Leveling Playing Field or Obfuscation: The Informational Role of Overconfident CEOs*

Transcription

1 Leveling Playing Field or Obfuscation: The Informational Role of Overconfident CEOs* Chishen Wei Nanyang Technological University Lei Zhang Nanyang Technological University October 12, 2016 Abstract We study the informational role of overconfident CEOs. We find that stocks led by overconfident CEOs have lower analyst forecast dispersion, higher breadth of ownership, and lower informed trading intensity. These results are consistent around CEO turnovers. We also show that overconfident CEOs level the playing field between short sellers and other investors, as the (good) news in short interest (Boehmer, Huszar and Jordan, 2010) exists only among stocks with non-overconfident CEOs and disappears among stocks with overconfident CEOs. These findings suggest that overconfident CEOs can help increase information efficiency and lower the mispricing of their company shares. Keywords: Overconfident CEOs, Differences of Opinion, Informed Trading, Short Interest, Information Efficiency JEL Classification: D80, G14, G32, G34 * We thank Zhanhui Chen, Stephen Dimmock, Chuan Yang Hwang, Oguzhan Karakas, Jun-koo Kang, Andy Kim, Neng Wang, Xiaoyun Yu and seminar participants at Nanyang Business School for helpful comments. Both authors are at Nanyang Business School, Nanyang Technological University, Singapore Chishen Wei cswei@ntu.edu.sg; phone: (+65) Lei Zhang zhangl@ntu.edu.sg; phone: (+65) All errors and omissions are our own.

2 Leveling Playing Field or Obfuscation: The Informational Role of Overconfident CEOs Abstract We study the informational role of overconfident CEOs. We find that stocks led by overconfident CEOs have lower analyst forecast dispersion, higher breadth of ownership, and lower informed trading intensity. These results are consistent around CEO turnovers. We also show that overconfident CEOs level the playing field between short sellers and other investors, as the (good) news in short interest (Boehmer, Huszar and Jordan, 2010) exists only among stocks with non-overconfident CEOs and disappears among stocks with overconfident CEOs. These findings suggest that overconfident CEOs can help increase information efficiency and lower the mispricing of their company shares. Keywords: Overconfident CEOs, Differences of Opinion, Informed Trading, Short Interest, Information Efficiency JEL Classification: D80, G14, G32, G34

3 1. Introduction Recent research in finance shows that managerial traits such as managers prior experience, personal characteristics, and cognitive biases affect corporate decisions (e.g., Bertrand and Schoar, 2003). Of the cognitive biases, CEO overconfidence has received perhaps the most attention (Malmendier and Tate, 2015). The extant literature examines the effect of overconfident CEOs on corporate policies such as investments, financing decisions, and innovation, 1 but their impact on a firm s information environment has gone unexplored. This is an important consideration because the information environment affects the firm s cost of capital, investor base, and price efficiency (e.g., Diamond and Verrechia, 1991). In this paper, we provide the first direct evidence on the overall impact of CEO overconfidence on investor disagreement, informed trading, and the information efficiency of stock prices. There are two competing views on the informational role of overconfident CEOs. First, extant literature shows that because overconfident CEOs view their company shares as undervalued, they tend to release more voluntary managerial disclosure and earnings guidance (Hribar and Yang, 2015), exhibit less asymmetric timeliness in the recognition of good news versus bad news 2 (Ahmed and Duellman, 2013), and are more likely to repurchase shares from the market (Banerjee, Humphery-Jenner, and Nanda, 2013). This increase in information disclosure may facilitate more timely incorporation of information into stock prices and lower information asymmetry between informed and uninformed investors. As a result, firms led by overconfident CEOs will have lower investor disagreement, lower informed trading intensity, and higher information efficiency of stock prices. We refer to this hypothesis as the leveling playing field hypothesis. Alternatively, studies also show that overconfident CEOs have a tendency to misstate earnings in corporate financial reporting (Schrand and Zechman, 2011), issue overly optimistic estimates (Hribar and Yang, 2015), pursue less conservative accounting practices (Ahmed and Duellman, 2013), and are more 1 See, among others: Malmendier and Tate (2005, 2008); Graham, Harvey, and Puri (2009); Hirshleifer, Low, and Teoh (2012). 2 Barth, Landsman, Raval and Wang (2015) find that higher asymmetric timeliness in news recognition decreases the speed with which equity investor disagreement and uncertainty resolve at earnings announcements. 1

4 likely to be involved in securities class action litigations (Banerjee, Humphrey-Jenner, Nanda, and Tham, 2015). From this perspective, overconfident CEOs may obfuscate information quality, potentially creating a more opaque information environment and increasing information asymmetry between informed and uninformed investors. This information obfuscation hypothesis predicts that firms with overconfident CEOs will have higher investor disagreement, higher informed trading intensity, and lower information efficiency of stock prices. To test these two competing hypotheses, we use a comprehensive sample of 30,149 firm-year observations in the S&P Execucomp database between 1992 and We identify overconfident CEOs by their willingness to hold deep in-the-money vested stock options. 3 In the first half of our analysis, we directly examine the informational role of overconfident CEOs by linking CEO overconfidence with measures of investor disagreement and informed trading intensity. Also, we examine how these measures change around CEO turnovers. In the second half our analysis, we analyze the effect of CEO overconfidence on the information efficiency of stock prices by focusing on a particular form of informed trading, short selling. Overconfident CEOs may marginalize the information advantages of short sellers if their actions level the playing field between informed and uninformed investors. On the other hand, short sellers may stand to benefit if the actions of overconfident CEOs obfuscates the information environment. We briefly summarize our main results. We find that overconfident CEOs are associated with lower analyst forecast dispersion, higher breadth of ownership, and lower informed trading intensity. Investor disagreement and informed trading increase when an overconfident CEO is replaced by a nonoverconfident CEO and decrease when a non-overconfident CEO is replaced by an overconfident CEO. These findings unambiguously support the leveling playing field hypothesis. Consistent with this evidence, our second set of results show that overconfident CEOs marginalize the information advantages of short sellers, as the relation between short interest and future returns disappears among stocks with 3 Specifically, we define a CEO as overconfident once he or she postpones the exercise of vested stock options that are at least 67% in the money. Once a CEO is identified as overconfident, he or she remains overconfident for the rest of the sample period. This commonly used option-based measure is based on the idea that managers who are overconfident in their firm s future prospects will voluntarily maintain excessive exposure to the idiosyncratic risk of the firm (e.g., Malmendier and Tate, 2005; Hirshleifer, Low, and Teoh, 2012; Ahmed and Duellman, 2013). 2

5 overconfident CEOs. Further supporting the leveling playing field hypothesis, these results suggest that CEO overconfidence helps to correct mispricing and improves the information efficiency of stock prices. We start by examining the effect of overconfident CEOs on investor disagreement among investors using two commonly used measures of differences of opinion. The results indicate that the stocks of overconfident CEOs have significantly lower analyst forecast dispersion (e.g., Diether, Malloy, and Scherbina, 2002). Stocks with overconfident CEOs display a 21% lower analyst forecast dispersion relative to the sample average. 4 One concern with examining analyst forecast dispersion is that analysts are not buy-side investors in the stock market. To address this concern, we use an alternative measure based on the breadth of ownership by active U.S. equity mutual funds (e.g., Chen, Hong, and Stein, 2002). We find that the stocks of overconfident CEOs display a 6% higher breath of ownership relative to the sample average compared to the ones with non-overconfident CEOs. We also examine two trading-based measures that capture the intensity of informed trading in the market. The evidence indicates that the stocks of overconfident CEOs have 5% lower probability of informed trading (PIN) relative to the sample average (e.g., Easley, Hvidkjaer, and O Hara, 2002; Brown and Hillegest, 2007). A potential concern with the PIN measure is that it is less applicable in highly automated markets. Therefore, we construct an alternative informed trading measure, C2, based on the return autocorrelation conditional on trading volume (e.g., Llorente, Michaely, Saar, and Wang, 2002). We find stocks of overconfident CEOs exhibit a significant reduction in C2 of 40% relative to the sample average. The strong association between CEO overconfidence and the firm s information environment supports the leveling the playing field hypothesis. However, simultaneity and reverse causality are potential sources of concern for the interpretation of our findings. These issues are partially attenuated because we examine aspects of the information environment that are formed by investor beliefs and trading behavior, not direct managerial actions. Concerns of unobserved firm characteristics are further 4 Our tests control for firm size, leverage, profitability, institutional ownership, stock volatility, past stock return and trading volume in our empirical analyses. The results remain robust with the inclusion of industry fixed effects or firm fixed effects. 3

6 alleviated by the use of firm fixed effects in our earlier tests. Nonetheless, we recognize the possibility that overconfident CEOs may self-select or are hired into firms with higher transparency and lower information asymmetry. To rule out this explanation, we follow the literature (e.g., Ahmed and Duellman, 2013) and examine changes in the information environment around CEO turnover. If the previous results are due to CEO-firm matching, we expect no change in our set of information asymmetry measures after the new CEO takes over. However, we find that investor disagreement and informed trading increase after an overconfident CEO is replaced by another non-overconfident new CEO and decrease after a nonoverconfident CEO is replaced by another overconfident new CEO. These patterns are consistent for both voluntary and forced CEO turnovers, which alleviates the simultaneity concern that poor recent performance worsens the information environment and causes the overconfident CEO to be fired. Overall, the evidence supports the leveling playing field hypothesis of the informational role of overconfident CEOs. Building on the overconfidence/transparency relation, we conduct a market based test on whether overconfident CEOs level the playing field between short sellers and other investors. This is an ideal setting for our research question because: 1) short sellers are generally viewed as more sophisticated and more informed; 2) short interest data is public information; 3) market participants directly observe short sellers action as revealed by short interest; 4) short interest predicts future returns (e.g., Boehmer, Jones, and Zhang, 2008; Boehmer, Huszar, and Jordan, 2010). Perhaps surprisingly, Boehmer, Huszar, and Jordan (2010) show that short interest predicts high future returns of lightly shorted stocks suggesting that the information content of short interest is not hindered by short sell constraints. To test the effect of overconfident CEOs on information efficiency, we compare the return predictability of short interest between stocks with and without overconfident CEOs. We find that the (good) news in short interest only exists within stocks with non-overconfident CEOs and disappears among stocks with overconfident CEOs. Results from independent sorts show that the long-short portfolio (sorted by short interest in the previous month) based on the 3-4

7 factor plus momentum factor model yields an insignificant alpha of 0.26% per month for stocks led by overconfident CEOs, compared to a highly significant alpha of 0.72% per month for stocks with nonoverconfident CEOs. The results are similar using a 5-factor plus momentum factor model, with the long-short portfolio yielding an insignificant alpha of 0.27% per month for overconfident CEOs compared to a highly significant alpha of 0.76% per month for non-overconfident CEOs. Moreover, we find that this difference in return predictability in short interest between overconfident CEOs and nonoverconfident CEOs is largely driven by the lowest short interest portfolio. This suggests that our findings are not explained by short sell constraints. Using a 5-factor plus momentum factor model, the lowest short interest portfolio yields an insignificant alpha of 0.18% per month for overconfident CEOs, compared to a highly significant alpha of 0.46% per month for non-overconfident CEOs. These results are similar using dependent sorts and matching approach based on firm characteristics. These results are important for two reasons. First, it provides direct market based evidence that CEO overconfidence levels the playing field among informed and uninformed investors. Second, from an information efficiency standpoint, it implies that CEO overconfidence is associated with less mispricing. For companies led by overconfident CEOs, the non-relation between short interest and future returns implies that the positive information in lightly shorted stocks has already been incorporated into the stock price. This suggests that overconfident CEOs play a role in correcting mispricing and improving the information efficiency of their stock. 5 We perform additional tests to ensure that our results are robust. First, our results are similar using alternative cutoffs for the identification of overconfident CEOs or alternative measures of short interest. Second, it is possible that firms tend to hire overconfident CEOs in good times and fire overconfident CEOs in bad times, creating additional firm-ceo matching issues when the CEO is first identified as overconfident. Following Hirshleifer, Low, and Teoh (2012), we address this concern by dropping the 5 It is important to note that the leveling playing field hypothesis does not predict the level of short interest in stocks with overconfident CEOs because lower level of short interest does not imply less informed trading. Rather, low short interest suggests that short sellers, being more informed and sophisticated, do not short the stocks that they have positive private information. Overconfident CEOs, by leveling the playing field, reduce the information advantage of informed vis-à-vis uninformed investors, which leads to less return predictability of short interest. 5

8 initial three years after a CEO is first identified as overconfident. Our results are not sensitive to this alternative specification. Our paper contributes to the literature on CEO overconfidence. Studies show that overconfident CEOs affect information disclosure but these channels often generate mixed implications on the overall informational role of overconfident CEOs. On the one hand, overconfident CEOs release more managerial earnings guidance (Hribar and Yang, 2015), exhibit less asymmetric earnings timeliness (Ahmed and Duellman, 2013), and are more likely to repurchase shares (Banerjee, Humphery-Jenner, and Nanda, 2013). On the other hand, overconfident CEOs are more likely to misstate earnings (Schrand and Zechman, 2011), issue overly optimistic estimates (Hribar and Yang, 2015), and attract class action litigation (Banerjee, Humphrey-Jenner, Nanda, and Tham, 2015). Our findings are unique in that we examine the overall informational role of overconfident CEOs and its consequence on the information efficiency of stock prices. We find that CEO overconfidence has a bright side in terms of improving information efficiency and lowering mispricing (underpricing). These results shed new light to our understanding of the puzzle on why firms are willing to hire or keep overconfident managers (e.g., Goel and Thakor, 2008; Gervais, Heaton, and Odean, 2011). Our results also add to the asset pricing literature on corporate transparency, informed trading, and short selling (e.g., Chen, Hong, and Stein, 2002; Diether, Malloy, and Scherbina, 2002; Zhang, 2006). Prior studies find evidence that stocks with high short interest subsequently underperform (e.g., Senchack and Starks, 1993; Asquith and Meulbroek, 1995; Desai, Ramesh, Thiagarajan, and Balachandran, 2002; Boehmer, Jones, and Zhang, 2008). More recently, Boehmer, Huszar, and Jordan (2010) show that lightly shorted stocks experience positive future abnormal returns. This raises a broader issues regarding market efficiency because it implies that information is slow to incorporate into prices, even for a long only strategy. Our evidence shows that the good news in short interest exist only within stocks with nonoverconfident CEOs and disappear among stocks led by overconfident CEOs. This suggests that overconfident CEOs help to level the playing field between informed and uninformed investors and increase the timeliness of (positive) information incorporation into stock prices. 6

9 Third, our results contribute to a broad literature on behavioral finance and market efficiency. A large body of evidence suggests that behavioral biases of investors affect asset prices and reduce market efficiency (e.g., Barberis and Thaler, 2003). We depart from this literature by examining the effect of managerial biases on asset prices with the existence of information frictions. To our knowledge, we are among the first to show that overconfident CEOs may help the market to correct mispricing, particularly undervaluation. This is important from a welfare perspective because undervaluation may increase the cost of capital and waste valuable investment opportunities. Our findings imply that a behavioral bias of CEO can actually reduce the barriers to arbitrage and improve informational efficiency. 2. Data and Variables This section describes the data sources and variable construction. We also provide summary statistics of our sample. 2.1 Data and variables The data on CEO option holdings and other CEO characteristics are obtained from S&P Execucomp database during the period of 1992 to We collect data on daily and monthly stock returns, monthly short interest, trading volumes and shares outstanding are from CRSP. The data on annual accounting information are from Compustat. The data on stock holdings of mutual funds and institutional investors are from Thomson Reuters. The data on analyst earnings forecasts are from IBES. To identify CEO overconfidence, we follow the most commonly used approach in the literature based on the exercise of vested CEO stock options. As we do not have the dataset used in Malmendier and Tate (2005, 2008), we estimate CEO overconfidence following the approach of Hirshleifer, Low, and Teoh (2012) and Campbell et al. (2011). 6 Specifically, we use the Execucomp data to construct the 6 Similar to Hirshleifer, Low, and Teoh (2012), we do not require that the CEO holds a 67% in the money option at least twice and define the CEO as overconfident after the first time he or she exhibits such a behavior. The reason is that this requires the use of forward-looking information, which is not appropriate for us to examine the return predictability of short interest. 7

10 overconfidence measure. First, we divide the value of exercisable unexercised options (Execucomp item: opt_unex_exer_est_val) by the number of exercisable unexercised options (Execucomp item: opt_unex_exer_num) and subtract this value from the stock price at the fiscal year end (Compustat item: PRCC_F) to obtain the average strike price per option. Second, we divide the value of exercisable unexercised options per option by the average strike price per option to calculate the average moneyness of the options. We define a CEO as overconfident once he or she postpones the exercise of vested stock options that are at least 67% in the money, following the cut-off in Malmendier and Tate (2005, 2008). The Options67 variable takes the value of 1 if the CEO is identified as overconfident and 0 otherwise. We only include the vested options held by the CEO because we are interested in options that the CEO can exercise. Once a CEO is identified as overconfident, he or she remains overconfident for the rest of the sample period. We construct Options100 following the same methodology as Options67, except that we identify a CEO as overconfident once he or she postpones the exercise of vested options that are at least 100% in the money. Similarly, once a CEO is identified as overconfident, he or she remains overconfident for the rest of the sample period. We construct two measures of investor disagreement. The first measure is analyst forecast dispersion. For a given firm-month, we define analyst forecast dispersion as the ratio between the standard deviation of analyst one-year EPS forecasts divided by the absolute mean forecast. We require at least five analysts to calculate the measure. Then, we calculate the yearly average across the monthly observations. The second measure is based on the breadth of ownership by active U.S. equity mutual funds. 7 For each firm-quarter, we define the breadth of ownership as the number of mutual funds holding the stock divided by the total number of mutual funds in that quarter. Then, we calculate the yearly 7 We only focus on U.S. domestic equity mutual funds. For data from 1993 to 1999, we select funds with IOC fund objective codes (available until 1999 in the Thomson data) 2, 3, 4, and 7: Aggressive growth, Growth, Growth & Income, and Balanced. For data from 2000 to 2013, we select funds with the Lipper objective codes that pertain to domestic equity funds from the CRSP Mutual Fund database: G, SG, MC, SP, I, B, GI, FX, EI, TK, H, S, CS, UT, TL, CA, DSB, LSE, ID, BM, and CG. We use the MFLinks data to match the Thompson data with the CRSP Mutual Fund data. 8

11 average across four quarters. Our two measures of informed trading intensity are PIN and C2. The probability of informed trading (PIN) measures the unconditional probability of information-based trading for a given stock based on the observed order flow. We follow the construction procedure detailed in Brown and Hillegeist (2007). 8 The second measure is the C2 coefficient based on stock return autocorrelation conditional on trading volume developed in Llorente et al. (2002). 9 For a given firm-month, we define Short Interest as the amount of shorted shares divided by the trading volume of the month or, alternatively, divided by the total shares outstanding. We control for the following firm characteristics. Firm Size is the log value of book assets. Marketto-book is market value of assets divided by book assets. Book Leverage is total debt divided by book assets. Profitability is operating income before depreciation divided by book assets. Sales Growth is current year sales less prior year sales scaled by prior year sales. Institutional Ownership is the number of shares held by all of the institutional investors divided by the total number of shares outstanding. Yearly Return is the cumulative stock return in a year. Return Volatility is the standard deviation of monthly stock returns in a year. Trading Volume is the average monthly trading volume in a year. We estimate monthly trading volume as the number of shares traded in the month divided by the total shares outstanding. The detailed definitions of each variable together with Compustat item symbols are available in the appendix. We also control for important CEO characteristics. CEO Tenure is the logarithm of the number of years since the CEO resumes office. CEO Age is the logarithm of the CEO s age. CEO Ownership is the number of stocks held by the CEO divided by the number of shares outstanding. Initial CEO is a dummy variable that is equal to 1 if the CEO is the same CEO when the firm first appears in the Execucomp database and 0 otherwise. Chairman CEO is a dummy variable that is equal to 1 if the CEO is also the 8 The quarterly data on the PIN measure from 1993 to 2010 are obtained from Stephen Brown s website at 9 Specifically, for each firm-year, we estimate the coefficient C2 in the time-series regression:, 0 1, 2,,,, where, is the daily stock return of firm i on day t+1, is the value-weighted market return,, is the detrended logarithm of stock turnover on day t, by subtracting a 100 trading day moving average. We calculate daily stock turnover as the number of shares traded on that day divided by the total number of shares outstanding. 9

12 chairman of the board and 0 otherwise. 2.2 Summary Statistics In Table I, we report the summary statistics of the main variables used in the subsequent analyses. The overall sample includes 30,149 firm-year observations during the period of 1992 to The mean value of Options67 is 0.58, meaning that 58% of CEOs are identified as overconfident. This figure is comparable to the sample average reported in Hirshleifer, Low, and Teoh (2012). The alternative measure Options100 has a mean of 0.46, implying that 46% of CEOs are identified as overconfident. After merging with firm characteristics and other CEO characteristics, the combined sample includes 28,086 firm-year observations. The mean and median size of book assets is $13.8 billion and $1.7 billion respectively, consistent with the fact that firms covered in Execucomp are relatively large firms. The average CEO tenure is 7.8 years and the average CEO age is 56 years. Around 56% of the CEOs serve as the chairman of the board, 2% are female CEOs, and the average CEO stock ownership is 1.8% of the firm s shares outstanding. Table II reports the percentage of overconfident CEOs year by year. In each year from 1992 to 2012, we report the mean and standard deviation of two CEO overconfidence measures, Options67 and Options100. The time-series pattern shows that CEOs tend to become more overconfident during market booming period and tend to be less so during market downturns. This pattern is also consistent with the one reported in Hirshleifer, Low, and Teoh (2012) Overconfident CEOs and the Information Environment The extant literature shows that CEO overconfidence affects information disclosure, but the implications on the informational role of overconfident CEOs are mixed. 11 For example, Hribar and Yang (2015) show 10 This may raise some firm-ceo matching issues that CEO overconfidence can be spuriously correlated with market conditions. It could be that firms tend to hire overconfident CEOs in good times and fire overconfident CEOs in bad times. Following Hirshleifer, Low, and Teoh (2012), we address this concern by dropping the first three years since a CEO starts to be identified as overconfident. We report these results in later analyses. 11 Confirming prior studies, we verify that overconfident CEOs are more likely to provide voluntary earnings 10

13 that overconfident CEOs are more likely to provide company-issued earnings guidance, but their managerial forecasts tend to be overly optimistic. Ahmed and Duellman (2013) find that overconfident CEOs use less conservative accounting in the form of both conditional and unconditional conservatism. The informational implications are unclear because less asymmetric timeliness in the recognition of good versus bad news may quicken the resolution of investor disagreement and uncertainty at earnings announcements (Barth, Landsman, Raval, and Wang, 2015). On the other hand, less accounting conservatism may be associated with higher earnings management (Beaver and Ryan 2005), resulting in a greater likelihood that overconfident CEOs misstate earnings (Schrand and Zechman, 2011). These findings raise the important question of the overall informational role of overconfident CEOs. The leveling playing field hypothesis suggests that the disclosure provided by overconfident CEOs improves the transparency of the information environment. In contrast, the information obfuscation hypothesis implies that information environment becomes more opaque as a result of the type of disclosure overconfident CEOs release. To analyze these two competing hypotheses, we examine two broad dimensions of the information environment: 1) investor disagreement and 2) informed trading intensity. We analyze issues associated with CEO-firm matching by studying the changes in information environment around CEO turnover. 3.1 Overconfident CEOs and investor disagreement We begin by examining the effect of overconfident CEOs on the investor disagreement by estimating regressions following equation (1): Investor Disagreementit, Overconfident CEOit, 1 X it, 1 it,. (1) Our main CEO overconfidence measure is Options67. Investor Disagreement is measured using either analyst forecast dispersion (e.g., Diether, Malloy, and Scherbina, 2002) or breadth of mutual fund ownership (e.g., Hong, Kubik, and Stein, 2002). We control for firm characteristics including book size, guidance, announce stock repurchases and use less conservative accounting in terms of asymmetric timeliness of earnings in our sample. The results are available in the Internet Appendix. 11

14 market-to-book, book leverage, profitability, sales growth, institutional ownership, return volatility, and trading volume. We also control for past stock returns following Hirshleifer, Low, and Teoh (2012) and Malmendier and Tate (2015). Our regressions include year fixed effects to capture time-varying macroeconomic shocks and industry fixed effects to capture industry factors. We also consider a specification with firm fixed effects to alleviate concerns that certain unobserved firm characteristics are behind our results. Standard errors are clustered by firm. Panel A of Table III shows a negative and significant effect of CEO overconfidence on analyst forecast dispersion. Relative to the sample average, the parameter estimate in Column (1) shows that an overconfident CEO is associated with a decrease in analyst forecast dispersion by 22% (t=-5.82). In Column (2) the results are unchanged (21%, t=-5.83) with the inclusion of industry fixed effects. As analyst dispersion may be related to other CEO characteristics, we include CEO tenure, age, ownership, founder, chairman, and female indicators. Column (3) shows the results are similar with the inclusion of these characteristics. We also estimate a specification with firm fixed effects to capture unobserved firm heterogeneity. The results remains similar (20%, t=-4.61) in Column (4), suggesting that unobserved firm heterogeneity is not behind our findings. The results also show a positive relation between analyst forecast dispersion and trading volume. This is consistent with the view that forecast dispersion measures investor disagreement because disagreement generates greater trading volume. Analyst forecast dispersion is also associated with higher return volatility and firm leverage, and smaller firm size, lower profitability and low past stock returns, consistent with the view that larger and more profitable companies have more transparent information environment. One concern with using analyst forecast dispersion to measure investor disagreement is that analysts are not buy-side investors in the stock market. We address this concern using an investor-based measure that captures the breadth of ownership by active U.S. equity mutual funds (e.g., Chen, Hong, and Stein, 2002). Diamond and Verrecchia (1991) show that a reduction of information asymmetry increases demand from many investors which leads to a higher breadth of ownership. 12

15 The results in Panel B of Table III show a significant positive effect of CEO overconfidence on the breadth of ownership. The results are consistent across different specifications. For example, with year and industry fixed effects in Column (3), firms with overconfident CEOs display a 4% higher breadth of ownership relative to the sample average with a t-statistic of We note that the effects are stronger in the specification with firm fixed effects in Column (4). Firms with overconfident CEOs display a 6% higher breadth of ownership relative to the sample average with a t-statistic of This shows that the effect of CEO overconfidence is unlikely to be driven by certain unobserved firm characteristics. Among the control variables, firm size, market-to-book, profitability, institutional ownership, and past stock return are positively related to breadth of ownership, while trading volume, book leverage, and CEO ownership are negatively related to breadth of ownership. In sum, these results support the leveling playing field hypothesis and refute the information obfuscation hypothesis. Another implication of the leveling playing field hypothesis is that stocks with overconfident CEOs should display a lower level of informed trading intensity because of a decrease in the information advantage of informed investors over uninformed investors. We next examine the effect of overconfident CEOs on informed trading intensity. 3.2 Overconfident CEOs and informed trading intensity To examine the relation between CEO overconfidence and informed trading intensity, we estimate regressions following equation (2): Informed Trading Intensityi,t Overconfident CEOi,t-1 X it, (2) 1 it, We focus on the main CEO overconfidence measure Options67. We consider two general measures of informed trading intensity: 1) the probability of informed trades (PIN) based on observed order flow, 2) the return-volume coefficient C2 based on autocorrelation of stock returns conditional on trading volume. In Panel A of Table IV, the dependent variable is the probability of information-based trading (PIN) based on the observed order flow. The procedures to construct PIN are detailed in Brown and Hillegeist 13

16 (2007). We calculate the yearly average across 4 quarters for a given firm-year. Column (1) reports the base-line specification. Column (2) includes industry fixed effects at the two-digit SIC level. Column (3) further controls for CEO characteristics. In column (4), we include firm fixed effects. The standard errors are always clustered at the firm level. The results show a significantly negative relation between CEO overconfidence and the probability of informed trading. The results are consistent across different specifications. Under the most complete specification with year and industry fixed effects, firms with overconfident CEOs display a 3% lower intensity of informed trades relative to the sample average with a t-statistic of This result is consistent in the specification with firm fixed effects. In this case, firms with overconfident CEOs display a 5% lower intensity of informed trades relative to the sample average with a t-statistic of Among the control variables, firm size, market-to-book, profitability, sales growth, institutional ownership, past stock return and trading volume are negatively related to PIN, while book leverage, CEO age and CEO ownership are positively related to PIN. In Panel B, Table IV, the dependent variable is the return-volume coefficient C2 based on autocorrelation of stock returns conditional on trading volume (Llorente et al., 2002). They show that a correspondence exists between the cross-sectional variation in return-volume dynamics and the relative importance of informed trading in stock returns. The higher the C2 coefficient, the higher the intensity of information based trading. Panel B follows the same layout as in Panel A. The results show a significantly negative relation between CEO overconfidence and the C2 coefficient. The results are consistent across different specifications. In the specification with year and industry fixed effects, firms with overconfident CEOs display a 70% lower C2 relative to the sample average 12 with a t-statistic of This effect remains significant in the specification with firm fixed effects, in which firms with overconfident CEOs display a 40% lower C2 relative to the sample average with a t-statistic of Among the control variables, market-to-book, sales growth, past stock return and trading volume are negatively related to C2 across different specifications. 12 The average C2 is 0.01 in our sample, consistent with Llorente et al. (2002). 14

17 3.3 Changes in the information environment around CEO turnover The previous results indicate a strong association between overconfident CEOs and a firm s information environment. However, simultaneity and reverse causality are potential sources of concern for the interpretation of these findings. The endogeneity concerns are alleviated by the use of firm fixed effects in our tests. Still, we recognize that an alternative explanation is that overconfident CEOs may self-select or are hired into firms with more transparency and lower information asymmetry. To rule out this explanation, we examine changes in the information environment around CEO turnover. If our previous findings are explained by CEO-firm matching, we expect no change in the information environment measures after the new CEO takes over. We obtain the data on CEO turnover including both voluntary turnover and forced turnover for all firms in the S&P Execucomp database during the period of 1993 to 2010 from Jenter and Kanaan (2015). We focus on the five-year period both around turnover from an overconfident incumbent CEO to a nonoverconfident new CEO, and around turnover from a non-overconfident incumbent CEO to an overconfident new CEO. For example, in terms of analyst forecast dispersion, we report the average forecast dispersion in the two years before the turnover year, the average forecast dispersion in the two years after the turnover year, and we use a t-test to test the statistical differences. The results on breadth of ownership, PIN and C2 are reported accordingly. The number of observations and t-statistics are given in parentheses. Table V presents the results. In Panel A, we include all of the CEO turnover events. We find evidence of significant changes in the information measures around CEO turnovers, even in a significantly reduced sample. We find that investor disagreement and informed trading increase after an overconfident CEO is replaced by another non-overconfident new CEO, and decrease after a nonoverconfident CEO is replaced by another overconfident new CEO. For example, we find that when an overconfident incumbent CEO is replaced by another non-overconfident new CEO, analyst forecast dispersion increases from to 0.198, breadth of ownership decreases from 6.96% to 6.72%, PIN 15

18 remains relatively stable from to 0.127, and C2 increases from to When a nonoverconfident incumbent CEO is replaced by another overconfident new CEO, analyst forecast dispersion decreases from to 0.125, breadth of ownership increases from 6.31% to 6.93%, PIN decreases from to 0.128, and C2 decreases from to One remaining concern is that poor recent stock performance may simultaneously worsens the information environment and causes overconfident CEOs to be fired. Therefore, we examine voluntary CEO turnovers and forced CEO turnovers separately. In Panel B, we only consider voluntary CEO turnovers, and in Panel C, we focus on forced CEO turnovers. We find similar patterns for both voluntary and forced CEO turnovers. Overall, we conclude that the evidence supports the leveling playing field hypothesis of the informational role of overconfident CEOs. 4. Overconfident CEOs and the News in Short Interest Our second set of tests builds upon the overconfidence/transparency relation in Section 3. We examine the effect of overconfident CEOs on the information efficiency of their stock price by testing whether overconfident CEOs level the playing field among short sellers and other investors. Short sellers are an ideal investor group to examine this research question for the following reasons. First, it is a common view that short sellers are sophisticated and informed. Short-selling is risky and costly and requires superior information. Second, the data on short interest are publicly available information and observable by market participants. This is particularly important in our setting because it provides a direct measure of the informed traders in a stock. Third, evidence suggests that short interest predicts future abnormal returns (e.g., Boehmer, Jones, and Zhang, 2008; Boehmer, Huszar and Jordan, 2010), confirming the perception that short-sellers are informed. Fourth, Boehmer, Huszar, and Jordan (2010) show that lightly shorted stocks experience higher future returns. The undervaluation of low short interest stocks implies that positive information is slow to incorporate into prices. This is a surprising result because short interest is a publicly available signal and only requires a long position investment to capture the abnormal returns. 16

19 Using the short interest setting developed in Boehmer, Huszar, and Jordan (2010), we examine the effect of overconfident CEOs on the undervaluation of low short interest stocks. We expect no undervaluation among these stocks if overconfident CEOs are able to level the playing field between informed and uninformed investors. Before we begin our main analysis, we first examine the overall relation between short interest and future return in our sample. 4.1 The overall relation between short interest and future returns At each month-beginning from January 1993 to December 2013, stocks are sorted into quintiles based on the previous month-end short interest. We use the main measure on short interest defined as the amount of shorted shares divided by the trading volume of the month. Portfolio 1 has the lowest short interest. Portfolio 5 has the highest short interest. Equally-weighted returns for the five portfolios are calculated over the month. Panel A1 of Table VI reports the mean and standard deviation of short interest in each portfolio. We also report the percentage of overconfident CEOs in each portfolio. We focus on our main overconfidence measure Options67 throughout the table. The results show that the average short interest in Portfolio 1 is with a standard deviation of 0.029, while the average short interest in Portfolio 5 is with a standard deviation of The average percentage of firms with overconfident CEOs is similar across the five portfolios, being 54.1% in Portfolio 1 and 56.7% in Portfolio 5. In Panel A2, for each portfolio, we report the raw average portfolio return, the abnormal return (i.e., alpha) from the 3-factor (market factor, SMB, HML) model, and the alpha from the Fama- French 3-factor and the Carhart momentum factor model, and the alpha from the 5-factor (market factor, SMB, HML, RMW, CMA) and the Carhart momentum factor model. Long Portfolio 1 & Short Portfolio 5 is the difference in the returns between the lowest and highest short interest portfolios. We report the raw return and the alphas for the long-short portfolio accordingly. We find that short interest predicts future returns, consistent with Boehmer, Huszar, and Jordan (2010). Our sample extends their original time period by 8 years (until 2013) and only focus on the stocks 17

20 covered in the S&P Execucomp database. Using a 3-factor plus momentum factor model, the long-short portfolio yields a significant alpha of 0.48% per month, driven mostly by the positive alpha of 0.46% per month in the lowest short interest portfolio. Based on the 5-factor plus momentum factor model, the long-short portfolio yields a significant alpha of 0.51% per month, contributed both from a significant positive alpha of 0.31% in the lowest short interest portfolio and a marginally significant negative alpha of -0.20% in the highest short interest portfolio. 4.2 CEO overconfidence and the relation between short interest and future returns We perform our main test by separating firms with overconfident CEOs and firms with non-overconfident CEOs. We use Options67 throughout this analysis. A CEO is identified to be overconfident if the indicator variable Options67 is equal to one in the previous year. We examine raw and risk-adjusted returns of stock portfolios sorted by short interest, both independently and dependently conditioning on firms with overconfident and non-overconfident CEOs. We report the results in Panel B and C of Table VI. Panel B reports the results of independent sorting, in which we first sort stocks into five portfolios independently, without conditioning on whether the stock has an overconfident CEO. Then, we calculate portfolio returns separately for stocks with overconfident CEOs and stocks with non-overconfident CEOs. For each portfolio, we report the average short interest, the raw average portfolio return, the abnormal return (i.e., alpha) from the 3- factor model, and the alpha from the 3-factor plus the Carhart momentum factor model, and the alpha from the 5-factor plus the Carhart momentum factor model. Long Portfolio 1 & Short Portfolio 5 is the difference in the returns between the lowest and highest short interest portfolios. The evidence shows that the return predictability in short interest is nearly non-existent among stocks with overconfident CEOs, and strongly evident among stocks with non-overconfident CEOs. Panel B2 shows that for the non-overconfident sample of firms, the long-short portfolio is positive and statistically significant across all four alpha measures. For example, in the second to last column in Panel B2 (the 3-factor plus momentum factor model), the long-short portfolio earns an average 18

21 alpha of 0.72% (t=4.11) per month. In comparison, Panel B1 shows that the long-short portfolio alpha is statistically insignificant in three out of four specifications, and is notably weaker after controlling for momentum. This is consistent with the argument in Malmendier and Tate (2015) that it is important to control for past returns when using the Options67 measure. In the second to last column in Panel B1, the long-short portfolio earns an average of 0.26% (t=1.48) per month. The results are similar if using a 5-factor plus momentum factor model, where the long-short portfolio yields an insignificant alpha of 0.27% per month (t=1.43) for overconfident CEOs, compared to a significant alpha of 0.76% per month (t=3.99) for non-overconfident CEOs. Dependent sorts generate similar result as shown in Panel C of Table VI. The long-short portfolio based on the 3-factor plus momentum factor model yields an insignificant alpha of 0.26% per month (t=1.52) for overconfident CEOs, compared to a significant 0.75% per month (t=4.12) for nonoverconfident CEOs. The results are also similar if using a 5-factor plus momentum factor model, where the long-short portfolio yields an insignificant alpha of 0.30% per month (t=1.60) for overconfident CEOs, compared to a significant 0.79% per month (t=4.04) for non-overconfident CEOs. In Panel D, we test the statistical differences of raw and risk-adjusted returns between the long-short portfolio sorted among stocks with overconfident CEOs and the long-short portfolio sorted among stocks with non-overconfident CEOs. We report the results for both the independent sorting and the dependent sorting (as in Panel B and Panel C). The results show that the differences in raw and risk-adjusted returns between the long-short portfolio sorted from stocks with overconfident CEOs and the one sorted from stocks with non-overconfident CEOs are always statistically significant. Importantly, we find that this difference in the return predictability of short interest between overconfident CEOs and non-overconfident CEOs is largely driven by the lowest short interest portfolio. Using a 5-factor plus momentum factor model, in the case of independent sorting (Panel B), the lowest short interest portfolio yields an insignificant alpha of 0.18% per month (t=1.19) for overconfident CEOs, compared to a highly significant alpha of 0.46% per month (t=3.10) for nonoverconfident CEOs. Similarly, in the case of dependent sorting (Panel C), the lowest short interest 19

22 portfolio yields an insignificant alpha of 0.18% per month (t= 1.19) for overconfident CEOs, compared to a highly significant alpha of 0.49% (t=3.23) per month for non-overconfident CEOs. 4.3 Robustness tests Our results are also robust to scaling short interest by total shares outstanding. Panel A of Table VII shows that the independently sorted long-short portfolio yields an alpha of 0.36% per month (t=1.80) for overconfident CEOs, compared to a highly significant alpha of 0.69% per month (t=3.53) for nonoverconfident CEOs based on 5-factor plus momentum factor model. Similarly, Panel B shows that in the case of dependent sorting, the long-short portfolio yields an insignificant alpha of 0.28% per month (t=1.41) for overconfident CEOs, compared to a highly significant alpha of 0.80% (t=4.13) per month for non-overconfident CEOs. One potential concern is that the identification of overconfident CEO may be driven by other major firm characteristics. To deal with this concern, we perform matching sample analyses to examine the return predictability of short interest for the sample of firms with overconfident CEOs and for a matching sample matched by major firm characteristics but without overconfident CEOs. Specifically, we separately perform 1-on-1 propensity score matching (with replacement) by firm characteristics such as firm size, market-to-book, past stock return, stock volatility, trading volume, institutional ownership, CEO ownership and a combination of qualitative CEO characteristics (CEO tenure, CEO age, initial CEO, female CEO and chairman CEO). To ensure the quality of the matching, we calculate the absolute difference in the matching characteristics (propensity scores) between the overconfident CEO stocks and the matched non-overconfident CEO stocks. We drop those observations with the absolute differences above the sample top decile. We report the results in Table VIII. Panel A reports the average matching characteristics for the overconfident CEO stocks and the matched non-overconfident CEO stocks. For each matching sample, we report the average absolute difference in the firm characteristics used as a matching variable and we test the statistical differences with a t-test respectively. The results show that the overconfident CEO 20