Earnings Announcements

Transcription

1 Google Search Activy and the Market Response to Earnings Announcements Mary E. Barth Graduate School of Business Stanford Universy Greg Clinch The Universy of Melbourne Matthew Pinnuck The Universy of Melbourne September 2010 Note to UQ seminar participants: This manuscript is in a proto paper stage and is incomplete. It reports some very early and preliminary results only. The manuscript is short on economic stories and does not yet attempt to link to the relevant lerature. I will endeavour to fill in more background and some of our thinking in the seminar. Please do not ce or quote as the reported results are subject to the preliminary nature of our analyses, and may not survive the test of time. The use of the Google trademark does imply approval of the contents by Google Corp. We thank Mark Szabo and Sorabh Tomar for excellent research assistance. Please address all correspondence to Greg Clinch, clinchg@unimelb.edu.au.

2 1. Introduction In this paper we investigate the association between Google search activy and the market response to earnings announcements. We explo data relating to information searches via Google available from Google Insights for Search ( to investigate the extent to which Google search activy is associated wh stock returns and trading volume (i) at the time of an earnings announcement, and (ii) subsequent to an earnings announcement. We are motivated by the possibily that Google search activy reflects investors seeking information to help predict and/or understand earnings news released by companies. To the extent this is the case and also that the acquired information affects investors trading activy, we expect Google search activy to be associated wh the market response to earnings news. Our research investigates this possibily. Our preliminary results suggest that increased Google search activy occurs in earnings announcement weeks. Search activy in the announcement week appears to be associated wh the magnude of the returns-earnings relation only to the extent that is also associated wh trading volume. Finally, there is some evidence that post-earnings announcement returns over a thirteen week period are more strongly associated wh unexpected earnings when Google search activy is greater in the post announcement period. The remainder of the paper is organized as follows. Section 2 provides some background discussion relating to investor information search and the market response to information release. In section 3 we briefly describe our sample and data, including the Google search activy measures we employ. Section 4 reports inial descriptive results relating to Google search activy around earnings releases, while section 5 reports inial results relating to the stock return and trading volume response to earnings. Section 6 provides a brief conclusion. 1

3 2. Background discussion Our research is based on the premise that online search activy potentially reflects, in part, information search by investors. If this is the case then several questions naturally arise. First, to what extent does such search activy affect investor decisions in a way that is reflected in share market activy, in particular share prices? It is possible that online search activy reflects largely curiosy on the part of interested observers who have ltle direct interest in the share market. Or could represent information search by retail investors, and more sophisticated investors, whose investment decisions may or may not affect share prices. Our research inially investigates the extent to which information search through Google is related to share market activy. We focus on information search surrounding earnings announcements because such announcements are common, represent a major event for many companies, and are the subject of a substantial amount of previous research. 1 A second question that arises relating to online search activy is whether reflects, largely, the actions of relatively unsophisticated investors. If does, is possible that search activy could be associated wh circumstances where share prices appear to depart from efficient pricing. Specifically, in the case of earnings announcements, is possible that search activy is associated wh the existence and strength of post earnings announcement returns behaviour. Our research is aimed, in part, at investigating this possibily. 3. Sample and Data Our sample data covers the period To construct the sample we used the CRSP file as at December 31, 2003 to obtain an inial list of 6,912 firms. We employed a web crawling 1 We are aware of only one other paper at present that employs Google search data in a capal markets context: Da, Engelberg, and Gao (2009), In Search of Attention (unpublished manuscript). That paper employs Google search activy relating to IPOs. 2

4 program using ticker code and company name information from CRSP to obtain available weekly Google search activy data from the beginning of 2004 (the first date available) through March This was matched wh stock market data from CRSP for the period to December 2008 (the most recent CRSP file available). Because weekly Google search activy is measured from Sunday to Saturday, we constructed corresponding calendar week return and volume measures to match wh the Google data. 2 We obtained market capalization data from CRSP, quarterly earnings announcement dates, earnings and book value data from Compustat, and analyst forecast data from I/B/E/S. Our final sample consists of 5,618 firms, wh a total of 1,214,341 firm-weeks, and 95,157 quarterly earnings announcement dates. Table 1 provides some descriptive statistics for several firm characteristics pertaining to our sample. To obtain weekly search activy from Google Insights requires the specification of specific search strings that Google users might employ in their searches. We obtained search activy data using several alternative collections of search strings: (i) a company s ticker symbol, (ii) a company s name, (iii) eher a ticker symbol or company name, (iv) a company s ticker as well as one of the word combinations earnings, net income, eps, (v) eher a ticker symbol or company name as well as one of the earnings related word strings. We employ the ticker-based search string as our primary focus based on the expectation that investor-related search activy is more likely to use ticker symbols than company names. Also, company name-based search strings are likely to generate more spurious search activy unrelated to investor information search (e.g., Apple ) than are ticker-based search strings. 3 Google Insights also allows several options in determining the measure of weekly search activy, including the geographic area in 2 We defined a calendar week for return/volume purposes to be the week to the end of trade on each Friday. As a result, there is a slight, and unavoidable, mismatch between the time period covered by a google week and our return/volume weeks. 3 We plan to investigate the robustness of our results to using the alternative search strings. Search strings that include the earnings-related word strings generated only approximately 400 firms wh non-zero search activy, and thus would greatly reduce the available sample size and likely be concentrated on only large, visible companies. 3

5 which searches are iniated (e.g., Worldwide, or specific countries) and several broad categories of searches (e.g., all categories, finance and insurance, etc.). Our main results are based on Google search activy whin the Uned States under all categories. 4 Two addional features of the data provided by Google Insights affect our research. First, Google does not provide actual search activy. Instead reports weekly search activy for the supplied search string relative to the maximum weekly search activy for the same search string over the requested time period (in our case January 2004 through early April 2010) where the maximum search activy week is set at 100. We refer to this measure as unscaled search activy. As a result, for our ticker-based search string the unscaled search activy measure ranges between zero and 100 and represents weekly search activy relative to the maximum week of search activy over for each specific company. It does not allow for comparisons across firms in search activy. However, Google Insights also allows multiple (up to six) search strings to be specified concurrently, in which case the resulting search activy is measured relative to the maximum week across the multiple search strings. As a result, is possible to construct search a search activy measure that can be compared across companies. We achieve this by specifying for each company two concurrent search strings: s own ticker symbol and msft (the ticker symbol for Microsoft Corporation). The resulting search activy is scaled relative to the maximum weekly search activy across the two companies. When the maximum week across the two companies is not msft we rescale the measure by the ratio of the 4 We also obtained search activy data for other combinations of geographic area and categories: Worldwide/all categories, Uned States/insurance and finance, and Worldwide/insurance and finance. We plan to check the robustness of our reported results wh respect to these alternatives. 4

6 maximum msft week to the maximum non- msft week. This results in all weekly search activy measures being measured relative to the maximum msft search activy week. 5 The second feature of Google search activy measures that affects our analysis is that Google does not report search activy below a certain, undisclosed minimum. Instead values which fall below the minimum are disclosed as zero. As a result, values of zero should be interpreted as low rather than representing zero search activy. Table 2 provides some descriptive statistics regarding the Google search metric we employ. The full sample of available Google search activy data covers 5,660 firms and 1,247,874 firm-weeks. Table indicates a mean (median) unscaled search activy of 27.7 (24.0) across all firm-weeks. That is, on average search activy in a single week is 27.7% of the maximum weekly search activy over the period for each firm. Table 2 also indicates that there are a large number of weeks wh zero reported unscaled search activy: 515,363 firmweeks, representing approximately 41% of the sample. For rescaled search activy, the mean (188.3) is higher but the median (2.0) is lower than the corresponding mean and median for unscaled search activy. For the majory of firms/tickers, rescaling causes the search activy to decrease as their search activy is less than that for msft. But for a small number of tickers rescaling results in an increase in the search activy measure, and is these that cause the mean to increase. In addion, there are slightly more firm-weeks (approximately 7,000) wh zero 5 This procedure has the potential to magnify noise in the search activy measure for firms where the ticker symbol might represent a search string unrelated to investor information search for the specific company. For example, the firm Flag Financial Corporation has a ticker symbol of flag. Employing this as the search string will cause Google Insights to tabulate search activy based on any search strings that include the word flag. Many of these will be unrelated to Flag Financial Corporation. Moreover, a generic search string such as flag is likely to generate a greater maximum weekly search activy measure than for msft. This is, in fact the case. The flag search string generates a maximum weekly search activy measure of 100 in the week ending July 23, When flag and msft are specified as concurrent search strings, the maximum msft search activy is 1 in the week ending May 14, Thus the rescaled google activy measure for flag is for the week ending July 23, Rescaling generates the value of for 88 firm-weeks in our sample. These and other large values of rescaled search activy are likely to represent noise in our analysis. We migate this by employing the log of search activy in our analysis. Our future work will investigate other avenues for migation of noise in the search activy measure such as winsorising the data, omting extreme values, and employing influential observation diagnostics. 5

7 weekly rescaled search activy: including msft as a concurrent search string causes some firmweeks search activy to fall below the (undisclosed) minimum activy used by Google to report non-zero search activy. We employ two approaches to remove potential noise in the google search activy measure. First, we remove firms wh zero search activy across all weeks. This reduces the sample by 1,984 firms and 405,824 firm-weeks. Table 2 indicates that removing these observations increases the mean and median search activy measures, as expected. It also indicates that zero search activy weeks represent only 13% and 14% respectively of all available firm-weeks after the removal of firms wh no search activy in any week. The second approach we employ to migate noise in the search activy measure is to remove noisy tickers from the sample. We define noisy tickers as those for which a Google search based on the ticker symbol yielded search results such that neher the company s name or the word finance appeared in the first five generated search results. 6 Table 2 indicates that removing noisy tickers further reduces the sample by 2,326 firms and 499,727 firm-weeks, and causes the mean and median of both unscaled and rescaled search activy to decrease. 4. Google search activy around earnings release dates Figure 1 plots average unscaled weekly Google search activy in event time around quarterly earnings announcement weeks, both for all quarters and for 4 th quarters only. There is a clear spike in the earnings announcement week. There is also an apparent gradual increase in search activy as an earnings announcement approaches as well as a tapering off of search 6 Using this approach assumes that the non-appearance of eher the company s name or the word finance indicates an increased likelihood that the search string employed is picking up searches unrelated to the firm. 6

8 activy over the weeks subsequent to the earnings release week. The pattern is consistent wh an increase in Google search activy associated wh the release of earnings news. We investigate the association between search activy and earnings releases via a series of regressions. Our inial baseline regression is: LOGGOOGLE = a + bannweek + e (1) where LOGGOOGLE is the log of rescaled Google search activy for firm i in week t, and ANNWEEK is a dummy set equal to one for quarterly earnings announcement weeks. Table 3, panel A, reports summary statistics from estimating (1) for the full sample as well as the reduced sample of firms wh noisy tickers and zero search activy firms removed (the non-noisy sample). Consistent wh the plots contained in figure 1, the estimated coefficient on the announcement week dummy is posive and statistically significant. 7 We expand (1) by including several firm characteristics as addional explanatory variables: LOGGOOGLE = a + bannweek + c LOGMKTVAL 1 + c LOGANALYST + c BOOKMKT + e 2 3 (2) where LOGMKTVAL is the log of market value, following, and LOGANALYST is the log of (one plus) analyst BOOKMKT is the book-to-market ratio, all measured at the end of the previous quarter for firm i and week t. Table 3, panel A indicates that for the full sample, and BOOKMKT are significantly posively associated wh search activy while LOGMKTVAL 7 All test statistics for regressions wh the LHS based on Google search activy employ standard errors calculated based on clustering at the firm level. 7

9 LOGANALYST is significantly negatively associated wh search activy. When the non-noisy sample is used, however, LOGANALYST loses significance. In each case, the announcement week dummy remains significantly posively associated wh search activy. We further expand (2) by including absolute abnormal return (weekly return less the equally weighted market return), ABSABNRET, and VOLUME (weekly trading volume divided by shares outstanding) as addional RHS variables. Both of these variables likely reflect news arrival and will potentially be associated wh information search on Google. Table 3, panel A indicates that for the full sample both variables are significantly posively associated wh Google search activy. However, for the non-noisy sample only absolute returns are significantly associated wh search activy. For both samples inclusion of weekly absolute abnormal returns and trading volume in the regression cause the coefficient on the announcement week dummy to become negative and significant. To investigate whether earnings announcement week search activy is associated firm characteristics and the magnude of earnings news we interact the announcement week dummy wh LOGMKTVAL, LOGANALYST, BOOKMKT, and ABSSUEDEC (the decile ranking of absolute unexpected earnings deflated by prior quarter ending stock price, using earnings from the same quarter in the previous year as expected earnings). Panel A of Table 3 indicates that Google search activy in the earnings announcement week is posively and significantly associated wh firm size and the magnude of earnings news and negatively associated wh the book-to-market ratio. It is not significantly associated wh analyst following. The results reported in table 3, panel A indicate that Google search activy is associated both wh various firm characteristics as well as stock market activy measures such as the magnude of abnormal weekly returns and trading volume. As always, the likelihood remains 8

10 that firm and market characteristics important for explaining variation in Google search activy across firms and time have been omted. One approach to further controlling for across firm differences in search activy is to employ a measure that controls for recent normal levels of search activy for each company. Panel B of table 3 reports results from estimating the same regressions reported in panel A but wh ABNGOOGLE as the LHS variable, where ABNGOOGLE is abnormal (logged) rescaled search activy, measured as average LOGGOOGLE over the previous seven weeks for firm i and week t. LOGGOOGLE less A first observation relating to the regression results where abnormal Google search activy is employed as the LHS variable is that the R 2 statistics are lower than in panel A of table 3. As might be expected, removing a source of cross-sectional variation by differencing whin firms reduces the explanatory power linked to firm characteristics. This is supported by the reduced significance levels for each of the firm characteristics: size ( LOGMKTVAL ), analyst following ( LOGANALYST ), and book-to-market ( BOOKMKT ). In contrast, the announcement week effect remains statistically significant and posive across all three regression specifications where there are no interaction effects included to explain. This suggests that the negative or weak announcement week effect observed when the level of logged rescaled search activy is the LHS variable for some specifications is likely explained by firm-specific differences in search activy across firms rather than differences over time whin firms. Overall, the results reported in table 3 indicate that there is a statistically significant increase in (abnormal) Google search activy in quarterly earnings announcement weeks, and that this increase is posively associated wh firm size and analyst following, but negatively associated wh the book to market ratio. Interestingly, there is ltle reliable evidence that the increase in search activy is associated wh the magnude of earnings news as captured by 9

11 decile rankings based on seasonal quarterly random walk SUEs. Table 4 provides addional evidence relating to this and reports results from a regression based on quarterly announcement weeks only, and for the reduced non-noisy sample. In panel A, where Google search activy is measured in levels there is an apparent posive association between search activy and earnings surprise measured relative to a seasonal random walk model or relative to consensus analyst forecasts. However, when abnormal Google search activy is used as the LHS variable (panel B), these associations lose their significance. Again, there is no consistently persuasive evidence of an association between Google search activy in an earnings announcement week and the magnude of earnings surprise released. 5. Google search activy and the association between returns and earnings 5.1 Announcement week returns Table 5 reports results from regressions of abnormal returns and earnings in quarterly earnings announcement weeks. The regressions include standard risk factors (beta, size, and book-to-market) to control for potential risk differences across firms and time for firms in our sample. A loss dummy is also included consistent wh prior research. Our focus is on the coefficient on the earnings variable (based on deciles of unexpected earnings) and how this coefficient varies across firms/time in association wh Google search activy. Panel A of table 5 reports the results from several regressions. The first column is a standard (short window) ERC regression. As expected the estimated slope coefficient on the unexpected earnings variable is posive and strongly statistically significant (t = 11.73). In the second column we include two unexpected earnings variables one based on a seasonal quarterly 10

12 random walk expectations model, the other based on I/B/E/S consensus forecasts. 8 Both variables are posively and statistically significantly associated wh announcement week abnormal returns, as might be expected. Though the analyst forecast based measure exhibs a substantially larger slope coefficient. Our main interest is in the regressions that include interaction variables wh the unexpected earnings variables. The results indicate that Google search activy is associated wh a larger slope coefficient between returns and earnings, but only if trading volume is excluded as a potential interaction variable. That is, only to the extent that Google search activy is also associated wh increased trading volume is associated wh a greater magnude association between returns and earnings. Panel B of table 5 includes an addional Google search activy variable which represents search activy over the five weeks prior to a quarterly earnings announcement week. Although the pre-announcement search activy interaction variable has a negative coefficient in each regression, is not statistically significant. Thus there is no statistically reliable evidence that higher Google search activy prior to an earnings release results in a weaker announcement week returns-earnings association. 8 Including two unexpected earnings variables in the one regression potentially increases concerns wh multicollineary, wh a possible loss of statistical power. 11

13 5.2 Post-announcement period returns Table 6 reports results from regressions of post-announcement abnormal returns on unexpected earnings. In panel A abnormal returns are measured over the week after a quarterly earnings announcement week, while in panel B the measurement window is thirteen weeks. Consistent wh prior research relating to the post-earnings announcement drift, the baseline regressions (wh no interaction variables) exhib a posive and statistically significant association between post announcement abnormal returns and unexpected earnings. In panel A of table 6 no interaction variables exhib statistically significant coefficients. However, in panel B there is some evidence that long window post announcement returns are more strongly associated wh unexpected earnings when Google search activy in the post announcement period is higher. However, this result is not unambiguously strong. It is somewhat weakened when both unexpected earnings measures are included in the one regression, perhaps due to increased multicollineary and loss of statistical power. 6. Conclusion Our inial results indicate that Google search activy is increased around the release of quarterly earnings numbers. However there is no consistent persuasive evidence that the increased search activy is associated wh the magnude of the earnings surprise reported. Our results also suggest that the magnude of the returns-earnings relation in announcement weeks is associated wh announcement week Google search activy only to the extent that search activy is also associated wh trading volume. There is no persuasive evidence that the returns-earnings relation is associated wh search activy in the pre-announcement period. Finally, there is some evidence that long window abnormal returns subsequent to an earnings release week are more strongly associated wh earnings when post-announcement Google search activy is higher. 12

14 References [to be completed] 13

15 Table 1 Summary statistics for weekly sample firm characteristics. Mean Median Std Dev Min 25 th percenti le 75 th percenti le Max Market capalization Book-to-market Analyst following Absolute SUE analyst Absolute SUE seasonal random walk Absolute weekly abnormal return Weekly trading volume SUE is standardized reported earnings less forecast earnings (using eher the most recent consensus I/B/E/S forecast or the corresponding reported earnings from the prior year) deflated by stock price at the beginning of the relevant quarter. Market capalization, Book-to-market, Analyst following, and Absolute SUE are all measured on a quarterly basis. Weekly abnormal return is cumulated daily abnormal return using the CRSP equally weighted index as the benchmark return. Weekly trading volume is measured relative to shares outstanding. All variables are winsorised at the 1 st and 99 th percentiles. 14

16 Table 2 Summary statistics for weekly Google search activy. All Firms (5660 firms, firm-weeks) Unscaled Weekly Google Search Activy Scaled Weekly Google Search Activy Firms wh at least one week of non-zero activy (3676 firms, firmweeks) Unscaled Weekly Google Search Activy Scaled Weekly Google Search Activy Firms wh non-noisy tickers and at least one week of non-zero activy (1350 firms, firmweeks) Unscaled Weekly Google Search Activy Scaled Weekly Google Search Activy Mean Median Std Dev Min Max No. = Unscaled Google search activy is measured on a scale relative to the maximum weekly search activy for each firm over the January April 2010 time period. Scaled Google search activy is measured relative to the maximum weekly search activy for the ticker MSFT (Microsoft Corporation) over the January 2004-April 2010 time period, where the maximum weekly MSFT search activy is set to 100. For rescaling purposes, if Google reports maximum weekly MSFT search activy equal to zero we set equal to 1. Thus the maximum rescaled search activy for any ticker is Non-noisy tickers are tickers where we om any ticker/firm where eher the company s name or the word finance did not appear in the first five search results from a Google search of the ticker. 15

17 Table 3 Summary statistics for regressions explaining weekly Google search activy. Panel A: LHS variable is log of rescaled weekly Google search activy Coeff t Coeff t Coeff t Coeff t All firms (5618 firms, observations) Intercept ANNWEEK LOGMKTVAL LOGANALYST BOOKMKT ABSABNRET VOLUME ANNWEEK*LOGMKTVAL ANNWEEK*LOGANALYST ANNWEEK*BOOKMKT ANNWEEK*ABSSUEDEC R Firms wh zero Google activy and noisy tickers removed (1348 firms, observations) Intercept ANNWEEK LOGMKTVAL LOGANALYST BOOKMKT ABSABNRET VOLUME ANNWEEK*LOGMKTVAL ANNWEEK*LOGANALYST ANNWEEK*BOOKMKT ANNWEEK*ABSSUEDEC R

18 Table 3 (continued) Summary statistics for regressions explaining weekly Google search activy. Panel B: LHS variable is abnormal (log of) rescaled weekly Google search activy Coeff t Coeff t Coeff t Coeff t All firms (5549 firms, observations) Intercept ANNWEEK LOGMKTVAL LOGANALYST BOOKMKT ABSABNRET VOLUME ANNWEEK*LOGMKTVAL ANNWEEK*LOGANALYST ANNWEEK*BOOKMKT ANNWEEK*ABSSUEDEC R Firms wh zero Google activy and noisy tickers removed (1343 firms, observations) Intercept ANNWEEK LOGMKTVAL LOGANALYST BOOKMKT ABSABNRET VOLUME ANNWEEK*LOGMKTVAL ANNWEEK*LOGANALYST ANNWEEK*BOOKMKT ANNWEEK*ABSSUEDEC R

19 ANNWEEK is a dummy set equal to one for quarterly earnings announcement weeks. LOGMKTVAL is the log of market value, LOGANALYST is the log of analyst following, and BOOKMKT is the book-to-market ratio all measured at the end of the previous quarter. ABSABNRET is absolute weekly abnormal return measured relative to the equally weighted market index. VOLUME is weekly trading volume divided by shares outstanding. ABSSUEDEC is the decile ranking of absolute unexpected earnings deflated by prior quarter ending stock price, using earnings from the same quarter in the previous year as expected earnings. Standard errors are based on based on clustering at the firm level. 18

20 Table 4 Summary statistics for regressions explaining weekly Google search activy in earnings announcement weeks. Panel A: LHS variable is log of rescaled weekly Google search activy Coeff t Firms wh zero Google activy and noisy tickers removed (1169 firms, observations) Intercept LOGMKTVAL LOGANALYST BOOKMKT ABSABNRET VOLUME DUMLOSS ABSSUEDECRW ABSSUEDECANL R Panel B: LHS variable is abnormal (log of) rescaled weekly Google search activy Coeff t Firms wh zero Google activy and noisy tickers removed (1169 firms, observations) Intercept LOGMKTVAL LOGANALYST BOOKMKT ABSABNRET VOLUME DUMLOSS ABSSUEDECRW ABSSUEDECANL R LOGMKTVAL is the log of market value, LOGANALYST is the log of analyst following, and BOOKMKT is the book-to-market ratio all measured at the end of the previous quarter. ABSABNRET is absolute weekly abnormal return measured relative to the equally weighted market index. VOLUME is weekly trading volume divided by shares outstanding. DUMLOSS is a dummy set to one if the firm reports a loss. ABSSUEDECRW is the decile ranking of absolute unexpected earnings deflated by prior quarter ending stock price, using earnings from the same quarter in the previous year as expected earnings. ABSSUEDECANL is the decile ranking of absolute unexpected earnings deflated by prior quarter ending stock price, using the most recent I/B/E/S consensus earnings forecast as expected earnings. Standard errors are based on based on clustering at the firm level. 19

21 Table 5 Summary statistics for announcement week returns-earnings regressions. Panel A: The association between announcement abnormal Google search activy and the returns-earnings relation. Coeff t Coeff t Coeff t Coeff t Coeff t Coeff t Firms wh zero Google activy and noisy tickers removed (1173 firms, observations) Intercept BETA LOGMKTVAL BOOKMKT DUMLOSS SUEDECRW SUEDECRW*BETA SUEDECRW*LOGMKTVAL SUEDECRW*BOOKMKT SUEDECRW*DUMLOSS SUEDECRW*ABNGOOGLE SUEDECRW*VOLUME SUEDECANL SUEDECANL*BETA SUEDECANL*LOGMKTVAL SUEDECANL*BOOKMKT SUEDECANL*DUMLOSS SUEDECANL*ABNGOOGLE SUEDECANL*VOLUME R

22 Table 5 (continued) Summary statistics for announcement week returns-earnings regressions. Panel B: The association between pre-announcement week abnormal Google search activy and the returns-earnings relation. Coeff t Coeff t Coeff t Coeff t Firms wh zero Google activy and noisy tickers removed (1173 firms, observations) Intercept BETA LOGMKTVAL BOOKMKT DUMLOSS SUEDECRW SUEDECRW*BETA SUEDECRW*LOGMKTVAL SUEDECRW*BOOKMKT SUEDECRW*DUMLOSS SUEDECRW*ABNGOOGLE SUEDECRW*ABNGOOGLEPREANN SUEDECRW*VOLUME SUEDECANL SUEDECANL*BETA SUEDECANL*LOGMKTVAL SUEDECANL*BOOKMKT SUEDECANL*DUMLOSS SUEDECANL*ABNGOOGLE SUEDECRW*ABNGOOGLEPREANN SUEDECANL*VOLUME R The LHS variable is weekly abnormal return for the earnings announcement week measured relative to the equally weighted market index. BETA is a firm s beta. LOGMKTVAL is the log of market value and BOOKMKT is the book-to-market ratio, both measured at the end of the previous quarter. DUMLOSS is a dummy set to one if the firm reports a loss. SUEDECRW is the decile ranking of unexpected earnings deflated by prior quarter ending stock price, using earnings from the same quarter in the previous year as expected 21

23 earnings. ABNGOOGLE is the log of rescaled Google search activy less the average log of rescaled Google search activy over the previous seven weeks. ABNGOOGLEPREANN is average ABNGOOGLE over the five weeks prior to an earnings announcement. VOLUME is weekly trading volume divided by shares outstanding. SUEDECANL is the decile ranking of unexpected earnings deflated by prior quarter ending stock price, using the most recent I/B/E/S consensus earnings forecast as expected earnings. Standard errors are based on based on clustering at both the firm and week level. 22

24 Table 6 Summary statistics for post-announcement returns-earnings regressions. Panel A: LHS variable is abnormal return in the week following earnings announcement. Coeff t Coeff t Coeff t Coeff t Coeff t Coeff t Firms wh zero Google activy and noisy tickers removed (1173 firms, observations) Intercept BETA LOGMKTVAL BOOKMKT DUMLOSS SUEDECRW SUEDECRW*BETA SUEDECRW*LOGMKTVAL SUEDECRW*BOOKMKT SUEDECRW*DUMLOSS SUEDECRW*ABNGOOGLE SUEDECRW*ABNGOOGLEPOSTANN SUEDECRW*VOLUME SUEDECANL SUEDECANL*BETA SUEDECANL*LOGMKTVAL SUEDECANL*BOOKMKT SUEDECANL*DUMLOSS SUEDECANL*ABNGOOGLE SUEDECRW*ABNGOOGLEPOSTANN SUEDECANL*VOLUME R

25 Table 6 (continued) Summary statistics for post-announcement returns-earnings regressions. Panel B: LHS variable is abnormal return in the 13 weeks following earnings announcement. Coeff t Coeff t Coeff t Coeff t Coeff t Coeff t Firms wh zero Google activy and noisy tickers removed (1173 firms, observations) Intercept BETA LOGMKTVAL BOOKMKT DUMLOSS SUEDECRW SUEDECRW*BETA SUEDECRW*LOGMKTVAL SUEDECRW*BOOKMKT SUEDECRW*DUMLOSS SUEDECRW*ABNGOOGLE SUEDECRW*ABNGOOGLEPOSTANN SUEDECRW*VOLUME SUEDECANL SUEDECANL*BETA SUEDECANL*LOGMKTVAL SUEDECANL*BOOKMKT SUEDECANL*DUMLOSS SUEDECANL*ABNGOOGLE SUEDECRW*ABNGOOGLEPOSTANN SUEDECANL*VOLUME R The LHS variable is weekly abnormal return for the earnings announcement week measured relative to the equally weighted market index. BETA is a firm s beta. LOGMKTVAL is the log of market value and BOOKMKT is the book-to-market ratio, both measured at the end of the previous quarter. DUMLOSS is a dummy set to one if the firm reports a loss. SUEDECRW is the decile ranking of unexpected earnings deflated by prior quarter ending stock price, using earnings from the same quarter in the previous year as expected 24

26 earnings. ABNGOOGLE is the log of rescaled Google search activy less the average log of rescaled Google search activy over the previous seven weeks. ABNGOOGLEPOSTANN is average ABNGOOGLE over the same post announcement period as used for the LHS variable. VOLUME is trading volume divided by shares outstanding over the same post announcement period as used for the LHS variable. SUEDECANL is the decile ranking of unexpected earnings deflated by prior quarter ending stock price, using the most recent I/B/E/S consensus earnings forecast as expected earnings. Standard errors are based on based on clustering at both the firm and week level. 25

27 Figure 1 Average unscaled Google search activy around earnings announcements * Panel A: All Quarters 42.5 Average Google search activy Week relative to earnings announcement Panel B: 4 th Quarter Average Google search activy Week relative to earnings announcement * The figures present average unscaled Google search activy in event time relative to earnings announcement weeks. Unscaled Google search activy measures search activy relative to the maximum search activy week for each firm across the time period. 26