The Informativeness of Stale Financial Disclosures

Transcription

1 The Informativeness of Stale Financial Disclosures Michael S. Drake Marriott School of Management Brigham Young University Darren T. Roulstone Fisher College of Business The Ohio State University Jacob R. Thornock Foster School of Business University of Washington October 2012 We are especially grateful to Scott Bauguess and many others at the Securities and Exchange Commission and to Dick Dietrich for assistance in acquiring and implementing the proprietary data used in this study. We thank Scott Bauguess, Anne Beatty, Bob Bowen, Jonathan Brogaard, Nicole Cade, John Campbell, Ed dehaan, Thomas Gilbert, Frank Hodge, Dawn Matsumoto, Kathy Petroni, Terry Shevlin and workshop participants at the 2012 UBCOW Conference, 2012 Chicago Quantitative Alliance Annual Academic Competition, Brigham Young University, Michigan State University, the Division of Risk, Strategy, and Financial Innovation at the U.S. Securities and Exchange Commission, University of Georgia, and University of Hawaii for valuable comments. We thank Isaac Kelly and Joseph Tate for computational assistance and Bret Johnson and Nick Guest for excellent research assistance. Finally, we gratefully acknowledge the financial support of the Marriott School of Management, the Fisher College of Business, and the Foster School of Business.

2 ABSTRACT: Investors have at their fingertips an almost unlimited supply of financial disclosures. Some of these disclosures are recently released, but most have been in the public domain for months or even years. In this study, we evaluate whether such stale disclosures continue to be informative to investors. Using a novel dataset that tracks search requests on the SEC EDGAR database, we find evidence that the acquisition of stale disclosures from EDGAR is positively associated with absolute returns and trading volume within the next two hours. We investigate several potential explanations for why investors might find previously released disclosures informative and find evidence consistent with two explanations: first, investors appear to request stale information in order to provide context for new information releases (e.g., earnings announcements, 8-K filings); second, investors appear to request stale information to resolve cases of high valuation uncertainty. We find no evidence that the relation between requests for stale disclosures and market activity is driven by unsophisticated investors. Our results suggest that financial disclosures continue to remain useful beyond their initial release and highlight the value of historical disclosures and their archives to equity markets. Key Words: Information Acquisition, EDGAR, Trading Volume

3 1. Introduction The set of public information available to an investor is vast, diverse and growing. Decades of capital market research provides evidence that the release of new financial disclosures is associated with increased market activity, suggesting that the disclosures are informative. 1 However, new disclosures make up only a very small fraction of the information set available to investors the overwhelming majority of public information can be considered old. Theoretically, competitive forces quickly drive the net gains from trading on information to zero, thus rendering the information stale. 2 Given the dominance of old information in the total information set, and the considerable resources required to prepare and maintain financial disclosures, it is important to understand whether historical financial disclosures remain informative to investors. Accordingly, the objectives of this paper are twofold. First, we evaluate the informativeness of historical disclosures. Second, we investigate three potential reasons for the informativeness of historical disclosures. The question of whether historical financial disclosures remain informative has been relatively unexplored, perhaps because it is difficult to measure the timing of investors acquisition of historical information. The market s acquisition and use of new financial information is relatively straightforward to measure using standard event study methodologies: new information arrives to the market through an information release and we measure 1 For example, extant research has found that important corporate disclosures such as earnings releases (Beaver [1968]), dividend announcements (Aharony and Swary [1980]), SEC filings (Griffin [2003], Li and Ramesh [2009]), management forecasts (Foster [1973], Patell [1976]) and restatement announcements (Palmrose et al. [2004]) are all associated with short-window stock market activity. 2 We use the label stale because it is commonly used in the literature to describe information that has been in the public domain for a period of time and thus, in theory, should already be reflected in securities prices (e.g., Barberis et al. [1998], Tetlock [2011], and Gilbert et al. [2012]). Throughout the paper, we use the terms stale, old, previously filed, and historical interchangeably. 1

4 investors reaction during the event period (Brown and Warner [1985]; Kothari and Warner [2007]). The assumption is that once information is made publicly available, it is quickly acquired, processed, and used by investors to reallocate capital. However, outside the event window, it is more difficult to tie the actions of investors to information that was previously disclosed. We overcome this measurement problem using a novel dataset that tracks all investor requests for disclosures on the SEC s EDGAR database. The dataset records every click made by an investor to request a regulated filing, such as a 10-K or 8-K, from EDGAR and thereby provides a direct proxy for investor information acquisition. These data allow us to empirically observe the acquisition of previously disclosed financial information without relying on assumptions of horizons or information arrival, as required in an event study. Our proxy for historical information acquisition is the number of investor requests for disclosures that have been publicly available on EDGAR for at least 30 days at the time of the request. Our assumption is that if users request a particular company s disclosure, it is highly likely that they are interested in the information the disclosure contains about the company; in other words, they consider the disclosure to be potentially informative. Given that investors are acquiring the actual SEC filing, we also assume that they are interested in information unique to that filing that cannot be more easily obtained via other financial sources. 3 To test the informativeness of historical disclosures, we examine the association between investor requests for these disclosures from EDGAR and subsequent short-term market activity, 3 This assumption implies that a rational investor chooses the least costly method of acquiring financial information. For example, it is unlikely that an investor would pull a 10-K filing simply to acquire last year s earnings number given that this number is widely reported by other sources. 2

5 measured as absolute returns and trading volume. To strengthen the link between these two measures, we conduct our analysis at the intraday level using hourly requests for disclosures on EDGAR and hourly aggregations of equity market data from the Trade and Quote (TAQ) database. Given the intense computational requirements to analyze intra-day data, we carry out our analyses on a random sample of 200 firms. We interpret historical disclosures to be informative if requests for the disclosures are positively associated with subsequent short-term trading activity. We find that requests for historical disclosures are positively associated with absolute returns and total trading volume in the subsequent two hours. This result holds after we control for EDGAR requests for contemporaneous information, contemporaneous and lagged stock returns and trading volume, firm size, and the inclusion of hour-of-day fixed effects. When we focus on requests for periodic accounting filings (annual reports (10-Ks) and quarterly reports (10-Qs)), our results are uniformly stronger than results for all disclosures, underscoring the importance of historical, periodic accounting reports to equity markets. In sensitivity tests, we find that the association between historical disclosure acquisition and market activity holds even during periods of minimal firm-specific news (i.e., after removing periods when firmspecific news, such as earnings announcements, analyst forecasts, and news articles is released), which suggests that other news events are not entirely driving the association. Overall, our evidence is consistent with the notion that historical information is informative to equity markets. We test three potential reasons for the informativeness of historical financial disclosures. The first explanation is that investors acquire and use old information to provide important 3

6 context for new information releases and current events. For example, investors may turn to previously filed 10-Ks to assess how managers previous discussions of strategic initiatives played out in current periods. We test this notion by examining whether the observed positive association between requests for historical disclosures and subsequent market activity is particularly strong (i.e., more positive) in the presence of two important events, earnings announcements and 8-K filings, while controlling for the normal level of market activity that arises from these events. For earnings announcements, we find that the positive association between requests for historical disclosures and market activities is stronger during a short window around the earnings release. For 8-K filings, we find that the positive association between requests for historical disclosures and market activity is stronger on the date the 8-K is filed. In sum, the evidence is consistent with historical financial information being informative to equity markets when acquired and used in conjunction with current information releases. A second potential reason for the informativeness of historical financial disclosures relates to valuation uncertainty. Our intuition is that valuation is a difficult and subjective process to begin with and is made more difficult and subjective when there is greater uncertainty about the valuation inputs. For certain firms, investors need more time to process and contextualize information, thus investors may continue to find previously filed financial disclosures helpful for valuation assessments. We use three different proxies for valuation uncertainty and find that the positive association between historical information acquisition and market activity is even greater when valuation uncertainty is higher. The third explanation for the association between historical disclosure acquisition and market activity is that unsophisticated investors acquire and trade on historical information 4

7 because they are late to the game that is, they access information in an untimely manner and fail to realize that the information is no longer value-relevant. In our empirical model, we follow prior research in assuming that unsophisticated traders initiate smaller trades on average (e.g., Frankel et al. [1999]). However, we find that requests for historical disclosures are positively associated with the mean dollar value of individual trades in the subsequent two hours. Thus, the evidence is inconsistent with the notion that demand for historical disclosures is driven by smaller, less sophisticated investors. 4 Given the difficulty in reading and processing complex financial disclosures such as 10-Ks and our finding that historical disclosures are used in fundamental analysis, we conjecture that sophisticated investors are more likely to use the information in historical disclosures than unsophisticated investors. Our results are consistent with this conjecture. Our findings contribute to the emerging literature that investigates whether historical or redundant news is perceived as informative to at least some market participants. For example, Tetlock [2011] and Gilbert et al. [2012] find evidence consistent with individual investors trading on stale news, in the context of media articles and macro-economic indicators. Our paper extends the literature by examining the staleness of the firm s financial disclosures, which are arguably more prominent, and certainly more regulated and complex than media articles and macro-economic indicators. Given the importance of these disclosures to market participants, examining their informativeness is important in its own right. Another point of contrast between our paper and the Gilbert et al. [2012] and Tetlock [2011] papers is that they 4 We acknowledge that trade size is a noisy proxy for investor sophistication because large investors often break up their trades to reduce their impact on prices (Chakravarty [2001]). On the other hand, large trades are not likely to be executed by small investors. 5

8 examine reactions to the release of stale information (i.e., supply), we examine consequences of actual investor requests for stale information (i.e., demand). Finally, we find evidence that old information can be informative to rational investors, which stands in contrast the evidence in prior studies. In summary, this study revisits an old question of whether disclosures are informative, with an emphasis on historical disclosures. Providing initial evidence on this question is important because the vast majority of publicly available information is old. For example, at the start of our sample period (January, 2008), EDGAR hosted over 8.2 million financial disclosures, of which less than 0.2 million (representing 2.5% of all filings) had been in the public domain for less than 30 days and could be reasonably considered new. Prior research on the informativeness of financial disclosures has almost exclusively focused on the small segment of information that is new; in contrast this paper is, to our knowledge, the first to study the informativeness of historical financial disclosures, which comprise the majority of financial disclosures. Moreover, our evidence provides insights to regulators, who are concerned about the usefulness of financial information (Paredes [2008]). The current U.S. regulatory regime requires a broad spectrum of disclosures, from complex, comprehensive period reports (10-Ks, 10-Qs) to small and frequent current disclosures (8-Ks, Form 4s). These disclosures must strike a balance between being so large that investors cannot process their information content quickly and so small that frequent updates are needed. Our findings suggest that stale disclosures (especially 10-Ks and 10-Qs) continue to be informative, especially at times of new information releases. We view this as providing some justification for the substantial costs to firms and 6

9 regulators of preparing and archiving financial disclosures. Finally, our study provides evidence for the previously untested intuition that disclosures continue to remain important beyond their initial release. 2. The Informativeness of Historical Information A fundamental step in assessing the informativeness of a disclosure is to test whether investors acquire and use the information contained in that disclosure. Researchers in accounting and finance have long been interested in understanding the role that information acquisition plays in the price discovery process, but have been hampered by the lack of an available proxy for information acquisition. One line of research uses variation in broad firm characteristics such as firm size, analyst following, or institutional ownership to proxy for variation in incentives to acquire information about a particular firm (e.g., Atiase [1985], El- Gazzar [1998], Kirk [2011]). More recent research has begun to examine whether information acquisition can be inferred using different channels of information dissemination, including conference calls (Frankel et al. [1999]), investor conferences (Bushee et al. [2011]), and the media (Soltes [2009], Bushee et al. [2010], Engelberg and Parsons [2011]). In general, prior research infers information acquisition (and thereby informativeness) by examining how disclosure events are associated with differential trading activities by investors before, during and after the event. For example, Frankel et al. [1999] find that capital market activity, such as absolute returns and trading volume, is higher for all measures during conference calls than during the control period. Bushee et al. [2011] find significant short window increases in stock returns and trading volume during managerial presentation at 7

10 conferences. Busse and Green [2002] and Engelberg et al. [2012] show that information is acquired through media television programs by documenting increases in market activity during and after a stock is mentioned on the shows. Another stream of research proposes internet search as a more direct measure of information acquisition. For example, Da et al. [2011] find that weekly Google requests for corporate stock ticker symbols are positively associated with contemporaneous weekly abnormal returns and share turnover, while Gao et al. [2011] find that daily Google requests are positively associated with greater trading volume and trade sizes. Drake et al. [2012b] find that Google search requests increase around earnings announcements and provide information that aids in the pricing of earnings news. Thus, the general evidence from studies using Google search as a proxy for information acquisition is that when contemporaneous information is acquired by investors, it leads to increases in trading activities, which suggests that these acquisition events are informative to investors. Using EDGAR search requests as a measure of information acquisition, we aim to assess the informativeness of stale information. There are several reasons that one would not expect previously filed financial disclosures to be informative to markets. Given the volumes of macro-, industry-, and firm-specific information available on a more timely basis through other sources (e.g., analysts, the business press, the internet), some argue that financial information filed with the SEC is old news upon arrival to the market (Collins et al. [1994]); Ball and Shivakumar [2008]). Furthermore, the semi-strong form of the efficient markets hypothesis predicts that the information in disclosures that have been in the public domain for a period of time should be fully reflected in stock prices. This conjecture holds true especially for 8

11 disclosures that are costless to obtain and widely disseminated, as is the case with many of the disclosures available on EDGAR. On the other hand, for the reasons that we discuss below, it is possible that investors find stale disclosures to be informative because they provide context for new information releases or because they are helpful in resolving valuation uncertainty. This discussion provides the basis for our first hypothesis, stated in the alternative form: H1: The acquisition of stale information is associated with subsequent market activity. If we find evidence for H1, the important follow-up question is why the acquisition of stale financial information is associated with market activity. One explanation for the informativeness of stale disclosures follows from the standard approach to fundamental analysis, which calls for the use of historical information in estimating firms intrinsic values. A number of studies find that historical information is predictive of future returns (e.g., Ou and Penman [1989]; Abarbanell and Bushee [1998]; and Piotroski [2000]). Newly disclosed information generally requires comparison with previously disclosed information; i.e., the older information contextualizes the new information and provides expectations against which the new information can be measured. Hence, new disclosures can contain information that cannot be fully exploited unless it is analyzed in connection with information provided in previously filed disclosures. Similarly, information can exist in old disclosures that cannot be fully exploited until subsequent disclosures are made. Moreover, market efficiency does not preclude investors from basing their expectations on information in prior disclosures. In fact, in an efficient market, investors form price expectations based all available conditioning information, including prior financial disclosures. It then follows that current disclosures can 9

12 trigger demand for stale disclosures as investors seek context. This discussion leads to our second hypothesis, stated in the alternative form: H2: The association between stale information acquisition and market activity is positively influenced by contemporaneous information releases and contemporary information acquisition. Another explanation for the informativeness of stale disclosures is that there is high uncertainty about the value of the firm given the information available to investors. When there is greater valuation uncertainty, the market can rationally under-react to new information as investors need more time to process and contextualize the information. Consistent with this argument, prior research finds evidence of lower market reactions to earnings announcements when earnings quality is lower (Francis et al. [2007]) or when earnings credibility is lower (Teoh and Wong [1993]). Valuation uncertainty also leads to more frequent revision of investor beliefs. For example, Baker and Wurgler [2006] finds that the returns of hard-to-value firms are more affected by subjective measures (in their case, investor sentiment) which they attribute to the difficulty of objectively valuing these firms. Our intuition is that, all else equal, valuation uncertainty will result in investors need to re-evaluate historic information as time passes. Taken together, these findings and our intuition suggest that when information about investment payoffs is noisy or uncertain, or when a firm is difficult to value because of a lack of objective valuation inputs, investors will continue to examine and process financial information after it has been made public. This leads to our third hypothesis, stated in the alternative: H3: The association between stale information acquisition and market activity is positively influenced by valuation uncertainty. Our final explanation for the informativeness of stale disclosures is based on investor sophistication. On the one hand, it is possible that unsophisticated investors are trading on the 10

13 information contained in stale disclosures. Researchers have posited that a subset of unsophisticated investors are likely to use stale information because they (i) are slow to acquire and process disclosures, (ii) exhibit cognitive biases, and/or (iii) do not recognize when disclosures have become stale (Gilbert et al. [2012]). For example, Hand [1990] and Huberman and Regev [2001] provide examples of presumably unsophisticated investors underreacting to an important news announcement, and subsequently overreacting when the news announcement was re-released. Similarly, Gilbert et al. [2012] and Tetlock [2011] provide evidence that investor inattention is linked to trading on stale information, which leads to subsequent mispricing. On the other hand, it takes a degree of sophistication to know how to access, read and properly interpret financial disclosures. For example, the 2011 Form 10-K recently filed by General Electric is over 270 pages long and contains detailed information on the firm s operations, risks and performance, all of which would require a substantial level of financial literacy and understanding. Moreover, the ability to perform fundamental analysis using current and past disclosures requires investment knowledge and skill that an unsophisticated investors would not likely possess. This discussion leads to our final hypothesis: H4: Stale information acquisition is associated with investor sophistication. We now turn to discussion of our unique dataset and how we use this dataset to test the hypotheses above. 11

14 3. Data and Research Design 3.1 Data and Sample The primary data used in this study capture investor information acquisition of all disclosures archived on the SEC EDGAR servers. As these data are described in detail in Drake et al. [2012a], here we provide only a brief description. The SEC maintains server logs that record every request for financial disclosures hosted on the EDGAR servers. With the assistance of the SEC s division of Risk, Strategy & Financial Innovation, we obtained the server log for the first six months of Each entry in the server log allows us to observe the partial IP address of the user 6, the date and time of the request, the Central Index Key (CIK) of the company that filed the requested form, and a link to the particular filing. Following Drake et al. [2012a], we focus our analyses on requests made by investors, rather than those made by automated webcrawlers. 7 Drake et al. [2012a] provide evidence that webcrawlers are generally deployed for database building, pulling large volumes of historical filings. Our focus is on whether investors acquire and use historical filings in their short-term trading activities. We use these data as a direct measure of financial information acquisition. The data are subject to some caveats. First, we emphasize that our data on investor requests for filings in EDGAR represent a lower bound of total information acquisition of SEC filings. Many SEC filings are freely available to investors on company investor relations websites, financial websites (e.g., Yahoo! Finance), and brokerage websites. In addition, 5 To minimize processing costs, the SEC provided us with only a limited sample period, January 1, 2008 through June 30, For privacy reasons, the final octet of the IP address was removed by the SEC prior to our receiving the data. 7 Consistent with Drake et al. [2012a], we identify investor requests as coming from any IP address that makes no more than 5 requests per minute during a given 1 minute period of time. Also consistent with Drake et al. [2012a], we remove server log entries that reference an index or that reference a request for an image to avoid double counting of the filing request. 12

15 institutional investors can directly access SEC filings from commercial data aggregators such as Bloomberg, Capital IQ, or Morningstar Document Research. 8 Institutions can also directly subscribe to the filings through the Public Dissemination Service. 9 Second, we are unsure of the identity of the user. Although we cannot validate it, we reason that the typical EDGAR user is unlikely to be a novice or expert investor, but likely somewhere in between. Third, our sample period is limited to six months of data; thus, to the extent that information acquisition via EDGAR during this period is systematically different, our results may not generalize to other periods. On the other hand, the EDGAR dataset provides several key advantages, as described in Drake et al. [2012a], over previously used proxies for information acquisition, such as firm size, analyst following and institutional ownership. First, it reflects actual information acquisition of regulatory financial filings by interested parties. That is, it reflects actions undertaken by individuals to acquire mandatory disclosures. Second, it reflects investor choice given the myriad of financial disclosures at their fingertips, the EDGAR request data allow us to identify the specific piece of information (i.e., the actual filing) that the user chooses to download. Finally, the data come directly from the primary source of regulatory disclosure the SEC. In summary, these data provide a powerful tool that we deploy to help us understand investors 8 Two additional points regarding data aggregators are relevant to our study: first, the information data aggregators transmit is often only a portion of the total information available on EDGAR; as a result, investors may still need to access some filings directly from EDGAR. Second, the information transmitted by data aggregators is often delayed relative to the SEC filing time and date. Li, Ramesh, and Shen (2011) document that it takes on average 2.3 weekdays for the Dow Jones Newswires to post an alert after an SEC report is filed while D Souza, Ramesh, and Shen (2010) document that S&P takes on average 14 weekdays to post a company s quarterly financials on Compustat. Thus, sophisticated investors may benefit from accessing the EDGAR database through the SEC website even in the presence of commercial data aggregators. 9 See the following website for more details on the Public Dissemination Service: 13

16 acquisition of stale financial information and whether that acquisition increases subsequent trading activity. The EDGAR server log provides the precise time when a disclosure is acquired by an investor. This level of detail enables us to conduct our analyses at the intra-day level, which allows us to approximate a link between the acquisition of stale information in a given hour and market activity in the subsequent hours. We obtain intra-day data on trading activity from the TAQ database available on WRDS. Due to the extreme computational demands involved in analyzing intra-day trading and intra-day EDGAR server log data, we use a random sample of 200 firms with data available from TAQ and our EDGAR dataset over the January 1, 2008 through June 30, 2008 time period. In Table 1, Panel A we present descriptive statistics (means and medians) of selected firm characteristics for our random sample of 200 firms and for all firms with available data in the intersection of the COMPUSTAT and CRSP database. We assess the representativeness of our random sample by testing for statistical differences between the means and medians of each firm characteristic, which we define in Appendix A. As Table 1, Panel A reports, we find that the Market Value of Equity, Total Assets, Book-to-Market, Return-on-Assets, Leverage, Analyst Following and Institutional Ownership for the random sample are statistically indistinguishable from the universe of firms with available data. In Table 1, Panel B we compare the distribution of the randomly selected firms across industries (Fama-French Classification 17) to that of the universe of firms with available data and again find that the industry distributions are similar. Overall, we conclude from the results in Table 1 that our random sample of 200 firms is representative of the population of available firms in Compustat and CRSP. 14

17 3.2 Variables and Empirical Models To measure the timing of investor requests, we divide each day into 24 one-hour periods. Our primary measure of information acquisition of a disclosure is the count of investor requests made for a firm s disclosures in a particular hour. To separate information acquisition for stale disclosures from that for contemporaneous disclosures, we sum all investor requests for disclosures on EDGAR separately for disclosures that are publicly available for greater than or equal to 30 days (StaleDisc) and for disclosures that are publicly available less than 30 days (RecentDisc). 10 These two variables are the explanatory variables of interest in the study. The choice of a 30-day threshold is arbitrary; we choose that threshold under the assumption that a month in the public domain is sufficient time for a disclosure to become stale. 11 Our predictions of informativeness rely on increased market activity following disclosure requests in EDGAR. Following tests of informativeness in Beaver [1968] and others, we employ two measures of market activity as dependent variables: the absolute value of the raw stock return during the hour (AbsRet) and the total dollar value of trading volume during the hour (Volume). 12 If no trades are recorded during the hour, these values are set to zero. For all of these variables, the unit of observation is measured at the firm-hour level. One potential concern is that requests for stale disclosures are endogenously determined by firm and market events that also lead to increased market activity. Hence, we include as 10 In raw form, StaleDisc and RecentDisc are highly skewed; thus, we use the natural log of one plus the raw value of these variables in our tests. Our results, however, are robust to using the raw values. 11 In section 5, we provide evidence that the results are robust to different thresholds. 12 We use the absolute value of returns rather than the signed return because we do not have signed predictions. That is, we do not evaluate the market reaction to a news release (which could contain either positive or negative news), but rather evaluate whether, on average, the number of requests for stale disclosures is associated with increased market activity. 15

18 control variables two hourly lags of the dependent variable, two hourly lags of absolute returns, and the prior day s absolute return and share turnover. To the extent that hourly and daily measures of market activity are associated with events that trigger information acquisition, these controls will reduce the effects of endogenous, omitted variables. 13 We also control for requests for recent disclosures (those filed within the most recent 30 days) in EDGAR because investor requests for stale and contemporaneous information are likely to be correlated; excluding the requests for recent disclosures would induce a correlated omitted variable problem. We also control for the decile rank of the market value of equity (RankMVE). We include this control because firm size is correlated with both market activity (returns and volume) and with information acquisition (Drake et al. [2012a]). Investors make disclosure requests on EDGAR throughout the day, but TAQ records trades only from 4am through 8pm (EST) each day the market is open; thus we construct indicator variables for times when the market is closed. Morning is set to one from midnight to 4am and to zero otherwise; Evening is set to one from 8pm to midnight and to zero otherwise; Weekend is set to one during the weekend and on holidays and to zero otherwise. In estimating the empirical models, we interact these indicator variables with both the requests for stale disclosures (StaleDisc) and the requests for recent disclosures (RecentDisc). Finally, we include 13 An alternative measure of information events is the count of the number of news articles in the Wall Street Journal, the New York Times, USA Today, and the Washington Post that mention the firm on the prior day. In untabulated tests, we include such a control and find that it has virtually no effect on the results reported in our tables. We thank Eugene Soltes for providing the media count data used in Soltes [2009]. 16

19 hour-of-day fixed effects and we cluster the standard errors by day to control for cross-sectional residual correlation. 14 We estimate the following general model for all firms i (subscripts omitted) and hours t: Market Activityt = f (StaleDisct-1, StaleDisct-2, RecentDisct-1, RecentDisct-2, Market Activityt-1, Market Activityt-2, AbsRett-1, AbsRett-2, RankMVE, PriorDay_AbsRet, PriorDay_Turnover, Controls for Market Closed, Hour Fixed Effects), (1) where Market Activity Controls for Market Closed = one of two dependent variables: AbsRet or Volume as defined above; = set of control variables which includes the Morning, Evening, and Weekend indicator variables defined above, as well as the interaction of each of these indicator variables with StaleDisct-1, StaleDisct-2, RecentDisct-1, and RecentDisct-2; all other variables are defined above. H1 predicts that the acquisition of stale information should be associated with trading activity. The coefficients on StaleDisct-1 and StaleDisct-2 are the test variables for H1 in estimations of model (1). A significantly positive coefficient on those variables is consistent with investor acquisition of stale disclosures leading to significant capital market activity, which is consistent with H1. All subsequent analyses are tests for potential reasons for H1. H2 posits that the relation between stale information acquisition and market activity is positively influenced by contemporaneous information releases and acquisition. We test H2 using two information release events: earnings announcements and 8-K filings. We investigate earnings 14 We have also run the models excluding relatively time-invariant variables (such as firm size) and instead including firm fixed effects. All results are insensitive to this design choice. 17

20 announcements because they are one of the most important corporate disclosure events and they allow us to observe the precise time that the report was made public. From IBES, we obtain the timestamp of any quarterly earnings announcement issued by our sample firms during the sample period. We employ a very short window (+/- 5 hours) around the earnings announcement to increase the likelihood that the investor requests from EDGAR are motivated by the earnings announcement. The indicator variable (EarnAnnHour(-5,5)) is set equal to one for the 11-hour period centered on the earnings announcement hour and to zero otherwise. 15 When a firm undergoes a significant current event, such as an acquisition, executive hiring or firing, changes in auditor, or bankruptcy, it is required to file a Form 8-K with the SEC. 16 We employ the filing date of an 8-K as a proxy for the many important events that trigger demand for firm-specific information. 17 We obtain 8-K filing dates from the Master Index File available in EDGAR and construct an indicator variable (Form8-K) set equal to one on the filing date and to zero otherwise. 18 We then enter these variables separately into model (1) and interact them with the EDGAR request variables as a test of H2. We also interact RankMVE 15 Some researchers have noted that IBES earnings announcement timestamps are unreliable (Dong et al. [2011]). To improve reliability, we validate the earnings announcement times using three different sources, following dehaan et al. [2012]. 16 The SEC provides a list of events that trigger an 8-K filing at 17 Earnings announcements also require an 8-K filing; as a result, there is some overlap between EarnAnnHour(-5,5) and Form8-K. However, Lerman and Livnat [2010] report that earnings announcements make up only 27 percent of all 8-K filings, while other large corporate events such as acquisitions and executive turnovers also make up a large proportion of the totals. Examining earnings announcements separately also allows us to tighten the event window because the precise time of the earnings announcement is available in IBES. We return to this issue in Section 5 when we discuss sensitivity tests of our empirical results. 18 The EDGAR Master Index File is available at ftp://ftp.sec.gov/edgar/full-index. We thank Andy Leone for providing the PERL program used to download the index (see 18

21 with both EDGAR request variables in the model to ensure that the earnings announcement or 8-K indicator variable is not contaminated with a size effect. 19 The model is as follows: Market Activityt = f (StaleDisct-1, StaleDisct-2, RecentDisct-1, RecentDisct-2, CurrentEvent, StaleDisct-1 x CurrentEvent, StaleDisct-2 x CurrentEvent, RecentDisct-1 x CurrentEvent, RecentDisct-2 x CurrentEvent, RankMVE, StaleDisct-1 x RankMVE, StaleDisct-2 x RankMVE, RecentDisct-1 x RankMVE, RecentDisct-2 x RankMVE, Market Activityt-1, Market Activityt-2, AbsRett-1, AbsRett-2, PriorDay_AbsRet, PriorDay_Turnover, Controls for Market Closed, Hour Fixed Effects), (2) where CurrentEvent = one of two variables: EarnAnnHour(-5,5) or Form8-K as defined above; all other variables are defined above. In Model (2), the main effect CurrentEvent captures the general increase in market activity associated with earnings announcements or 8-K filings. Our focus is on the interaction of the current event and stale information acquisition. A positive coefficient on StaleDisc x EarnAnnHour(-5,5) or on StaleDisc x Form8-K is consistent with H2 and indicates that the association between requests for stale disclosures and market activity is even stronger when there is a current information release. In a final test of H2, we examine whether the association between investor acquisition of stale disclosures and subsequent market activity is influenced by investor EDGAR requests for current financial disclosures. We test this conjecture by interacting the two EDGAR request 19 In all models that include interactions, we also interact RankMVE with the EDGAR variables (RecentDisc and StaleDisc). As noted in Drake et al. [2012a], the number of EDGAR requests is strongly linked to firm size; in addition, our market variables (especially trading volume) are also strongly linked to firm size. Hence, the choice to interact the primary variables with firm size is important to mitigate the joint effect of firm size on both the dependent and independent variables. Excluding the interaction from the models would result in an omitted correlated variable problem that would bias our main coefficients and thus influence the interpretation of our results. 19

22 variables for each one hour time period (StaleDisc x RecentDisc) and by entering these interactions into model (1). Given that investors disclosure requests are highly positively correlated with firm size (Drake et al. [2012a]), we also interact RankMVE with both EDGAR request variables in the model as well. Our third model is as follows: Market Activityt = f (StaleDisct-1, StaleDisct-2, RecentDisct-1, RecentDisct-2, StaleDisct-1 x RecentDisct-1, StaleDisct-2 x RecentDisct-2, RankMVE, StaleDisct-1 x RankMVE, StaleDisct-2 x RankMVE, RecentDisct-1 x RankMVE, RecentDisct-2 x RankMVE, Market Activityt-1, Market Activityt-2, AbsRett-1, AbsRett-2, PriorDay_AbsRet, PriorDay_Turnover, Controls for Market Closed, Hour Fixed Effects) (3) where all variables are defined above. A positive coefficient on the StaleDisc x RecentDisc provides further support for H2 and indicates that the association between requests for stale disclosures and market activity is stronger when there are more requests for contemporaneous financial information. Our third hypothesis, H3, examines whether valuation uncertainty plays a role in explaining the association between EDGAR requests for stale financial disclosures and subsequent market activity. We employ three proxies for valuation uncertainty. The first proxy is based on the level of intangible assets, Intang, under the assumption that firms with more intangibles assets are harder to value. We calculate Intang as the proportion of assets not related to capital assets following Kumar [2009]. The second proxy for valuation uncertainty relates to earnings quality and is based on the extent to which current accruals map into cash flows (e.g., Dechow and Dichev [2002], Francis et al. [2005], Francis et al. [2007]). We select an earnings-based measure of information uncertainty because earnings relates directly to equityinvestment payoffs. We estimate information uncertainty, EarnQuality, using the residuals from the following model for all firms i in years T: 20

23 CurrAccrT = f (CFOT-1, CFOT, CFOT+1, ΔREVT, PPET), (4) where CurrAccr CFO ΔREV PPE = total current accruals; = cash flow from operations; = change in revenues; and = gross value of property, plant, and equipment. We provide more detailed variable definitions in Appendix A. We follow Francis et al. [2007] in estimating model (4) for each industry (Fama-French 48 classifications) with at least 20 observations and for each year T. We capture the residual for each sample firm and estimate the standard deviation of the residuals (EarnQuality) over the past 5 years (years T-4 through T). Larger variation in the residuals suggests that cash flows map poorly into accruals and thus indicates greater information uncertainty. The final proxy is based on dispersion in analyst forecasts of earnings, which are an important component of value. We assume that greater disagreement among analysts about future performance is indicative of greater valuation uncertainty. We define Dispersion as the standard deviation of analyst annual earnings forecasts at the beginning of the year, scaled by stock price. We test H3 by interacting the EDGAR request variables for each one hour time period with our measures of valuation uncertainty (StaleDisc x ValUncertain and RecentDisc x ValUncertain) and entering these interactions into model (1). Given that valuation uncertainty is likely negatively correlated with firm size we also interact RankMVE with the EDGAR requests variables (StaleDisc x RankMVE and RecentDisc x RankMVE). The model is as follows: Market Activityt = f (StaleDisct-1, StaleDisct-2, RecentDisct-1, RecentDisct-2, ValUncertain, StaleDisct-1 x ValUncertain, StaleDisct-2 x ValUncertain, RecentDisct-1 x ValUncertain, RecentDisct-2 x ValUncertain, RankMVE, StaleDisct-1 x RankMVE, StaleDisct-2 x RankMVE, RecentDisct-1 x RankMVE, RecentDisct-2 x RankMVE, Market Activityt-1, 21

24 Market Activityt-2, AbsRett-1, AbsRett-2, PriorDay_AbsRet, PriorDay_Turnover, Controls for Market Closed, Hour Fixed Effects), (5) where ValUncertain = one of three variables: Intang, EarnQuality, or Dispersion as defined above; all other variables are defined above. A significantly positive coefficient on the StaleDisc x ValUncertain is consistent with H3 and indicates that the association between requests for stale disclosures and market activity is stronger when there is greater information uncertainty in the firm. Our final hypothesis, H4, posits that the relation between stale disclosure acquisition and market activity is related to sophisticated trading. Following prior research (e.g., Frankel et al. [1999]), we use the average trade size during the hour (TradeSize) as a proxy for sophisticated trading under the assumption that larger trades are unlikely to be initiated by unsophisticated investors. We enter TradeSize as the dependent variable in model (1) as follows: TradeSizet = f (StaleDisct-1, StaleDisct-2, RecentDisct-1, RecentDisct-2, TradeSizet-1, TradeSizet-2, AbsRett- 1, AbsRett-2, RankMVE, PriorDay_AbsRet, PriorDay_Turnover, Controls for Market Closed, Hour Fixed Effects), (6) where all variables are defined above. A negative (positive) coefficient on StaleDisct-1 and/or StaleDisct-2 is consistent stale information acquisition being associated with less (more) sophisticated trading. 4. Findings 4.1 Descriptive Statistics and Correlations 22

25 In Table 2, we provide descriptive statistics for disclosure requests by hour of the day. In Panel A of Table 2 we present the mean, standard deviation, and maximum number of stale and recent disclosure requests for any form in EDGAR during the hour, aggregated across all 200 firms. Broadly, we find that both stale and recent disclosures are in highest demand during normal trading hours, and in particular, in the hours just before the markets close, which is consistent with our data picking up the information acquisition activity of human investors (rather than robots). We also find that across all time periods the mean number of requests for stale disclosures is greater than the mean number of requests for recent disclosures. This finding should certainly be viewed in light of the fact that there is a much greater supply of stale financial disclosures available on EDGAR; however, it does underscore the value of historic financial information and the benefits of storing these disclosures in an archival database such as EDGAR. Untabulated findings show that the median number of days between the filing date and the request date is 370 days for the stale disclosure sample and 1 day for the recent disclosure sample. In Panel B of Table 2 we present descriptive statistics for the three dependent variables by hour of the day. Consistent with Kelly and Tetlock [2012], we find that AbsRet and Volume are highest during the middle hours of the day (from 10:00 to 15:00 EST). In contrast, TradeSize is generally greatest during the opening and closing hours of the market (around 9:00 EST and 16:00 EST). As described above, in our empirical models we include hour of day fixed effects to control for these observed differences in market activity throughout the hours of the day. In Table 3 we present pairwise Spearman correlations for all of the variables used in the regression models. We find that StaleDisct-1 is positively associated with RecentDisct-1 with a 23

26 correlation coefficient of 0.32, providing univariate evidence that disclosures filed at different times are requested together. We also find that StaleDisc is positively associated with the market activity variables AbsRet and Volume, and with prior day absolute returns and turnover. We do not discuss the correlations between all of variables, but note that the pairwise correlations are all significantly greater than zero at the 10 percent level. We now turn to the test results of the formal hypotheses. 4.2 Tests of H1 Figure 1 shows evidence that investors continue to download 10-Ks for a number of weeks after they are filed with the SEC. A visual inspection of the figure shows that investor requests for a 10-K are highest right after it is filed, but continue at levels above a baseline amount for about 18 months after it is filed. After that point, investor downloads of disclosures continue at roughly the same (lower) rate for years after the filing. To the extent that investor downloads of stale disclosures are evidence of their informativeness, we interpret the findings in this figure as prima facie evidence of the informativeness of stale disclosures. We next turn to tests of whether the acquisition of stale disclosures is associated with subsequent market activity. In Table 4 we present the estimation results for model (1). In Panel A of Table 4, we estimate the regressions using stale and recent disclosure requests for all disclosures in EDGAR; in Panel B of Table 4, we focus on requests for periodic accounting disclosures (i.e., requests for 10-Ks and 10-Qs only). The purpose of Panel B is to examine whether stale accounting disclosures are informative to equity markets. For parsimony, we do 24