel B reports correlation results for residual media coverage, where the residuals are obtained from rolling month-to-month OLS regressions of raw media coverage on firm size, S&P 500, NASDAQ, and analyst coverage. Firm size is the natural log of market capitalization. S&P 500 and NASDAQ are dummy variables which take the value of 1 when the firm is a member of the S&P 500 index or is listed on the NASDAQ. Analyst coverage is the natural log of (1+number of earnings estimates). Book-to-market is the firm’s equity book-to-market ratio, turnover is the average share volume divided by shares outstanding and IVOL is the residual of a Fama/French three factor model. For both, turnover and IVOL, we use daily data over the last six months

35

and take the natural log. AbsRetpr6 is the absolute value of the stock’s return over the previous six months. Price is the share price in CRSP, Amihud is the Amihud (2002) illiquidity ratio, and firm age is based on the first appearance of the stock’s permco in CRSP. Further construction details are in the appendix. The sample period covers M1:1989-M12:2010. Correlations media variables Firm

Book-to-

Variables

size

market

S&P 500

Analyst

Raw media

0.49

-0.04

0.43

Residual media

0.00

0.02

0.00

NASDAQ

Turnover

coverage

Raw

Residual

IVOL

AbsRetpr6

Price

Amihud

Age

media

media

-0.15

-0.04

0.21

-0.42

0.18

1.00

0.86

0.06

-0.08

0.00

-0.01

Panel A: Raw media coverage -0.20

0.13

0.33

Panel B: Residual media coverage in skipped month 0.00

0.04

0.00

0.10

1.00

Table 3: Residual media coverage and momentum: Baseline results

This table presents momentum returns for stock portfolios sorted on residual media coverage. To construct the momentum portfolios we use a formation and holding period of six months each and skip one month in between. Momentum profits are derived from a dependent double sorting procedure where stocks are first sorted into five equally sized residual media coverage portfolios. Residual media coverage is computed from rolling crosssectional OLS regressions of ln (1+number of articles in the skipped month) on firm size, analyst coverage and dummies for S&P 500 and NASDAQ membership. Within each quintile of residual media coverage, the winner (loser) portfolio then consists of all stocks with a return during the formation period above the 70th percentile (below the 30th percentile). Momentum returns (reported in % per month) are based on overlapping portfolios which are equally weighted as in the original work of Jegadeesh and Titman (1993). The sample period covers M1:1989-M12:2010. Panel A shows raw returns, panel B risk-adjusted returns. 1F refers to the CAPM, 3F to the Fama and French (1993) model, 4F to the Carhart (1997) model, and 6F to a model augmented with factors for short-term and long-term reversal. T-statistics (in parentheses) are adjusted for serial autocorrelation using West and Newey (1987) standard errors with a lag of five months. * indicates significance at the 10% level, ** indicates significance at the 5% level and *** indicates significance at the 1% level.

Residual media

Loser

Winner

Momentum

Momentum

coverage portfolios

return

return

return

t-stat

Panel A: Double sorts and raw returns 1

0.77

1.18

0.41

(1.61)

2

0.70

1.12

0.43*

(1.80)

3

0.70

1.30

0.59**

(2.36)

4

0.61

1.43

0.81***

(3.30)

5

0.61

1.46

0.85***

(3.13)

Return 5-1

-0.16

0.28**

0.44***

(-1.42)

(2.22)

(4.00)

t-stat 5-1

Panel B: Double sorts and risk-adjusted returns Factor model

1F

3F

4F

6F

N

264

264

264

264

0.41***

0.46***

0.42***

0.43***

(3.44)

(4.12)

(3.86)

(3.85)

Intercept 5-1 Intercept t-stat

36

Table 4: Residual media coverage and momentum: Robustness results

This table presents the results of various robustness tests in which we modify our baseline specification. In (1) stocks are sorted into media quintiles based on the average residual from the OLS media coverage regression in the skipped month and the previous three months. Similarly, in (2) the media sorting variable is the average residual over the skipped month and the previous six months (i.e. the formation period). In (3) we exclude all stocks which had zero media coverage over the skipped month and the formation period. Residual media coverage is then computed for all stocks with at least one press article over that period. In (4) and (5) we scale residual media coverage by 1+ln(1+ number of articles) and the standard deviation of the residual in t-6 to t-1, respectively. In (6), we repeat the analysis counting only articles with a LexisNexis “relevance score” of 90% or more. In (7) we augment the basis OLS model with dummies for size deciles in order to estimate residual media coverage. (8), (9), (10), and (11) report our results for further adjustments to the OLS estimation, particularly the inclusion of Fama/French 48 industry dummies, book-to-market, turnover, and maximum daily return (within the skipped month) as independent variables. In (12), we run a panel regression with firm-fixed and month-fixed effects. In (13) we re-estimate residual media coverage using articles in both national and local audiences. (14) uses local and national newspaper reports, and additionally controls for regional differences with 35 region dummies in the regression. In (15) we only consider local newspaper reports, and again run the regression with 35 region dummies. (16) summarizes momentum differences between top and bottom media deciles instead of quintiles. (17) reverses the sorting procedure (first formation period returns and second residual media coverage) and (18) employs an independent sorting procedure. In (19) momentum portfolios are constructed excluding stocks with the most extreme formation period returns (return deciles one and ten). Finally, (20) and (21) report the results for two subperiods of the entire sample (M1:1989-M12:1999 and M1:2000-M12:2010). The variable of interest is the difference in momentum returns (top 30% winner portfolio minus bottom 30% loser portfolio) between the highest and the lowest residual media coverage quintile. Momentum returns reported (in % per month) are based on overlapping portfolios which are equally weighted as in the original work of Jegadeesh and Titman (1993). We report the raw return difference as well as the corresponding risk-adjusted return differences. 1F refers to the CAPM, 3F to the Fama and French (1993) model, 4F to the Carhart (1997) model and 6F to a model augmented with factors for short-term and long-term reversal. T-statistics (in parentheses) are adjusted for serial autocorrelation using West and Newey (1987) standard errors with a lag of five months.* indicates significance at the 10% level, ** indicates significance at the 5% level and *** indicates significance at the 1% level. [continued overleaf ]

37

Table 4: Residual media coverage and momentum: Robustness results (continued) Robustness Specification

Raw

1F

3F

4F

6F

difference

alpha

alpha

alpha

alpha

Panel A: Changes in model estimation and calculation of residual media coverage 1) Residual media coverage 3+1

0.46***

0.38***

0.47***

0.45***

(3.45)

(2.66)

(3.52)

(3.43)

0.47*** (3.51)

2) Residual media coverage 6+1

0.58***

0.46***

0.56***

0.57***

0.57***

(3.65)

(2.77)

(3.8)

(3.72)

(3.75)

3) Excluding firms with zero coverage

0.54***

0.46***

0.57***

0.55***

0.57***

(3.04)

(2.52)

(3.46)

(3.39)

(3.48)

4) Scaled by 1+ln(1+ number of articles)

0.45***

0.42***

0.47***

0.43***

0.44***

(4.02)

(3.46)

(4.10)

(3.90)

(3.88)

5) Scaled by vol(residual) in t-6 to t-1

0.47***

0.41***

0.48***

0.45***

0.45***

(4.07)

(3.29)

(4.22)

(3.76)

(3.70)

6) Relevance Score 90+

0.43***

0.40***

0.44***

0.41***

0.42***

(3.90)

(3.34)

(3.84)

(3.64)

(3.61)

7) Basis model + size decile dummies

0.38***

0.37***

0.41***

0.33***

0.33***

(3.37)

(3.07)

(3.63)

(3.01)

(3.03)

8) Basis model + industry dummies

0.30***

0.25***

0.28***

0.24***

0.25***

(3.04)

(2.68)

(3.02)

(2.71)

(2.74)

9) Basis model + book-to-market

0.48***

0.45***

0.49***

0.45***

0.46***

(4.32)

(3.82)

(4.42)

(4.05)

(4.08)

10) Basis model + turnover

0.36***

0.34***

0.37***

0.36***

0.37***

(3.46)

(2.93)

(3.42)

(3.35)

(3.37)

11) Basis model + maximum daily return

0.34***

0.30***

0.33***

0.40***

0.41***

(3.37)

(2.69)

(3.09)

(3.83)

(3.86)

12) Panel regression with firm-fixed effects

0.29***

0.24**

0.27**

0.28**

0.27**

(2.63)

(2.01)

(2.41)

(2.37)

(2.42)

0.45***

Panel B: Inclusion of local media articles 13) National and local media articles 14) National and local media articles with region dummies 15) Local media articles with region dummies

0.42***

0.41***

0.43***

0.42***

(3.73)

(3.35)

(3.87)

(3.82)

(3.88)

0.39***

0.38***

0.41***

0.41***

0.43***

(3.54)

(3.24)

(3.7)

(3.78)

(3.79)

0.32***

0.31**

0.34***

0.34***

0.35***

(2.83)

(2.54)

(2.89)

(2.93)

(2.95)

0.50***

Panel C: Changes in sorting procedure and stock sample 16) Media deciles

0.48***

0.47***

0.54***

0.48***

(3.13)

(3.03)

(3.89)

(3.51)

(3.59)

17) Reverse dependent sort

0.36***

0.30***

0.35***

0.38***

0.39***

(3.52)

(2.83)

(3.51)

(3.87)

(3.92)

18) Independent sort

0.35***

0.30***

0.35***

0.38***

0.40***

(3.20)

(2.62)

(3.25)

(3.74)

(3.76)

19) Excluding return deciles 1 and 10

0.37***

0.38***

0.40***

0.31***

0.33***

(3.75)

(3.46)

(4.04)

(3.12)

(3.39)

20) Subperiod M1:1989-M12:1999

0.58***

0.48***

0.50***

0.47***

0.48***

(3.73)

(3.39)

(3.83)

(3.27)

(3.22)

21) Subperiod M1:2000-M12:2010

0.30**

0.30*

0.32**

0.32**

0.29*

(2.02)

(1.90)

(2.28)

(2.25)

(1.95)

38

Table 5: Momentum, residual media coverage, and other stock characteristics: Fama/MacBeth regressions

Following Bandarchuk and Hilscher (2013) we run Fama/MacBeth regressions of momentum profits on stock characteristics, idiosyncratic volatility (denoted as IVOL rank ), and extreme past returns (denoted as Momentum strength). Stock characteristics include residual media coverage as well as a set of additional factors that have been shown to impact the magnitude of momentum returns in previous work (size, book-to-market, analyst coverage, turnover, firm age, a NASDAQ dummy, the Amihud (2002) illiquidity ratio, and the stock’s share price). Momentum strength is calculated as exp(absolute difference between a stock’s formation period log return and the median formation period log return of all stocks in the sample)-1. As in Bandarchuk and Hilscher (2013) we use log returns to achieve comparability in returns for extreme winners and extreme losers. IVOL is the residual return from a month-to-month rolling Fama and French (1993) three factor regression using daily data. The residuals are then averaged over the six month formation period. We sort stocks into 25 different volatility portfolios and use the resulting rank as independent variable (1-25). Momentum return, the dependent variable, is the stock’s forward return minus the sample median return, multiplied with a dummy variable being 1 (-1) if the stock was a winner (loser). This calculation is also based on Bandarchuk and Hilscher (2013). Regressions aim at predicting the momentum return in t+1 (left-hand side) and the average momentum return in t+1 to t+6 (right-hand side). The sample period covers M1:1989-M12:2010. T-statistics (in parentheses) are adjusted for serial autocorrelation using West and Newey (1987) standard errors with a lag of five months.* indicates significance at the 10% level, ** indicates significance at the 5% level and *** indicates significance at the 1% level.

Variable Residual media coverage

I 0.12*** (3.66)

Momentum return in t+1 II III IV 0.09*** 0.10*** (2.73) (2.97)

V 0.09*** (3.16)

Average momentum return from t+1 to t+6 I II III IV V 0.09*** 0.08*** 0.07*** 0.06*** (3.86) (3.36) (3.48) (2.98)

0.43** (2.34)

0.42** (2.31)

0.39** (2.24)

0.39** (2.25)

0.35** (2.33)

0.34** (2.30)

0.32** (2.21)

0.31** (2.19)

IVOL (rank)

0.03*** (5.43)

0.03*** (5.37)

0.03*** (2.71)

0.02*** (2.71)

0.01*** (3.77)

0.01*** (3.64)

0.00 (0.79)

0.01 (1.08)

Firm size

-0.16 (-1.57)

-0.18* (-1.83)

0.10 (1.16)

0.07 (1.02)

Book-market

0.02 (0.42)

0.05 (0.93)

-0.03 (-0.86)

-0.02 (-0.47)

Analyst coverage

-0.05 (-1.02)

-0.02 (-0.40)

0.01 (0.2)

0.03 (0.85)

Turnover

-0.08 (-0.77)

-0.09 (-0.88)

0.12* (1.75)

0.11* (1.76)

-0.12*** (-4.02)

-0.11*** (-4.01)

-0.08*** (-3.97)

-0.06*** (-4.03)

NASDAQ dummy

-0.08 (-1.57)

-0.08* (-1.80)

-0.02 (-0.53)

-0.03 (-0.96)

Amihud

-0.08 (-1.11)

-0.09 (-1.27)

0.11 (1.64)

0.09* (1.67)

Price

0.24*** (2.7)

0.21*** (2.56)

0.07 (1.16)

0.08 (1.43)

no

yes

no

yes

39

Mom strength

Age

Fama-French 48 industry dummies

no

no

no

no

no

no

Table 6: The impact of corporate news on media-based momentum

This table presents the results of various robustness tests in which we modify our baseline specification of residual coverage to control for firm events. To facilitate comparability, panel A shows findings obtained from the baseline model over 1995 to 2010, which is also the sample period in panels B and C. In panel B, we augment the baseline model with an earnings announcement dummy which is one if a firm reported earnings in the month under consideration. Alternatively, we augment the model with ln(1+number of 8-K filings) in that month. The panel also shows findings when we include both controls. In panel C, we exclude all articles published in the seven-day period centered around firm earnings announcements, 8-K filings, or both, when we compute residual media coverage for that month. The variable of interest is the difference in momentum returns (top 30% winner portfolio minus bottom 30% loser portfolio) between the highest and the lowest residual media coverage quintile. Momentum returns reported (in % per month) are based on overlapping portfolios which are equally weighted as in the original work of Jegadeesh and Titman (1993). We report the raw return difference as well as the corresponding risk-adjusted return differences. 1F refers to the CAPM, 3F to the Fama and French (1993) model, 4F to the Carhart (1997) model and 6F to a model augmented with factors for short-term and long-term reversal. T-statistics (in parentheses) are adjusted for serial autocorrelation using West and Newey (1987) standard errors with a lag of five months.* indicates significance at the 10% level, ** indicates significance at the 5% level and *** indicates significance at the 1% level.

Specification

Raw

1F

3F

4F

6F

difference

alpha

alpha

alpha

alpha

Panel A: Baseline model from 1995 to 2010 as benchmark Residual media coverage

0.41***

0.38**

0.43***

0.40***

0.40***

(3.01)

(2.57)

(3.07)

(2.97)

(2.78)

Panel B: Inclusion of earnings announcements and 8-K filings as controls in baseline model Earnings announcements

8-K filings

Both earnings announcements and 8-K filings

0.43***

0.40***

0.44***

0.43***

0.43***

(3.19)

(2.73)

(3.13)

(3.11)

(2.95)

0.40***

0.37**

0.42***

0.42***

0.42***

(3.02)

(2.57)

(3.13)

(3.22)

(3.14)

0.41***

0.38**

0.42***

0.42***

0.43***

(3.13)

(2.65)

(3.11)

(3.25)

(3.14)

Panel C: Exclusion of articles around (t-3,t+3) earnings announcements and 8-K filings Earnings announcements

8-K filings

Both earnings announcements and 8-K filings

0.43***

0.39***

0.43***

0.42***

0.42***

(3.26)

(2.72)

(3.14)

(3.09)

(2.86)

0.41***

0.36**

0.39***

0.41***

0.40***

(3.07)

(2.55)

(2.77)

(2.97)

(2.73)

0.38***

0.33**

0.36***

0.39***

0.38***

(3.00)

(2.41)

(2.67)

(2.89)

(2.64)

40

Table 7: Medium-run momentum and long-term reversal

This table presents momentum returns over different holding periods for stock portfolios sorted on residual media coverage. See table 3 for a detailed description of how portfolios are constructed. Momentum returns (reported in % per month) are based on overlapping portfolios which are equally weighted as in the original work of Jegadeesh and Titman (1993). The sample period covers M1:1989-M12:2010. 1F refers to the CAPM, 3F to the Fama and French (1993) model, 4F to the Carhart (1997) model and 6F to a model augmented with factors for short-term and long-term reversal. T-statistics (in parentheses) are adjusted for serial autocorrelation using West and Newey (1987) standard errors with a lag of five months. * indicates significance at the 10% level, ** indicates significance at the 5% level and *** indicates significance at the 1% level. Residual media Loser Winner Momentum Momentum coverage portfolios return return return t-stat Panel A: Momentum effect from t+1 to t+12 1 0.86 1.03 0.17 (0.78) 2 0.79 1.03 0.23 (1.15) 3 0.84 1.08 0.24 (1.07) 4 0.78 1.20 0.42* (1.79) 5 0.82 1.20 0.38 (1.57) Return 5-1 -0.04 0.17 0.21** t-stat 5-1 (-0.39) (1.58) (2.10) Factor model 1F 3F 4F 6F Intercept 5-1 0.20* 0.23** 0.19* 0.19* Intercept t-stat (1.86) (2.40) (1.96) (1.94) Panel B: Momentum effect from t+13 to t+36 1 1.23 1.01 -0.22 (-1.45) 2 1.31 1.12 -0.19* (-1.70) 3 1.47 1.17 -0.31** (-2.15) 4 1.49 1.04 -0.44*** (-2.81) 5 1.35 1.01 -0.35** (-2.27) Return 5-1 0.12 -0.01 -0.13** t-stat 5-1 (1.34) (-0.07) (-2.23) Factor model 1F 3F 4F 6F Intercept 5-1 -0.11** -0.10* -0.10** -0.10** Intercept t-stat (-1.97) (-1.90) (-2.21) (-2.03) Panel C: Momentum effect from t+1 to t+36 1 1.12 1.03 -0.09 (-0.80) 2 1.17 1.10 -0.06 (-0.71) 3 1.29 1.15 -0.14 (-1.40) 4 1.29 1.12 -0.17* (-1.88) 5 1.20 1.08 -0.11 (-0.95) Return 5-1 0.08 0.05 -0.02 t-stat 5-1 (0.91) (0.74) (-0.65) Factor model 1F 3F 4F 6F Intercept 5-1 -0.01 0.01 0.00 -0.01 Intercept t-stat (-0.30) (0.40) (-0.11) (-0.29)

41

Table 8: Individualism and media-based momentum: Portfolio-level analysis

This table compares media-based momentum in collectivistic and individualistic areas. See the text for a more detailed description of this analysis. Both collectivistic and individualistic areas are constructed in a way that they each contain about 30% of sample firms. The table displays portfolio momentum returns stemming from a double sort of residual media coverage (three portfolios in ascending order) and formation period returns. Residual media coverage is based on the number of articles published during the skipped month in both national and local newspapers. Momentum returns (reported in % per month) are based on overlapping portfolios which are equally weighted as in the original work of Jegadeesh and Titman (1993). The sample period covers M1:1989-M12:2010. Panels A to C differ in the way residual media coverage is estimated and in the way the sorting procedure is carried out. In panel A, the model of residual coverage is the baseline model (see table 3) augmented with 35 region dummies. We consider the whole sample stock universe to run this regression and, subsequently, to sort firms into residual media coverage portfolios. In panel B, we rely on the baseline model (without region dummies), which we again run for all sample firms. The subsequent sorting procedure, however, is done for each region separately. Residual media coverage portfolios are then constructed from the pooled region portfolios. In panel C, we again rely on the baseline model (without region dummies). However, we now run the regression and sorting procedure separately for all firms in the collectivistic area and all firms in the individualistic area. T-statistics (in parentheses) are adjusted for serial autocorrelation using West and Newey (1987) standard errors with a lag of five months. * indicates significance at the 10% level, ** indicates significance at the 5% level and *** indicates significance at the 1% level.

Residual media coverage portfolios

1 2 3 Return 3-1 t-stat 3-1 Factor model Intercept Intercept t-stat

1 2 3 Return 3-1 t-stat 3-1 Factor model Intercept Intercept t-stat

1 2 3 Return 3-1 t-stat 3-1 Factor model Intercept Intercept t-stat

Collectivistic areas Individualistic areas Momentum Momentum Momentum Momentum return t-stat return t-stat Panel A: Baseline model with region dummies Raw returns 0.64** (1.97) 0.29 (1.34) 0.70** (2.48) 0.71*** (3.02) 0.90*** (3.03) 0.84*** (3.15) 0.26 0.55*** (1.63) (3.78) Risk-adjusted differences in media-based momentum 1F 3F 4F 6F 0.34* 0.34* 0.27 0.27 (1.81) (1.81) (1.52) (1.50) Panel B: Local ordering Raw returns 0.68** (2.11) 0.25 (1.17) 0.65** (2.29) 0.73*** (3.10) 0.94*** (3.17) 0.90*** (3.31) 0.26* 0.66*** (1.66) (4.49) Risk-adjusted differences in media-based momentum 1F 3F 4F 6F 0.42** 0.43** 0.32 0.32 (2.16) (2.21) (1.62) (1.58) Panel C: Separate estimation Raw returns 0.71** (2.16) 0.22 (0.99) 0.64** (2.25) 0.80*** (3.42) 0.90*** (3.16) 0.81*** (3.05) 0.19 0.59*** (1.25) (4.05) Risk-adjusted differences in media-based momentum 1F 3F 4F 6F 0.43** 0.44** 0.35* 0.34* (2.23) (2.25) (1.75) (1.74)

42

Difference in media-based momentum

0.29 (1.42)

0.40* (1.89)

0.40* (1.91)

Table 9: Individualism and media-based momentum: Fama/MacBeth regressions

This table shows monthly Fama/MacBeth regressions of area-specific media-based momentum returns on the area-specific collectivism score and a number of control variables. The sample period covers M1:1989-M12:2010. In each month, there are 15 areas whose composition is described in detail in figure 3. Control variables include a number of firm characteristics, each averaged across all firms in the respective area. As there are only 15 observations per regression, models 1 to 8 consider only one control variable each. In model 9, we include all control variables which have been proven meaningful in the univariate regressions. T-statistics (in parentheses) are adjusted for serial autocorrelation using West and Newey (1987) standard errors with a lag of five months. * indicates significance at the 10% level, ** indicates significance at the 5% level and *** indicates significance at the 1% level. Regression specification Variable Collectivism score

Size

43 Mom strength

I

II

III

IV

V

VI

VII

VIII

IX

-0.0110

-0.0115*

-0.0140*

-0.0159**

-0.0131*

-0.0095

-0.0126*

-0.0120*

-0.0151**

(-1.64)

(-1.76)

(-1.89)

(-2.26)

(-1.84)

(-1.37)

(-1.77)

(-1.74)

(-2.01)

-0.5751

-0.51725

(-1.60)

(-0.86) 0.1462 (0.14)

Idio Vola

Turnover

Book-market

Analyst coverage

1.5503***

-0.0888

(2.73)

(-0.05) 0.8724**

0.9813

(2.38)

(1.15) -1.7920*

-1.1169

(-1.69)

(-0.80) 0.0064 (0.01)

Residual media coverage

-0.7038

-0.7592

(-1.13)

(-1.14)

Table 10: Fund manager overconfidence and media-based momentum

Panel A displays descriptive statistics for the fund manager overconfidence index, whose design closely follows the approach proposed in Nikolic (2011). More precisely, the index is comprised of the following twelve characteristics: Turnover, residual turnover (with respect to year and fund size), Herfindahl index of the funds portfolio positions, active share, prior 36-months four factor alpha, fund idiosyncratic risk, average market capitalization/book-to-market/idiosyncratic volatility decile rank of the funds stock holdings, manager gender, manager tenure, and a team-managed dummy. Detailed information about the construction of the fund sample is given in the appendix. In panel B, we report findings from regressing media betas on the overconfidence index. To this end, we first follow Fang et al. (2011) in calculating the dollar value of a funds buys for each stock in a given quarter (from t-1 to t). For each fund-quarter with at least eight available observations we then regress ln(1+buy) on a measure of current excess media coverage. More precisely, based on the monthly average of residual media coverage measured over the particular quarter, each stock is assigned a portfolio rank of one to five. Portfolio five here represents the 20% stocks with the highest residual press coverage in the four nation-wide newspapers. This procedure results in fund-quarter-level media betas. In the second step, we pool these values in each quarter and regress the cross-section of fund media betas on the fund manager overconfidence index. Inferences are drawn using the time-series of the coefficients. In panel C, we first distinguish whether the manager had a positive or negative signal about a given stock before running the analysis as described above. More specifically, for a given stock, fund, and quarter, a positive (negative) signal is defined as a non-negative (negative) dollar value of the funds net trading activity in that stock in the previous quarter. In both panels, reported coefficients are standardized and t-statistics (in parentheses) are adjusted for serial autocorrelation using West and Newey (1987) standard errors with seven lags. In panel D, we present main findings from Fama/MacBeth regressions analogously to the ones displayed in table 5. More specifically, we run regressions of average monthly momentum profits on a firm’s media coverage, the value-weighted overconfidence index for managers holding that stock, the interaction effect and on the full set of controls as described in detail in table 5. In all panels, the sample period covers M1:1989-M12:2010. * indicates significance at the 10% level, ** indicates significance at the 5% level and *** indicates significance at the 1% level. Panel A: Descriptive statistics of overconfidence index N Mean Median Std 60,943 7.38 7.38 1.33 Min P10 P90 Max 2.83 5.60 9.13 11.50 Panel B: Media betas and manager overconfidence Sample Coefficient t-statistic All stocks 0.030*** (4.10) Panel C: Media betas and manager overconfidence, depending on signal in previous period Previous Period Coefficient t-statistic Positive signal (buy>= 0) 0.034*** (5.20) Negative signal (sell> 0) -0.056*** (-5.07) Panel D: Manager overconfidence and media-based momentum Full set of controls as in table 5 Momentum return in t+1 Residual media coverage 0.04*** -0.132 rank 1-5 (3.52) (-1.30) Overconfidence index 0.03 -0.032 (0.81) (-0.60) Overconfidence index* 0.024* Residual media coverage (1.68) Average momentum return from t+1 to t+6 Residual media coverage 0.03*** -0.124* rank 1-5 (4.34) (-1.82) Overconfidence index -0.00 -0.063* (-0.12) (-1.87) Overconfidence index* 0.021** Residual media coverage (2.19)

44

Appendix for “Media Makes Momentum”

February 2013

Abstract This appendix contains supplemental material for the study “Media makes momentum”. It starts with an overview of the variables used in the paper. It then provides a discussion of our findings in contrast to those of Chan (2003). In this context, tables 1 to 3 compare our findings with those of Chan. Figure 1 shows the return distribution of the media-based momentum effect uncovered in the study. Table 4 shows summary statistics of national media coverage. Table 5 presents results of the baseline month-to-month OLS regression, carried out to obtain residual media coverage as well as various extensions to this model. Table 6 shows descriptive statistics for the five residual media coverage portfolios. Table 7 displays state rankings and other descriptive statistics of the U.S. collectivism score developed in Vandello and Cohen (1999). Table 8 shows similar descriptive statistics for US regions constructed partly from more than one state to assure a minimum number of firms in each region. Table 9 explores the role of investor collectivism on media-based momentum by running multivariate Fama/MacBeth regressions. Table 10 estimates how fund manager overconfidence is related to the long-term return reversal effect in media-based momentum returns.

1

Appendix: Variable Definitions In the following we outline how main variables used in this study are constructed. When regressing media coverage on a set of firm variables in order to obtain residual media coverage, S&P 500 membership and NASDAQ membership are measured in the skipped month, i.e. the month between formation and evaluation period for the momentum portfolios. To compute firm size and analyst coverage, we consider the average value over the formation period. Unless noted otherwise, data used for variable constructions are taken from CRSP and printed in italics. Amihud Illiquidity: We first compute a daily illiquidity measure defined as the absolute value of the daily stock return divided by scaled total daily dollar volume (prc ∗ vol/106 ). In our analysis, we then rely on the natural log of the equally weighted average of the daily illiquidity ratios over the last month of the formation period. Analyst Coverage: Following Hong et al. (2000), analyst coverage is the natural log of (1+number of estimates for the firm’s earnings next year). The number of estimates is from IBES (numest). Book-to-Market: Book-to-Market is the book value of equity (ceq) divided by market value of equity from Compustat. The market value of equity is measured as common shares outstanding (csho) times price at fiscal year end (prccf ). Book-to-Market is computed as the average value over the formation period. Firm Size: Firm size is defined as the natural log of end-of-month market capitalization, which is calculated as the number of shares outstanding (shrout) times price (prc). Firm Age: Age is computed as the natural log of the number of months since the firm’s PERMCO first appeared in CRSP (starting in 1925). Firm age is measured in the skipped month. Idiosyncratic Volatility: Idiosyncratic volatility is the residual from a Fama and French (1993) three factor regression (with mktrf from CRSP as excess market return) using daily return data from the previous six months lagged by one month (i.e. the formation period for the momentum portfolios). Momentum Strength As in Bandarchuk and Hilscher (2013), momentum strength is defined as exp(absolute difference between a stock’s formation period log return and the median formation period log return of all stocks in the sample)-1.

2

Retpr6: Retpr6 is the return of the stock over the previous six months lagged by one month (i.e. the formation period for the momentum portfolios). NASDAQ: NASDAQ is a dummy variable taking the value of 1 if the firm is listed on NASDAQ (exchcd==3). S&P 500: S&P 500 is a dummy variable taking the value of 1 if the firm’s stock is a constituent of the S&P 500 index. Data is obtained from the “Index Constituents” database available in Compustat. Price: Price is computed as the natural log of the equally weighted average of the daily closing price (prc) over the last month of the formation period. Turnover: Turnover is share volume (vol) divided by shares outstanding (shrout). We use the following modifications for NASDAQ stocks (see e.g. Anderson and Dyl (2005)): We multiply turnover by 0.5 before 1.1.1997 and by 0.62 afterwards. Based on daily data, we then calculate average turnover (for all stocks) during the previous six months lagged by one month and take the natural log.

3

Appendix: Mutual Fund Data The mutual fund holdings data and general information comes from two sources: The CRSP Survivorship Bias Free Mutual Fund Database (CRSP MFBD, henceforth) and the Thomson Reuters Mutual Fund Holdings database. Both databases are merged using MFLinks from Wharton Research Data Services (WRDS). In the first step, we largely follow the filter rules of Kacperczyk et al. (2008) to restrict the sample to domestically investing equity funds in CRSP MFBD. Specifically, we use the classification systems from Lipper, Strategic Insight, Wiesenberger and the variable ”policy”. We declare all funds with Lipper classification codes LCCE, LCGE, LCVE, MLCE, MLGE, MLVE, SCCE, SCGE, SCVE, MCCE, MCGE, or MCVE as equity. Further, funds are also classified as equity if they have AGG, GMC, GRI, GRO, ING, or SCG as Strategic Insight classification code, or GRO, LTG, MCG, or SCG as Wiesenberger Objective code. Finally, if the policy variable equals CS (common shares), we also classify the fund as equity product. After this first classification, we delete all funds that are by none of the above criteria classified as equity. To address the problem of incubation bias (see Evans (2010)), we then exclude all observations where the month for the observation is prior to the reported fund starting month in CRSP MFBD. We also drop observations in CRSP MFDB where the fund had less than $5 million under management in the previous month since incubated funds tend to be smaller. The resulting sample is merged with MFLinks to obtain the Wharton Financial Institution Center Number (wficn) for each share class in CRSP MFBD and investment objective codes (ioc) from the Thomson Reuters database. Fund share classes with the same wficn are aggregated into one fund and funds without a record in MFLinks are dropped from the sample. Further, we exclude funds that do not primarily invest in common shares according to Thomson Reuters (ioc 1, 5, 6, 7, or 8). Even after applying these filter criteria, the resulting sample still maintains a substantial number of nonequity and international funds. To address this problem, we first use the percentage invested in common stocks (per com) from the CRSP MFBD annual summary files and exclude funds that on average hold less than 80% or more than 105% in common stocks. This rule is also taken from Kacperczyk et al. (2008). Second, we merge the Thomson Reuters holdings file with monthly stock data from CRSP and calculate the percentage of the funds’ total assets invested in CRSP stocks. If this percentage is on average also below 80% we delete the funds from the sample. The last criterion primarily excludes international funds. Finally, to enhance the robustness of the MFLinks merge, we compare total fund assets (TNA) in CRSP 4

MFBD to total fund assets in Thomson Reuters:

Absolute dif f erence = |T N ACRSP − T N AT homson |/(T N ACRSP ).

(1)

If the median absolute difference for a particular fund over all overlapping data observations is larger than 1.3 or smaller than 1/1.3, we also drop the fund from the sample. After applying these steps the sample consists of 3,062 different funds and 100,763 fund quarters over the period from 1989 to 2010. Next, we identify and also exclude index funds and ETF since our focus is on actively managed funds. To this end, we use fund names and search for any of the following strings: ”Index”, ”Idx”, ”Ix”, ”Indx”, ”Nasdaq”, ”Dow”, ”Mkt”, ”DJ”, ”S&P 500”, ”S&P, 500”, ”Barra”, ”DFA”, ”Vanguard”, ”ETF”, ”SPDR”, ”ETN”, ”Powershares”, ”Wisdomtree”, ”Tracker”, and ”Profunds”. The coding routine is based on the work of Gil-Bazo and Ruiz-Verdu (2009) but uses some additional character strings to better reflect the increased importance of ETF and the appearance of new index fund and ETF providers towards the end of our sample period. This selection criterion reduces our final mutual fund sample to 2,719 funds (89,248 fund quarters).

Appendix: Comparison with the findings of Chan (2003) In his analysis, Chan (2003) relies both on newspapers (Wall Street Journal, Chicago Tribune, Los Angeles Times, New York Times, Washington Post, USA Today) and newswires (Associated Press Newswire, Gannett New Service, Dow Jones Newswire). While his newspaper database has a considerable overlap with ours, the additional use of newswires may lead to his news sample being considerably different from ours. Descriptive statistics show that this is indeed the case. In calculating these statistics, we aim at following his approach to the extent possible. Note that this is not perfectly possible, due to e.g. a different stock universe. For instance, Chan (2003) focusses on a random subsample of about a quarter of all CRSP stocks during his sample period (1980-2000). In his sample, about one-half of stocks has some news in each month. In the later years of the sample period, this number increases to even close to two thirds. In contrast, relying on our national newspapers from 1980 to 2000, the coverage is only about 20%. Even if we focus on a more recent time period and also take local newspapers into account, the fraction of firms with news in a given month is well below 30%. It thus seems that most news events are driven by the use of newswires. Other statistics confirm this impression. For instance, 14% of firms have news in 90% or more months of their lifespan, according 5

to Chan (2003). The respectively value in our analysis is 1%. Moreover, in the news data set of Chan (2003), only 8% of firms have news in 10% or less months. In contrast, in our newspaper data set, the respective value is more than 50%. Similar differences apply to other statistics, such as the probability of news in month t+1, conditioned on news in t. Despite these differences, we first aim at replicating the findings of Chan (2003) by using his methodology and our sample of newspapers. Note that this implies, for instance, the use of cumulative raw returns. The following table shows the main findings. Despite the considerable different data base, our results are broadly in line with his. There is a significant positive return difference, both in the medium run (momentum) and in the long run (potential reversal effects), between the winner minus loser spread in news stocks and the winner minus loser spread in no news stocks. Table 1: Momentum and reversal in news and no news stocks (replication of table 4 in Chan (2003), similar methodology and similar stock data) 1980-2000 News Stocks Holding Period

No News Stocks

Winner

Loser

Momentum

Winner

Loser

Momentum

1

1.05

1.18

-0.13

0.67

1.65

-0.99

3

3.88

3.06

0.82

3.17

3.74

-0.56

6

7.75

5.94

1.81

6.98

6.82

0.16

12

16.41

11.80

4.61

15.28

12.95

2.33

24

31.79

27.25

4.54

30.00

29.38

0.62

1980-2000 Holding Period

Difference between news and no news stocks Winner

T-Stat

Loser

T-Stat

Momentum

T-Stat

1

0.38

3.11

-0.47

-3.54

0.85

7.57

3

0.71

2.23

-0.67

-1.73

1.38

6.49

6

0.77

1.25

-0.88

-1.18

1.65

5.30

12

1.13

0.91

-1.15

-0.79

2.28

4.77

24

1.79

0.69

-2.13

-0.72

3.91

5.10

We next redo this analysis, but now focus on the subsample of stocks that we rely on in our paper. That is, we exclude stocks in the first NYSE size decile and stocks with prices below $5 at the end of the formation period to make sure that our findings are not driven by small and illiquid stocks. As stocks traded at the NYSE are typically large, our procedure implies that about 50% of CRSP firm months are dropped. Thus, many of the small stocks which Chan considers and in which the reported return drift is particularly pronounced will not enter our analysis.1 As can be seen from table 2, this change considerable reduces the long run differences between the news and the no news sample. For instance, 1 Note

that, despite this strict screening process, our final sample contains more firms (about 7,800) than Chan’s randomly

selected CRSP subsample (about 4,200).

6

after 24 months, the difference between both samples is now only 1.41% as opposed to 3.91% if we do not exclude small firms. In other words, the findings become more in line with the results reported in our study. It is also worth noting that there is no loser drift anymore (fourth column, lower half of the table) as it was the case in table 1. The effect now tends to come more from the winner side, as it the case in our study. Thus, the use of the firm universe is important. Table 2: Momentum and reversal in news and no news stocks (replication of table 4 in Chan (2003), similar methodology, but exclusion of small stocks) 1980-2000 News Stocks Holding Period

No News Stocks

Winner

Loser

Momentum

Winner

Loser

Momentum

1

1.24

1.26

-0.01

1.01

1.48

-0.47

3

4.10

3.54

0.56

3.66

3.57

0.09

6

8.14

6.80

1.34

7.53

6.59

0.94

12

16.77

12.93

3.85

15.40

11.84

3.57

24

32.14

27.99

4.15

28.98

26.25

2.73

1980-2000 Holding Period

Difference between news and no news stocks Winner

T-Stat

Loser

T-Stat

Momentum

T-Stat

1

0.24

2.30

-0.22

-2.30

0.46

4.90

3

0.44

1.86

-0.03

-0.11

0.47

3.07

6

0.62

1.34

0.21

0.45

0.41

1.95

12

1.37

1.56

1.09

1.26

0.28

0.82

24

3.16

1.77

1.75

1.06

1.41

2.73

Note that there also a number of methodological differences in our paper and the study by Chan (2003). For instance, the ways media coverage are measured can hardly be compared. Chan relies on a binary analysis quantifying whether a stock was mentioned in the headline or lead paragraph in a given month. In contrast, we consider the exact number of all firm-specific articles as indicated by LexisNexis. Note that this does not necessarily imply that the firm is mentioned in the headline or the first paragraph. Moreover, we rely on a measure of excess coverage which measures the unexpectedly high or low weight the media attaches to a given firm, holding important characteristics as firm size or analyst coverage fixed. In addition, while Chan considers news stories over the previous month, we (partly) rely on the previous six months. The same holds true for the length of the formation period. Chan relies on one month, we on six months, which is the period predominantly used in the momentum literature. Note also that the return computation scheme and the sample periods differ. In addition, the descriptive statistics shown above give rise to the question to what extent the large impact of newswires might lead to a qualitatively different sample of news events. For instance, 90% of firms that make an earnings announcement in a specific month are included in the news sample of Chan (2003), but 7

only 43% are included in our solely newspaper-based replication. This lends support to the idea that the newswire sample is more likely to catch actual news events, i.e. the arrival of new information, whereas our newspaper articles might, in the first place, proxy for attention effects not necessarily related to time-sensitive valuable news. The following table provides strong support for this notion. We follow the methodology of Chan with our newspaper sample, but now include the additional constraint that only newspaper articles in firm months without an earnings announcement and without an 8-K filing are taken into account. This procedure is intended to isolate the true impact of newspaper articles, controlling for firm-specific news, which are likely to be reflected in newswire articles. We exemplarily select a sample period from 1995 to 2010, as 8-K filings only become available from 1995 on. Table 3: Momentum and reversal in news and no news stocks based on newspaper articles in months without earnings announcements and 8-K-filings 1995-2010 News Stocks Holding Period

No News Stocks

Winner

Loser

Momentum

Winner

Loser

Momentum

1

0.71

1.14

-0.42

0.59

1.84

-1.25

3

2.59

2.89

-0.30

2.84

4.26

-1.42

6

5.85

5.24

0.61

6.21

7.74

-1.53

12

11.57

11.20

0.37

13.57

14.64

-1.07

24

20.99

24.53

-3.54

26.65

30.44

-3.79

Winner

T-Stat

Loser

T-Stat

Momentum

T-Stat

1

0.12

0.66

-0.70

-2.62

0.83

3.06

3

-0.25

-0.54

-1.36

-2.26

1.11

2.33

6

-0.35

-0.40

-2.50

-2.06

2.15

2.53

12

-2.01

-1.16

-3.45

-1.50

1.44

1.00

24

-5.66

-1.54

-5.91

-1.32

0.25

0.13

1995-2010 Holding Period

Difference between news and no news stocks

The table shows an intriguing result. Now, and in contrast to Chan (2003), firms with news show a pronounced long-term reversal. Due to this fact, the return difference between news stocks and no news stocks is insignificant in the long run, which is qualitatively similar to our baseline findings in the paper. In contrast and as verified in untabulated tests, if we only focus on the months during which there were earnings announcements or 8-K filings, then there are long-tern return differences. Together, these findings suggest that the behavior of returns in the long run depends on the type of news studied. If one focusses primarily on articles which are likely to contain value-relevant news (as the newswires in Chan (2003), earnings announcements, or 8-K filings), then there does not seem to be a

8

long-term reversal effect. This is in line with the idea of slow information diffusion. In contrast, if one focusses primarily on pure media coverage not necessarily related to news about fundamentals (as in our baseline analysis or in the exercise above), then momentum returns tend to reverse in the long run. This is in line with the idea of investor attention leading to overreaction. Note that this might also partly explain why Chan (2003) finds stronger findings in the short leg of the portfolio, whereas our findings are stronger in the long leg (see also Barber and Odean (2008)). Another reason for this discrepancy might be the impact of small and illiquid firms, as suggested by the comparison of tables 1 and 2. Together, the crucial differences in sample firms, news articles, and in the methodology are likely to explain while the two studies arrive at partly different conclusions.

9

Figure 1: Return distribution of the media-based momentum effect This figure shows the historical return distribution of the media-based momentum effect uncovered in the paper. Media-based momentum is defined as the difference between the momentum effect in the portfolio of stocks with the highest residual media coverage and the momentum effect in the portfolio of stocks with the lowest residual coverage. To construct the momentum portfolios we use a formation and holding period of six months each and skip one month in between. Momentum profits are derived from a dependent double sorting procedure where stocks are first sorted into five equal-sized residual media coverage portfolios. Residual media coverage is computed from rolling cross-sectional OLS regressions of ln (1+number of articles in the skipped month) on firm size, analyst coverage and dummies for S&P 500 and NASDAQ membership. Within each quintile of residual media coverage, the winner (loser) portfolio then consists of all stocks having a return during the formation period above the 70th percentile (below the 30th percentile). Momentum returns are based on overlapping portfolios which are equally-weighted as in the original work of Jegadeesh and Titman

0

10

Frequency 20

30

(1993). The sample period covers M1:1989-M12:2010.

−.1

−.05

0 media_based_momentum

10

.05

.1

Table 4: Summary statistics on media coverage

This table presents summary statistics for the national media coverage of our sample firms. National media coverage is based on articles written about a firm in one of the following four U.S. newspapers: New York Times, USA Today, Wall Street Journal, and Washington Post. Articles are obtained from LexisNexis using the company search function and a “relevance score” of at least 80%. Panel A shows the total number of firms (N) and the percentage of firms having at least one article (%) in a given year for a subset of years and the whole sample (”All years”). The statistics are also shown for sub-samples divided in NYSE and NASDAQ stocks. Conditioned on firms having at least one article in a year, Panel B reports distribution details for the yearly number of articles. The sample period covers M1:1989-M12:2010. Panel A: Unconditional media coverage statistics All stocks

NYSE stocks

NASDAQ stocks

Year

N

%

N

%

N

%

1990

2,671

50

1,113

70

1,310

35

1995

4,235

51

1,547

73

2,466

40

2000

4,889

42

1,612

62

3,068

33

2005

2,935

47

1,368

62

1,482

34

2010

2,553

30

1,203

43

1,284

20

78,986

45

31,234

64

43,787

33

All years

Panel B: Distribution conditional on media coverage All years

Mean

Median

p99

p75

p25

p1

15.85

3

271

9

1

1

11

Table 5: Multivariate regressions to explain media coverage

This table presents the results of the baseline month-to-month OLS regression carried out to obtain residual media coverage (specification I) as well as various extensions. The variable Mom Strength refers to Bandarchuk and Hilscher (2013). Further construction details are provided in the variable definitions. The sample period covers M1:1989-M12:2010. Following Fama/MacBeth (1973), coefficients are calculated as time-series averages of monthly estimates, and t-statistics (in parentheses) are based on the time-series average and standard deviation. We adjust t-statistics for serial autocorrelation using West and Newey (1987) standard errors with a lag of five months.* indicates significance at the 10% level, ** indicates significance at the 5% level and *** indicates significance at the 1% level.

12

Basis-Model I

I + Book Market II

II + Turnover III

III + Idio Vola IV

IV + Mom Strength V

V + Industry Dummies VI

Firm size

0.1603*** (36.30)

0.1700*** (42.16)

0.1768*** (42.36)

0.2056*** (41.13)

0.2072*** (42.21)

0.2080*** (40.46)

NASDAQ Dummy

0.0167*** (2.84)

0.0232*** (4.07)

0.0289*** (4.44)

-0.0177** (-2.53)

-0.0161** (-2.30)

-0.0227*** (-3.69)

S&P 500 Dummy

0.2601*** (23.16)

0.2513*** (23.26)

0.2496*** (23.42)

0.2302*** (21.41)

0.2266*** (20.95)

0.2170*** (21.65)

Analyst coverage

-0.0274*** (-19.23)

-0.0352*** (-23.01)

-0.0581*** (-23.38)

-0.0482*** (-21.68)

-0.0477*** (-20.94)

-0.0475*** (-25.02) )

0.0178*** (4.59)

0.0233*** (6.48)

0.0371*** (9.91)

0.0419*** (10.22)

0.0565*** (10.64)

0.0355*** (11.67)

-0.0061 (-1.45)

-0.0074* (-1.76)

-0.0097*** (-2.60)

0.1859*** (21.28)

0.1744*** (21.33)

0.1645*** (17.63)

0.0372*** (9.31)

0.0387*** (9.80)

Book-to-market

Turnover

IVOL

Momentum strength

Constant

No. of observations adj. R Fama-French 48 industry dummies

-1.9472*** (-27.10)

-2.0748*** (-30.78)

-2.1064*** (-30.08)

-1.7966*** (-30.81)

-1.8739*** (-31.07)

-1.8973*** ( -29.84)

264 0.2581 no

264 0.2621 no

264 0.2652 no

264 0.2779 no

264 0.2791 no

264 0.2983 yes

Table 6: Descriptive statistics of residual media coverage portfolios

This table presents descriptive firm-level statistics (average monthly number of press articles in the four national newspapers, firm size for each of the five residual media coverage portfolios. Panel A to C differ only in the way residual media coverage portfolios are constructed. In panel A, we rely on the baseline model as described in detail in the paper. This implies that the sorting procedure for residual media coverage portfolios is based on residual coverage for articles in the skipped month. In panel B, we rely on specification 2 of table 4 in the paper. This implies that the sorting procedure is based on residual media coverage over the formation period and the skipped month, i.e. effectively over the past seven months. Panel C is identical to panel B, except that we first exclude all firms with zero coverage in the previous seven months before the computation of residual media coverage takes place. The portfolio containing the firms with zero coverage is denoted as portfolio 0. Period: M1:1989-M12:2010

Average monthly No.

Average monthly No.

Firm Size

Firm Size

NYSE Size Decile

of articles

of articles

(in million $)

(in million $)

Decile

NYSE Size Decile Decile

(skipped month)

(previous 7 months)

Mean

Median

Mean

Median

Panel A: Specification as in the baseline model Residual Media Coverage Portfolio

13

1

0.01

0.31

5,629

2,944

7.70

7.60

2

0.04

0.16

1,480

915

5.02

4.87

3

0.04

0.13

886

441

3.41

3.16

4

0.05

0.11

701

237

2.45

2.02

5

3.81

3.22

9,707

1,098

5.73

5.52

1

0.15

0.14

5,353

2,821

7.54

7.57

2

0.13

0.13

1,683

855

5.01

4.61

3

0.13

0.13

1,069

431

3.64

3.12

4

0.17

0.17

906

254

2.98

2.05

5

3.38

3.36

9,437

611

5.20

4.18

0

0.00

0.00

1,220

466

4.00

3.24

1

0.04

0.65

8,325

4,737

8.16

8.12

2

0.14

0.50

3,431

1,237

5.74

5.48

3

0.26

0.52

3,293

496

4.18

3.32

4

1.18

1.07

6,193

1,392

5.68

5.78

5

7.38

6.20

15,600

2,303

6.36

6.69

Panel B: Specification as in table 4, robustness test 2 Residual Media Coverage Portfolio

Panel C: Specification as in table 4, robustness test 3 Residual Media Coverage Portfolio

Table 7: Individualism and states: Descriptive statistics

This table presents descriptive statistics for individual U.S. states. Collectivsm Score refers to the collectivism index proposed in Vandello and Cohen (1999). N umber of f irms refers to those firms in our final sample whose headquarter is located within a given state. N umber of articles counts news stories in both local and nationwide newspapers, averaged across all firms in a given state in a given month. State

ID

Collectivism

Number of firms

Number of articles

ccore

mean

std

min

max

mean

std

min

max

22.31

6.30

10

35

1.09

0.66

0.00

3.93 4.50

Alabama

1

57

Alaska

2

48

1.84

0.52

0

3

0.12

0.39

0.00

Arizona

3

49

31.32

8.18

16

48

1.11

1.18

0.04

9.03

Arkansas

4

54

15.16

1.85

12

20

11.70

7.46

0.33

32.69

California

5

60

425.63

96.66

280

731

3.04

1.15

1.00

6.15

Colorado

6

36

50.21

11.54

36

86

1.42

0.77

0.33

3.90

Connecticut

7

50

73.09

17.06

45

102

1.79

0.59

0.86

5.20

Delaware

8

55

11.26

3.05

6

17

3.25

1.70

0.86

12.86

Florida

9

54

95.31

25.37

54

151

1.34

0.61

0.39

4.42

Georgia

10

60

78.73

17.64

48

112

3.61

1.33

0.60

9.15

Hawaii

11

91

6.02

1.62

3

10

0.11

0.17

0.00

1.11

Idaho

12

42

8.20

1.83

4

12

1.01

1.03

0.00

6.00

Illinois

13

52

147.73

28.19

103

194

3.64

1.07

1.52

6.53

Indiana

14

57

39.36

8.18

26

55

0.84

0.54

0.05

4.16

Iowa

15

39

20.09

5.23

12

29

0.27

0.25

0.00

1.72

Kansas

16

38

15.48

2.56

11

23

2.69

2.03

0.00

13.57

Kentucky

17

53

21.14

3.97

15

32

1.07

0.71

0.06

3.73

Louisiana

18

72

20.47

2.38

15

26

1.68

0.96

0.00

5.94

Maine

19

45

5.13

1.09

3

8

0.25

0.49

0.00

4.33

Maryland

20

63

56.84

10.33

41

82

2.30

0.80

0.70

5.11

Massachusetts

21

46

137.45

38.02

90

237

1.00

0.39

0.24

3.01

Michigan

22

46

58.86

11.65

34

80

7.92

3.50

1.64

20.00

Minnesota

23

41

75.22

16.09

51

106

2.30

0.98

0.49

4.96

Mississippi

24

64

10.72

3.20

6

17

0.29

0.35

0.00

3.86

Missouri

25

46

55.98

9.09

42

74

2.08

0.71

0.71

7.09

Montana

26

31

3.04

1.23

1

6

0.08

0.20

0.00

1.50

Nebraska

27

35

13.04

2.33

8

19

2.96

1.80

0.22

9.89

Nevada

28

52

20.69

4.73

13

32

1.89

1.59

0.00

12.06

New Hampshire

29

43

11.46

6.08

1

23

0.19

0.25

0.00

2.00

New Jersey

30

59

108.35

19.87

71

150

2.22

0.98

0.73

5.73

New Mexico

31

51

3.53

1.59

1

7

0.37

0.60

0.00

4.83

New York

32

53

224.81

38.10

174

326

6.72

2.34

2.51

15.09

North Carolina

33

56

55.83

10.14

36

74

6.08

2.22

1.66

23.00

North Dakota

34

37

1.86

0.68

1

3

0.08

0.26

0.00

2.00

Ohio

35

45

111.97

22.82

74

152

2.75

1.10

0.42

6.10

Oklahoma

36

42

18.36

3.53

9

24

2.42

1.36

0.00

6.33

Oregon

37

33

25.37

5.95

15

38

2.90

1.36

0.67

8.55

Pennsylvania

38

52

136.98

20.26

96

172

1.30

0.52

0.27

3.75

Rhode Island

39

48

9.03

2.59

5

14

5.64

3.72

0.38

25.20

South Carolina

40

70

17.50

5.92

8

28

0.32

0.26

0.00

1.23

South Dakota

41

36

4.49

0.90

2

6

0.40

0.79

0.00

5.60

Tennessee

42

56

45.20

9.62

22

64

1.23

0.58

0.27

4.41

Texas

43

58

242.73

45.70

164

345

3.85

1.40

1.13

7.62

Utah

44

61

14.58

3.95

8

24

0.87

0.72

0.00

4.45

Vermont

45

42

4.81

1.98

1

9

0.46

1.07

0.00

10.14

Virginia

46

60

71.88

11.99

52

100

3.43

1.15

1.31

8.94

Washington

47

37

48.22

11.06

30

81

5.94

3.00

0.97

17.11

West Virginia

48

48

5.98

1.81

4

13

0.18

0.37

0.00

4.80

Wisconsin

49

46

49.27

9.16

33

66

1.07

0.53

0.10

3.02

Wyoming

50

35

0.30

0.60

0

2

0.12

0.46

0.00

3.00

14

Table 8: Individualism and U.S. regions: Descriptive statistics

This table presents descriptive statistics for 35 U.S. regions. A region is equivalent to a state unless there are less than 15 sample firms on average or less than 9 sample firms in any month located in this state. If this is the case, regions are constructed from at least two neighboring states with a similar collectivism score. Collectivsm Score refers to the (average of the state-specific) collectivism index proposed in Vandello and Cohen (1999). N umber of f irms refers to those firms in our final sample whose headquarter is located within a given region. N umber of articles counts news stories in both local and nationwide newspapers, averaged across all firms in a given region in a given month. Region Alabama Arizona, New Mexiko Arkansas California Colorado Connecticut, Rhode Island Delaware, Pennsylvania Florida Georgia, South Carolina Idaho, Montana, Nebraska, North Dakota, South Dakota, Wyoming Illinois Indiana Iowa Kansas Kentucky Louisiana, Mississippi Maine, New Hampshire, Vermont Maryland Massachusetts Michigan Minnesota Missouri Nevada New Jersey New York North Carolina Ohio, West Virginia Oklahoma Oregon Tennessee Texas Utah Virginia Washington Wisconsin

ID

Collectivism Score

mean

1 2 3 4 5 6 7 8 9 10

57 50 54 60 36 49 53.5 54 65 36

22.31 34.85 15.16 425.63 50.21 82.12 148.24 95.31 96.24 30.92

6.30 8.67 1.85 96.66 11.54 19.48 23.06 25.37 22.34 4.32

11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35

52 57 39 38 53 68 43 63 46 46 41 46 52 59 53 56 46.5 42 33 56 58 61 60 37 46

147.73 39.36 20.09 15.48 21.14 31.19 21.41 56.84 137.45 58.86 75.22 55.98 20.69 108.35 224.81 55.83 117.96 18.36 25.37 45.20 242.73 14.58 71.88 48.22 49.27

28.19 8.18 5.23 2.56 3.97 3.81 8.57 10.33 38.02 11.65 16.09 9.09 4.73 19.87 38.10 10.14 23.60 3.53 5.95 9.62 45.70 3.95 11.99 11.06 9.16

15

Number of firms std min

max

Number of articles mean std min max

10 19 12 280 36 50 104 54 56 22

35 53 20 731 86 116 188 151 140 41

1.09 1.06 11.70 3.04 1.42 2.20 1.44 1.34 3.04 1.58

0.66 1.10 7.46 1.15 0.77 0.72 0.53 0.61 1.15 0.88

0.00 0.04 0.33 1.00 0.33 0.92 0.37 0.39 0.47 0.13

3.93 8.57 32.69 6.15 3.90 5.30 3.64 4.42 7.74 4.88

103 26 12 11 15 24 7 41 90 34 51 42 13 71 174 36 79 9 15 22 164 8 52 30 33

194 55 29 23 32 42 36 82 237 80 106 74 32 150 326 74 158 24 38 64 345 24 100 81 66

3.64 0.84 0.27 2.69 1.07 1.20 0.26 2.30 1.00 7.92 2.30 2.08 1.89 2.22 6.72 6.08 2.63 2.42 2.90 1.23 3.85 0.87 3.43 5.94 1.07

1.07 0.54 0.25 2.03 0.71 0.64 0.29 0.80 0.39 3.50 0.98 0.71 1.59 0.98 2.34 2.22 1.06 1.36 1.36 0.58 1.40 0.72 1.15 3.00 0.53

1.52 0.05 0.00 0.00 0.06 0.03 0.00 0.70 0.24 1.64 0.49 0.71 0.00 0.73 2.51 1.66 0.40 0.00 0.67 0.27 1.13 0.00 1.31 0.97 0.10

6.53 4.16 1.72 13.57 3.73 3.48 2.47 5.11 3.01 20.00 4.96 7.09 12.06 5.73 15.09 23.00 5.84 6.33 8.55 4.41 7.62 4.45 8.94 17.11 3.02

Table 9: Individualism and media-based momentum: Firm-level analysis

This table shows findings obtained from an analysis similar to the one displayed in table 5 in the paper. See the paper’s table description for a detailed explanation how the analysis is carried out. We now augment the regression model with additional explanatory variables to explore the effect of geographic variation in investor individualism. More specifically, we add the state-level collectivism score as developed in Vandello and Cohen (1999) and the interaction effect between the firm-level residual media coverage rank (1 to 3) and the collectivism score. The sample period covers M1:1989-M12:2010. T-statistics (in parentheses) are adjusted for serial autocorrelation using West and Newey (1987) standard errors with a lag of five months.* indicates significance at the 10% level, ** indicates significance at the 5% level and *** indicates significance at the 1% level.

Variable Residual media coverage rank 1-3

I 0.10*** (4.63)

Momentum return in t+1 II III IV 0.28** 0.28** (2.07) (2.16)

V 0.26** (2.00)

Average momentum return from t+1 to t+6 I II III IV V 0.06*** 0.2151*** 0.1850** 0.1736** (4.06) (2.74) (2.37) (2.24)

Mom strength

0.41** (2.20)

0.41** (2.20)

0.35** (2.04)

0.36** (2.09)

0.33** (2.13)

0.33** (2.13)

0.29** (2.01)

0.29** (2.03)

IVOL (rank)

0.03*** (5.51)

0.03*** (5.41)

0.03** (2.58)

0.02** (2.38)

0.01*** (3.64)

0.01*** (3.47)

0.00 (0.61)

0.00 (0.88)

0.00 (1.33)

0.01 (1.30)

0.00 (0.89)

0.01** (2.20)

0.00* (1.66)

0.00 (1.19)

-0.0041* (-1.67)

-0.0040* (-1.68)

-0.0035 (-1.49)

-0.0033** (-2.28)

-0.0026* (-1.82)

-0.0025* (-1.75)

Size

-0.13 (-1.34)

-0.15 (-1.60)

0.10 (1.22)

0.08 (1.08)

Book-Market

0.02 (0.45)

0.05 (0.83)

-0.04 (-1.15)

-0.03 (-0.77)

Analyst Coverage

-0.05 (-1.14)

-0.03 (-0.52)

0.00 (0.03)

0.02 (0.73)

Turnover

-0.05 (-0.52)

-0.07 (-0.61)

0.12* (1.80)

0.11* (1.80)

-0.13*** (-4.18)

-0.11*** (-4.24)

-0.09*** (-4.03)

-0.07*** (-4.05)

NASDAQ dummy

-0.05 (-0.98)

-0.06 (-1.24)

-0.01 -(0.32)

-0.03 (-0.78)

Amihud

-0.07 (-0.90)

-0.07 (-1.02)

0.11* (1.67)

0.10* (1.69)

Price

0.24*** (2.71)

0.00** (2.49)

0.07 (1.13)

0.08 (1.38)

no

yes

no

yes

Collectivism Score

16

Collectivism Score * Residual Coverage Rank

Age

Fama-French 48 industry dummies

no

no

no

no

no

no

Table 10: Fund manager overconfidence and media-based momentum: Long-term reversal

This table presents main findings from Fama/MacBeth regressions analogously to the ones displayed in tables 5 and 9 in the paper. However, in contrast to table 10, we do not look at momentum returns (i.e. the time period from t+1 to t+6), but at long-term reversal effects (i.e. the time period from t+13 to t+36). More specifically, we run regressions of average firm-level returns over t+13 to t+36 on a firm’s media coverage, the value-weighted overconfidence index for managers holding that stock, the interaction effect and on the full set of controls as described in detail in table 5. In all panels, the sample period covers M1:1989-M12:2010. * indicates significance at the 10% level, ** indicates significance at the 5% level and *** indicates significance at the 1% level.

Manager overconfidence and media-based long-term reversal Full set of controls as in tables 5 and 9 in the paper Average momentum return from t+13 to t+36 Residual media coverage

0.06

rank 1-5

(1.61)

Overconfidence index

-0.001 (-0.05)

Overconfidence index*

-0.009*

Residual media coverage

(-1.69)

17

References Anderson, A.-M., and E. A. Dyl, 2005, “Market structure and trading volume,” Journal of Financial Research, 28, 115–131. Bandarchuk, P., and J. Hilscher, 2013, “Sources of momentum profits: Evidence on the irrelevance of characteristics,” Review of Finance, 17, 809–845. Barber, B. M., and T. Odean, 2008, “All that glitters: The effect of attention and news on the buying behavior of individual and institutional investors,” Review of Financial Studies, 21, 785–818. Chan, W. S., 2003, “Stock price reaction to news and no-news: Drift and reversal after headlines,” Journal of Financial Economics, 70, 223–260. Evans, R. B., 2010, “Mutual Fund Incubation,” Journal of Finance, 65, 1581–1611. Fama, E. F., and K. R. French, 1993, “Common risk factors in the returns on stocks and bonds,” Journal of Financial Economics, 33, 3–56. Fama, E. F., and J. D. MacBeth, 1973, “Risk, return, and equilibrium: Empirical tests,” Journal of Political Economy, 81, 607–636. Gil-Bazo, J., and P. Ruiz-Verdu, 2009, “The Relation Between Price and Performance in the Mutual Fund Industry,” Journal of Finance, 64, 2153–2183. Hong, H. G., T. Lim, and J. C. Stein, 2000, “Bad news travels slowly: Size, analyst coverage, and the profitability of momentum strategies,” Journal of Finance, 55, 265–295. Jegadeesh, N., and S. Titman, 1993, “Returns to buying winners and selling losers: Implications for stock market efficiency,” Journal of Finance, 48, 65–91. Kacperczyk, M., C. Sialm, and L. Zheng, 2008, “Unobserved Actions of Mutual Funds,” Review of Financial Studies, 21, 2379–2416. Vandello, J. A., and D. Cohen, 1999, “Patterns of individualism and collectivism across the United States,” Journal of Personality and Social Psychology, 77, 279–292. West, K. D., and W. K. Newey, 1987, “A simple, positive semi-definite, heteroskedasticity and autocorrelation consistent covariance matrix,” Econometrica, 55, 703–708.

18