Online Search Activity and Market Reaction to Earnings Announcements

Abstract

This paper leverages Google Trends search volume data from 2004 to 2008 as a proxy for investor information demand. The analysis documents that greater search activity prior to earnings announcements is positively associated with future market reaction to earnings announcements, pre-earnings announcement drift, and buying pressure. The results are consistent with investors gathering value-relevant information through online research, which is subsequently incorporated into prices through trading around earnings announcements. Notably, search volume is positively associated with market reaction to earnings announcements and pre-announcement drifts for more obscure firms where data is scarce. Overall, this paper provides large-sample evidence validating theoretical models where dispersed private information is incorporated into stock prices. The findings suggest that broader data access may facilitate pricing efficiency by promoting more informed market participation.

1. Introduction

Stock prices aggregate information dispersed across diverse geographic locations and investors, thereby enhancing their informativeness (S. Grossman, 1976). This aggregative process is further enriched by the heterogeneous research efforts of individual investors, whose collective analysis embeds new insights into stock prices (Kim & Verrecchia, 1997). Empirically observing when investors engage in information gathering has been challenging, as such behavior is inherently private. However, the emergence of digital platforms has enabled researchers to proxy for investor attention using online search behavior. The existing literature has used Google search volume data as a signal of information demand. Da et al. (2011) demonstrate that spikes in Google search volume index (SVI) predict short-term returns and partial reversals. Drake et al. (2012) extend this analysis to earnings announcements, showing that abnormal search activity prior to releases predicts return drift and anticipatory trading patterns. These studies suggest that investor attention plays a critical role in aggregating dispersed information into prices.

While prior attention in research has largely focused on price effects, less is known about how attention translates into trading volume and order flow—the mechanisms through which prices adjust. Theoretically, if attention-driven searches reflect informed trading, one should observe both pre-announcement price drift and elevated, directionally skewed trading volume. Attention often leads to asymmetric trading behavior: individual investors, for example, tend to be net buyers of attention-grabbing stocks (Barber & Odean, 2007). Thus, unusually high search interest may induce buy–sell imbalances if investors act on bullish insights. Conversely, if negative information is uncovered, retail investors are less likely to short-sell, leading to a muted response. Whether such imbalances result in permanent or temporary price impacts depends on whether trades are informed or sentiment-driven.

Importantly, Drake et al. (2012) provide evidence that speaks directly to whether search captures informed or noise trading. They document a “preemption” effect: when pre-announcement Google searches are elevated, more of the earnings information is incorporated into prices before the announcement, resulting in a diminished price reaction to the announcement itself. As Drake et al. (2012, p. 1002) explicitly state, this finding “is inconsistent with the measure solely capturing the behavior of less sophisticated retail investors (or noise traders).” If search activity merely reflected uninformed attention, it would not systematically preempt earnings information content. The preemption result suggests that elevated search activity reflects genuine information acquisition that is subsequently incorporated into prices through informed trading.

This paper makes three distinct contributions beyond the foundational work of Da et al. (2011) and Drake et al. (2012). First, while Da et al. (2011) document price reversals consistent with noise trading, we find no post-announcement reversals—price effects are permanent rather than transient, suggesting information aggregation rather than sentiment-driven trading. Second, we extend beyond price outcomes to examine the microstructure mechanism through order flow decomposition, showing that search activity predicts directional buying pressure. Third, we document pronounced heterogeneity: effects are approximately eight times larger for small firms (coefficient 0.0205) compared to large firms (coefficient 0.0026), validating theoretical predictions that information acquisition is most valuable in information-scarce environments (S. J. Grossman & Stiglitz, 1980; Hong et al., 2000).

This paper contributes new evidence on how investor search activity affects trading dynamics around earnings announcements. Specifically, we show that abnormal Google search volume in the weeks preceding an earnings release predicts (1) pre-announcement return drift; (2) CAR around earnings announcements; and (3) buy–sell imbalances indicative of buying pressure. These patterns are strongest among smaller, less-followed firms, where information is scarcer and the marginal value of research is greater (DellaVigna & Pollet, 2009; Hong et al., 2000). In such cases, online search activity appears to surface value-relevant insights not yet reflected in prices, allowing informed investors to act ahead of the earnings news.

The 2004–2008 window offers several advantages that make it particularly well-suited for studying the relationship between online search behavior and price formation. First, our sample period aligns with foundational work that established Google Trends as a valid proxy for investor attention. Da et al. (2011) used data from January 2004 through June 2008 in their Journal of Finance paper showing that search volume predicts IPO returns. Drake et al. (2012) analyzed the same period in the Journal of Accounting Research when studying information demand around earnings announcements. By using comparable data, we can directly compare our methodological extensions to these benchmark studies.

Regarding concerns about the 2008 financial crisis potentially confounding results, we note that Da et al. (2011) explicitly tested subperiod robustness by splitting their sample into January 2004–May 2006 versus June 2006–June 2008. They report that “the regression results are qualitatively similar in the two subsample periods” (Table VII, p. 1483). This direct test from the foundational paper provides strong precedent that SVI–return relationships are robust to the inclusion of crisis-period observations. Additionally, our regression specifications include year-quarter time-fixed effects, which absorb all aggregate time-varying shocks common across firms—including crisis-induced volatility and systematic changes in investor behavior. Our identification relies on cross-sectional variation in search activity within each year-quarter, not on time-series variation that could be confounded by crisis dynamics.

The 2004–2008 sample period offers an additional methodological advantage: it largely predates social media influence on financial markets. Twitter launched in July 2006 but achieved mass adoption among retail investors only years later; Reddit was founded in 2005, but the WallStreetBets community became influential only around 2012; Facebook was restricted to college students until September 2006; StockTwits launched in 2008 at the very end of our sample. Thus, the social media-driven herding and coordinated retail activity that characterize recent markets were essentially absent during our sample period. This represents a feature of our research design: we study information-seeking behavior in an environment before coordinated social media activity could confound the signal.

We include a number of robustness checks confirming the tangible influence of investor information seeking in markets. The relationship holds even when controlling for past returns and news coverage that could also drive search activity and trading (Engle et al., 2021). We also consider three alternative explanations that may explain the results. First, it is possible that the abnormal search index is merely driven by news stories or by the efforts of a company through advertising expenses. However, the results are robust to the inclusion of news stories and measures of advertising expenses. Second, the “liquidity provision” hypothesis states that retail investors provide liquidity to institutional investors by taking the other side of institutional trades. Retail investors demand a premium for providing liquidity and thus exert temporary pressure on prices (Kaniel et al., 2008). In such a case, we should see price reversals. Third, it is possible that current-period retail buying predicts more retail buying in subsequent periods (autocorrelated flow hypothesis). The buying pressure leads to high future returns (Dorn et al., 2008; Barber et al., 2009). These mechanisms predict a transient impact on prices, and any price increases resulting from them should subsequently reverse. By examining price reversals, we eliminate these mechanisms from the analysis.

The paper is organized as follows: Section 2 provides the literature review and motivates the hypotheses; Section 3 describes the data; Section 4 presents the methodology and results; Section 5 discusses the results; Section 6 provides limitations and future research directions; and Section 7 concludes the paper.

2. Literature Review and Hypothesis Development

2.1. Literature Review

2.1.1. Theoretical Underpinnings

Information acquisition incentives outlined by S. J. Grossman and Stiglitz (1980) predict investors will research stocks when personal benefits outweigh costs. Their gathered insights then enter prices when sufficiently profitable to trade on (Kim & Verrecchia, 1997). These models formally articulate predicted mechanisms linking dispersed knowledge gathering to pricing through investors individually optimizing research and then collectively revealing information. Specifically, S. J. Grossman and Stiglitz (1980) mathematically state investors will acquire knowledge until marginal costs equal marginal benefits. As more investors research, prices become more informative. Kim and Verrecchia (1997) build on these incentives to formally model how privately informed trading aggregated across market participants reveals information through price movements.

The S. J. Grossman and Stiglitz (1980) framework yields specific testable predictions. First, the marginal value of information acquisition is highest when prior public information is scarce. This predicts stronger search–return relationships for small firms with limited analyst coverage and media attention. Second, information value is higher when uncertainty is elevated. This predicts stronger effects when analyst forecast dispersion is high. Third, if a search captures informed trading, price effects should be permanent; if a search captures noise trading, price effects should reverse. We test each of these predictions empirically.

2.1.2. Google Trends

Google Trends data has been used in various disciplines and settings. Carneiro and Mylonakis (2009) show that one can estimate the level and outbreak of flu activity in different geographic areas by looking at search terms related to flu. Choi and Varian (2009) use the Google search index as a proxy for the population’s demand for information. Google Trends has become an increasingly valuable data source for finance and accounting research. Various papers demonstrated its potential by showing that the search volume contains information on investor attention and predicts market activity. Specific applications include using company-name searches to predict trading volumes (Da et al., 2011; Drake et al., 2012), detecting industry concentration risk from keyword correlations (Dimpfl & Jank, 2016), and relating search intensity to IPO underpricing (Blankespoor et al., 2017).

This paper builds on pioneering work by Da et al. (2011) utilizing Google Trends data from 2004 to 2008. They find that Google search volume index (SVI) of stock tickers predicts positive price pressure on stocks and contributes to IPO underperformance. This paper is also related to Drake et al. (2012) who analyze investor demand for information around corporate announcements and find that when pre-announcement Google searches are high, some portion of the information content of earnings is preempted.

While these foundational studies establish the validity of Google search as an attention proxy, important gaps remain. Da et al. (2011) focus on the Price Pressure Hypothesis, documenting that retail attention creates temporary buying pressure that subsequently reverses—consistent with noise trading rather than informed trading. Their primary finding of short-term price spikes followed by underperformance contrasts sharply with our documentation of permanent price effects. Drake et al. (2012) examine price and volume reactions but do not decompose order flow to identify whether attention translates into buying or selling pressure. Neither study examines the microstructure mechanism through which attention enters prices. Our analysis addresses these gaps by (1) testing for reversals and finding none, (2) decomposing trading volume into buyer-initiated and seller-initiated trades, and (3) documenting pronounced heterogeneity across firm information environments.

2.2. Hypothesis Development

In this section, we empirically test the predicted mechanism that increased information seeking prompts informed trading, revealing a positive relationship between search volume and subsequent abnormal returns and buying pressure. Examining effects around earnings announcements provides a clear informational event when investors are likely seeking and trading on new insights.

The earnings announcement process allows companies to consolidate their sales and profit figures and disclose these results publicly. An earnings surprise occurs when a company’s performance exceeds or falls short of market expectations. This surprise represents new information that, though dispersed in fragmented form across the company’s various employees and customers, only becomes available to the market once management has compiled and released it on the announcement date.

Changes in the search index also mirror the above aggregation of information. A multitude of customers and employees of the company observe the company’s products and services firsthand. When they get excited about the company, they may decide to research the company further before buying its stock.

Hence, we can expect that the search index of a firm will be associated with future market reaction to earnings announcements. We utilize the cumulative abnormal return (CAR) centered on the earnings announcement as a measure of market reaction to the earnings announcement. The CAR reflects the impact of new information that is disclosed by the earnings announcement after taking into account market expectations. An increase in searches for a firm indicates that retail investors are indulging in information-seeking behavior and are actively seeking information about the company. This behavior could have been triggered by some private information or observation about the firm or it can even be a mere noise trading reaction (Friedman & Zeng, 2022). If their research sheds a positive light on the company, they end up buying the stock. However, if the investors uncover something negative, they tend not to do anything, since retail investors tend not to short-sell. Thus, we state our first hypothesis as:

Hypothesis 1.

Investor information seeking prior to earnings announcements is associated with positive abnormal returns around the earnings announcements.

Information about smaller companies is more challenging to find because they receive less media and institutional coverage (Hong et al., 2000). We test if search activity is more predictive for smaller stocks, where data is limited (DellaVigna & Pollet, 2009). Thus, we analyze whether search impacts will be stronger for such lesser-known stocks, based on the difficulty of gathering insights. Detecting a moderating effect would confirm that investors disproportionately leverage searches for hard-to-analyze companies, profiting from uncovering rare information (Tetlock, 2010). Thus, our second hypothesis is as follows:

Hypothesis 2.

The relationship between investor information seeking before earnings announcements and returns will be stronger for smaller companies where information is harder to find.

Earnings announcements represent important informational events, yet stocks under-react, exhibiting continued earnings announcement drift (Bernard & Thomas, 1989). The search data provides a proxy for investors gathering valuable information before announcements. We empirically test if search activity predicts pre-announcement drift, highlighting informed trading (Drake et al., 2012). This hypothesis specifically examines if higher information-seeking behavior by investors before earnings announcements is associated with pre-announcement returns. Detecting this link would confirm pre-announcement investor-level research, potentially driven by anticipation of valuable insights, manifests in trading and buying pressure on the stocks (Hirshleifer et al., 2009). Thus, our third hypothesis is as follows:

Hypothesis 3.

Pre-earnings announcement returns are positively associated with greater investor information-seeking behavior before earnings.

In the above analysis, we argue that market prices are informative since investors trade based on their information. In such a case, we should observe the effect of the search index on trading volume. To link the search index with actual trading activity, we examine the effect of the search index on future trading volume. It can be argued that unsigned volume merely represents the change in ownership of shares; that is to say, for every buyer, there is a seller. However, on average, stock prices tend to rise in periods of high volume and fall in periods of low volume, as shown by Karpoff (1987). If an increase in the search index leads to an increase in buying activity before the earnings announcement, we should also see a subsequent increase in trading volume. Specifically, given the short-selling constraints faced by individual investors (Barber & Odean, 2007), the effect should be stronger for buy orders. Thus, our fourth hypothesis is as follows:

Hypothesis 4.

Buy orders around earnings announcements are positively related to greater investor information-seeking behavior before earnings.

3. Data

3.1. Google Trends Data

This study utilizes Google Trends search volume indexes to proxy for investor information acquisition from 2004 to 2008. Google Trends reports search popularity for keywords as percentage deviations from 1 January 2004 levels, rather than absolute volumes. We use the same time period (2004–2008) as Da et al. (2011) and Drake et al. (2012) to extend the results of these papers.

Google Trends normalizes SVI to a 0–100 scale where 100 represents peak search interest for that term during the sample period. Because this index is relative rather than absolute, we use changes in SVI (SVIChange) as our primary independent variable, measuring the deviation from a firm’s typical search intensity. A one-standard-deviation increase in SVIChange corresponds to approximately 15 index points in our sample.

Google Trends provides a normalized search volume index constructed from Google web search frequency data. For our analysis, we download the weekly search index for S&P 1500 firms, representing almost 90% of U.S. market capitalization. To construct a meaningful metric, we create a weekly search index reflecting the volume of searches related to a particular stock. A higher search index value indicates heightened interest from a sizable audience seeking information about a company.

To ensure our search volume measure captures investment-related information seeking rather than non-financial queries, we implement systematic ticker filtering following Da et al. (2011). We cross-referenced all S&P 1500 ticker symbols against standard English dictionaries and excluded 178 tickers corresponding to common words or abbreviations that could generate non-investment searches. Examples include “GAP” (clothing retailer versus the common word), “CAT” (Caterpillar versus the animal), “LOVE,” “BABY,” “LAWS,” and “TRY.” Single-letter tickers and highly ambiguous short symbols were also removed. This conservative filtering minimizes false positives from non-financial queries. Additionally, our panel specification with firm-fixed effects identifies effects from within-firm variation in search intensity over time, absorbing any consistent non-investment noise associated with particular tickers.

3.2. Earnings Announcements

Publicly owned companies are required by law to file quarterly earnings reports with the SEC. Many companies announce the earnings announcement dates in advance, but several companies miss these dates and report late. To determine the actual date of earnings announcements we use Compustat. Compustat records the earnings date as the date when the earnings report appears in the Wall Street Journal or other newspapers. However, the coverage of the Wall Street Journal is biased towards well-known large stocks, and it is possible that earnings announcement dates of smaller stocks are not accurate. To be certain that we have the correct earnings announcement dates, we cross-check the earnings dates for smaller stocks against the Factiva database. We check for news stories to confirm that the date in Compustat is the actual earnings announcement date. In case of conflict, we use the date provided by the news story in Factiva.

3.3. Cumulative Abnormal Return and Standardized Unexpected Earnings (SUE)

CAR centered on the earnings announcement is used as a measure of new information content in the earnings announcement. A market-based measure such as CAR allows one to capture the full information content of earning’s announcement. For example, firms provide extensive disclosure through financial reports, footnotes, management’s discussion of the results and the competitive environment, forecasts, and other forward-looking information. In addition, many firms engage in conference calls where top executives of the firm present the last quarter’s results and answer questions from investors and analysts. This is consistent with the findings of Francis et al. (2002) who find that earnings announcements increasingly serve as a conduit for information different from merely the earnings numbers.

We use standardized unexpected earnings (SUE) as a proxy for market expectations surrounding the earnings announcement. Analyst forecast data is from Institutional Brokers’ Estimate System (I/B/E/S) database in Wharton Research Data Services (WRDS). SUE is defined as [Actual Earnings Per Share (EPS) − Expected Earnings Per Share (EPS)]/(Standard deviation of analyst forecasts for that quarter). Expected EPS is the mean of the analyst forecasts prior to the announcement. If an analyst has made multiple forecasts in this period, only the most recent forecast is used. If there are fewer than two forecasts for that quarter, it leads to the exclusion of the firm for that quarter. The absolute magnitude of SUE measures the degree of earnings surprise, while the sign of SUE signifies if the actual earnings were above or below the consensus forecast. That is, a positive SUE implies that the actual values of EPS came out to be above the analysts estimates and a negative SUE signifies that the earnings numbers came below the mean analyst estimate.

3.4. News Data

We collect weekly news data for S&P 1500 companies from 2004 to 2008. We utilize Google News Archive to search for news (in English only) related to the ticker and company name of a firm. For each week, from the beginning of 2004 to the end of 2008 (260 weeks for each of the S&P 1500 companies), we search for news stories for the company ticker and the company name. The number of news stories for each firm, for a particular week, is recorded and constitutes the variable “News Volume”.

There are various advantages to using Google News data. First, it is arguably the most comprehensive news data available. It is better than the DJI and Lexis Nexis news data sources since it includes many news sources that other databases do not include. For example, many online news sources and local newspapers are not part of DJI and Lexis Nexis. Prominent journal papers that look at the news data typically only look at the top three to five newspapers. Moreover, Google News data is most relevant to this paper as it captures the information available to an online investor.

3.5. Volume Data

We also look at the impact of the searches for a company ticker on the subsequent trading volume of the company’s stock. We use two primary measures of volume: daily buys and sells volume, and daily buys and sells dollar volume (Chordia et al., 2002). Daily buys are calculated using data from the New York Stock Exchange’s (NYSE) Trade and Quote (TAQ) database. The trades for each stock are signed based on the Lee and Ready (1991) algorithm. The quote rule and tick rule are used to identify the trades as buyer-initiated or seller-initiated. If the trade price is above the midpoint of the most-recent bid–ask quote, the trade is classified as buyer-initiated. Additionally, if the trade price is above the last executed trade price, the tick rule classifies the trade as buyer-initiated. Appendix A lists the primary variables and their difinitions.

4. Research Methodology and Empirical Results

4.1. Quarterly Earnings Announcements

One way to test whether the search index has any price impact is to look at events where there is a dissemination of information relevant to stock prices, such as earnings announcements. In this paper, we focus on quarterly earnings announcements for several reasons. All publicly traded companies are required by law to have four quarterly earnings announcements every year. This gives us approximately 30,000 firm-quarter data points to test our hypotheses. Since all public companies must make these announcements, sample selection biases are considerably reduced. The actual number of observations is reduced from 30,000 to 15,345 for two main reasons: First, the Google Trends algorithm sets the search index to zero for tickers with infrequent searches or below their minimum volume threshold. Second, we exclude ticker symbol keywords that match common words.

Table 1. Summary statistics for primary variables.

	(1)	(2)	(3)	(4)	(5)	(6)
Percentile	SVI	NewsVol	Assets ($M)	CAR5day	CAR3day	CAR30daypre
1%	0.0000	0.00	123	−0.2425	−0.2283	−0.2914
5%	0.0000	0.00	303	−0.1308	−0.1175	−0.1449
10%	0.0000	0.25	508	−0.0866	−0.0786	−0.0965
25%	0.0000	0.67	1113	−0.0347	−0.0307	−0.0418
50%	0.1217	1.75	3042	0.0056	0.0049	0.0031
75%	0.2638	4.75	10,310	0.0503	0.0454	0.0496
90%	1.1051	14.75	31,069	0.1010	0.0947	0.1004
95%	1.7055	28.25	57,869	0.1407	0.1310	0.1432
99%	2.0782	84.50	287,583	0.2256	0.2168	0.2483
Mean	0.3078	6.29	19,843	0.0062	0.0061	0.0017
Variance	0.5093	14.99	91,577	0.0865	0.0798	0.0951
Min	0.0000	0	30	−0.7337	−0.6647	−0.9207
Max	2.3366	148	1,817,943	0.7577	0.6491	0.7443
Number of obs	15,345	15,345	15,345	15,345	15,345	15,345

Notes: The above table shows the summary statistics of the variables used in the earnings regressions that follow this table. SVI is the average of the one-month search index prior to the earnings announcement. NewsVol is the number of news stories about a firm that appear in the Google news archive search in the one-month period prior to the earnings announcement. Assets are the total assets of the firm in millions. CAR5day is five-day CAR centered on the earnings date. CAR3day is three-day CAR centered on the earnings date. CAR30daypre is calculated from day −35 to day −5, where day 0 is the day of earnings announcement. The sample period is from January 2004 to December 2008.

shows the salient summary statistics and distribution for the primary variables used in the earnings-related regressions. SVI is the average of the one-month search index prior to the earnings announcement. Five calendar days prior to the earnings announcement are excluded so that there is no effect of earnings leakages on the search index. NewsVol (News Volume) is the total number of news stories that include the name of the company and its ticker symbol, in the one-month period prior to the earnings announcement.

News coverage is heavily skewed towards larger companies. While the top one percentile of companies have an average of 84 or more news stories in the one-month time period prior to the earnings announcement date, almost half the sample firms have on average less than two news stories in the same time period. Mean and median three-day and five-day CAR centered on the earnings announcement date is slightly positive. One would expect the CAR to be zero since the market should not systematically under-react to the earnings information. However, the variance of all the three specifications of CAR is quite high; hence, the positive mean and median are not significant in any of the specifications.

For most of the companies, the earnings announcement process is an opportunity to aggregate sales and profit numbers and share the results with the public. Earnings surprise indicates that the company performed better or worse than the market expectations. Earnings surprise is the new information that, while present in disaggregate form among the diverse employees and customers of the company, gets revealed to the market only when the information has been aggregated by the management of the company and shared with the public on the day of the earnings announcement.

Shifts in search activity similarly reflect this information aggregation process. Numerous customers and employees observe the company’s products and services directly. When they become enthusiastic about the company, they may choose to investigate it further before purchasing its stock.

Hence, we can expect that the search index of a firm will be associated with market reaction to future earnings announcements of the firm. As a proxy for the market reaction to earnings announcements, we utilize the cumulative abnormal return (CAR) centered on the earnings announcement. CAR centered on earnings announcement reflects the impact of new information that is disclosed by the earnings announcement after taking into account market expectations.

The selection of CAR windows follows established conventions in the earnings announcement literature. CAR[−2, +2] and CAR[−1, +1] capture immediate announcement reactions, consistent with standard practice (Cready & Gurun, 2010; Stice, 1991). Pre-announcement windows (CAR[−45, −5] and CAR[−35, −5]) capture gradual information incorporation during the period when investors may be researching upcoming announcements. The five-day exclusion buffer before the announcement prevents contamination from information leakage or insider trading. Window lengths are deliberately chosen to avoid overlap with adjacent quarterly announcements, which occur approximately every 90 days.

General specification for the regression equation is as follows:

$${\displaystyle \sum _x^y}{CAR}_{i,t=}{\alpha }_{i,t}+FirmDummies+TimeDummies+{\beta }_{1}SearchInde{x}_{i,a,b}+ControlVariables+{\epsilon }_{i,t}$$

(1)

Number of firms i = 1, 2,… N; number of quarterly earnings announcements t = 1, 2,…Nq.

${\displaystyle \sum _x^y}{CAR}_{i,t}$ is the cumulative abnormal return for firm i from day t + x to day t + y and is also represented by the suffix [x, y]. Abnormal return is defined as the return of the stock minus the value-weighted market return from CRSP. The difference is summed up over each of the trading days in the period.

The search index for a firm i is the average search index from time period t + a to t + b. We exclude the five days prior to the earnings announcement to mitigate the effects of insider trading or the leakage of earnings information to the market. We also include the average search index from the month before the earnings announcement. This helps account for any consistent characteristics of the search index over time and controls for the previous level of search activity related to the stock. Since we are using firm-fixed effects, the independent variable is the change in the search index from its long-term mean. Hence, we are testing if for a given firm, the change in the level of search index is associated with future abnormal returns.

Table 2. Returns around the earnings announcements.

	Five Day CAR			Three Day CAR			Post 30-Day CAR	Post 45-Day CAR
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
SVIChanget−1	0.007224 **	0.007081 **	0.008036 **	0.007302 **	0.007096 **	0.007777 **	0.001179	0.0007817
	(0.0034)	(0.0034)	(0.0035)	(0.0031)	(0.0032)	(0.0032)	(0.0037)	(0.0043)
LogSizet−1	−0.02434 ***	−0.02328 ***	−0.02397 ***	−0.02323 ***	−0.02239 ***	−0.02447 ***	−0.01983 ***	−0.03307 ***
	(0.0055)	(0.0055)	(0.0057)	(0.0055)	(0.0056)	(0.0057)	(0.0064)	(0.0071)
MarketToBookRatiot−1	−0.01473 ***	−0.01492 ***	−0.01533 ***	−0.01218 ***	−0.01251 ***	−0.01241 ***	−0.01836 ***	−0.02171 ***
	(0.0024)	(0.0024)	(0.0026)	(0.0021)	(0.0021)	(0.0022)	(0.0021)	(0.0027)
Leveraget−1	0.02450 **	0.02779 ***	0.02448 **	0.01721	0.02043 *	0.02015 *	0.004620	0.01035
	(0.0107)	(0.0106)	(0.0111)	(0.0105)	(0.0105)	(0.0107)	(0.0158)	(0.0184)
ReturnOnAssetst−1	0.02375	0.01953	0.04453	0.02275	0.02239	0.05739	0.006758	−0.05318
	(0.0897)	(0.0889)	(0.0927)	(0.0926)	(0.0895)	(0.0925)	(0.0863)	(0.1099)
SG&A Ratiot−1	0.007086	0.005671	−0.01582	0.0001457	−0.002441	−0.02370	−0.09547 ***	−0.1145 ***
	(0.0231)	(0.0229)	(0.0258)	(0.0219)	(0.0226)	(0.0249)	(0.0335)	(0.0358)
TwoMonthLagReturn	−0.02287 ***	−0.02325 ***	−0.02375 ***	−0.02041 ***	−0.02053 ***	−0.02107 ***	−0.01132 *	−0.02356 ***
	(0.0073)	(0.0075)	(0.0078)	(0.0072)	(0.0074)	(0.0077)	(0.0066)	(0.0082)
SUE		0.001709 ***	0.002171 **		0.001474 **	0.001853 *	0.0001222	−0.0002371
		(0.0007)	(0.0010)		(0.0006)	(0.0010)	(0.0001)	(0.0002)
Forecast Stdev		0.0005702	0.0005620		0.005790	0.005815	−0.02356	0.02577
		(0.0195)	(0.0203)		(0.0197)	(0.0204)	(0.0192)	(0.0261)
NewsVolumet−1			0.00008158 ***			0.00007982 ***	−0.00003448	−0.00003825
			(0.0000)			(0.0000)	(0.0000)	(0.0000)
Constant	0.2058 ***	0.1978 ***	0.2090 ***	0.1920 ***	0.1858 ***	0.2035 ***	0.2029 ***	0.3099 ***
	(0.0391)	(0.0400)	(0.0416)	(0.0389)	(0.0399)	(0.0414)	(0.0462)	(0.0503)
Observations	11,767	11,340	10,792	11,767	11,340	10,792	10,792	10,792
Adjusted R-squared	0.0174	0.0440	0.0542	0.0149	0.0384	0.0471	0.0228	0.0344
Number of Firms	770	766	729	770	766	729	729	729
Firm-Fixed Effects	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Time-Fixed Effects	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes

Notes: The above table reports the results of the regressions with dependent variable CAR around the earnings announcements. Specifications 1, 2 and 3 show the results of quarterly regression of the five-day cumulative abnormal return, centered on the earnings announcement date, on the lagged month-over-month change in SVI. Abnormal return is calculated by subtracting the value-weighted market return from the stock return. SVIChange is the difference between average search index from day −35 to −5 and the average search index from day −66 to −36, where day 0 is the day of the earnings announcement. NewsVolume is the number of news stories that mention the firm in the (−36 to −66) time period. SUE is standardized unexpected earnings, defined as actual EPS minus the mean of analyst forecasts normalized by the standard deviation of analyst forecasts (Forecast Stdev). The following variables are based on the previous quarter’s numbers: Log Size is the natural log of sales. MarketToBook ratio is defined as the sum of total assets and market value of equity minus the book value of equity divided by total assets. Leverage is defined as long-term debt plus current liabilities divided by long-term debt plus current liabilities plus book value of common equity. ReturnOnAssets is income before extraordinary items divided by lagged assets value. SG&ARatio is the SG&A expense divided by sales. TwoMonthLagReturn is the return in the period from (−60, −120). For specifications 7 and 8, CAR is calculated from day (+5 to +30) and day (+5 to +45), where day 0 is the day of the earnings announcement. All days are calendar days. Standard errors in parentheses are White standard errors clustered by firm. The sample period is from January 2004 to December 2008. ***, **, * denote 1%, 5%, and 10% significance levels, respectively.

shows the results for the above specification. Specification (1) in Table 2 shows that the one-month search index prior to the earnings announcements predicts five-day CAR at a 5% level of significance. A greater increase in the search index is positively associated with a larger market reaction to the earnings announcement. The results are economically significant. A 10% increase in the search index is associated with an increase of almost 8 basis points in the three-day CAR. Considering the mean market capitalization of about USD 29 billion for an S&P 1500 company, this translates to approximately USD 2.3 million over a three-day period.

It is possible that the results are being driven by investors who get overly excited when they see a stock price going up prior to the earnings announcement and then trade excessively, pushing the price up temporarily. Thus, it is the extreme past returns (Barber & Odean, 2007) that are driving the results. We control for the possibility that past returns may be driving the results by including a lagged two-month return. Results are robust to the inclusion of one to six months of lagged returns. Specifications in Table 2 also include other control variables commonly used in the earnings announcement regressions (Kothari, 2001).

It is possible that the search index is capturing the information content in the analyst reports. Financial analysts collect information from diverse sources, studying the past and current performance of the companies they follow and making forecasts about the company. The search index could be merely substituting for the information and news stories about the company that analysts use to predict earnings and is already available to investors. It is important to check if the results hold, even while controlling for the analysts forecasts about the earnings. Analysts tend to follow companies closely and their reports should be able to capture part of the information contained in the search index. We find that in the case of the market reaction to earnings announcements, the effect is mitigated by the addition of a proxy for the analyst information prior to the earnings announcement (SUE (standardized unexpected earning)) indicating that analysts are indeed relaying to the markets part of the information that is contained in the search index.

Table 2 (specification 2) shows the results once we include the SUE in the regression equation. To control the information environment and uncertainty surrounding the earnings announcement, we also include the standard deviation of analysts forecast prior to the announcement. SUE is highly significant, as one would expect, and it does have an impact on the search index in terms of diminishing the magnitude of the coefficient of the search index variables. However, the search index retains most of its explanatory power with coefficient of 0.007081 and is still significant at the 5% level. The above results indicate that the search index has some of the information content that is being captured by the analysts.

To check if prior news coverage is driving the results, we include the variable NewsVol in specification 3. NewsVol is the number of news stories that mention the firm in the one-month period prior to the earnings announcement date. The coefficient of NewsVol variable is highly significant, indicating that, for the same firm, the periods in which it gets higher news coverage are associated with higher CAR around the future earnings announcement date.

Market to book ratio (MTB) and lagged returns have negative and significant coefficients. It seems that in the periods where there is a run up in the stock price prior to the earnings announcements there is a negative effect on the CAR around the date of an earnings announcement, indicating SG&A proxies for the advertising expense incurred by a firm. A higher number of advertisements may lead consumers to research a company. SG&A is included to control for the possibility that it is merely advertising that is driving the results.

To control for any spurious effects that could be contaminating the stock price in the five-day window, we also look at a shorter window specification. Specifications 4, 5 and 6 detail the results for the three-day CAR, centered on earnings announcement. The results are robust to the alternative specifications of the event window size.

As a further robustness check, we consider the possibility that investors get overly excited about the company for reasons unrelated to information, which causes a subsequent run-up in prices prior to and around the earnings announcement (Sifat & Thaker, 2020). In such a scenario, one can expect reversals to occur in the stock price after the announcement. We test for price reversals up to 45 days after the announcement date. Table 2 (specifications 7 and 8) shows the result of the regression on CAR from day +5 to day +45 and from day +5 to day +35. The coefficient of the search index prior to the earnings announcement is positive for the 30-day CAR and negative for the 40-day CAR. However, the coefficient is insignificant in both specifications, and it does not appear that any significant price reversals are taking place.

The absence of post-announcement reversals documented in specifications 7 and 8 is theoretically significant and consistent with the “preemption” effect documented by Drake et al. (2012). Their study found that when pre-announcement search activity is elevated, the market reaction at the earnings announcement is attenuated because a larger portion of the earnings news has already been incorporated into prices during the pre-announcement period. This pattern—information flowing into prices before the announcement rather than being followed by reversals after—characterizes informed trading rather than noise trading. If search-driven price movements were attributable to sentiment or uninformed attention, we would expect two observable patterns: (1) no preemption of earnings news, since noise traders lack information, and (2) subsequent price reversals as temporary demand pressure dissipates. The empirical evidence contradicts both predictions. Drake et al. (2012) document significant preemption, and we document no reversals. Together, these findings provide convergent evidence that elevated search activity proxies for information acquisition that permanently enters prices, consistent with the theoretical framework of S. J. Grossman and Stiglitz (1980).

Overall, the various specifications and results in Table 2 support Hypothesis 1. Models by S. J. Grossman and Stiglitz (1980), Holthausen and Verrecchia (1990), and Kim and Verrecchia (1997) predict that an investor will choose to become informed through information acquisition if the expected benefit exceeds the cost. Hence, if it is the information content of searches that is driving the results, then we expect the results will be stronger for firms where information is less readily available.

To test if the impact of searches is higher for smaller firms, we rerun the regression in

Table 3. Returns around the earnings announcements by size.

	Five Day CAR		Three Day CAR		Post 30 CAR		Post 45 CAR
	Small Firms	Large Firms	Small Firms	Large Firms	Small Firms	Large Firms	Small Firms	Large Firms
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
SVIChanget−1	0.02051 ***	0.002528	0.01873 ***	0.001499	0.01224 *	−0.006525	0.01878 **	−0.006502
	(0.0072)	(0.0057)	(0.0072)	(0.0052)	(0.0070)	(0.0064)	(0.0086)	(0.0074)
LogSizet−1	−0.03447 ***	−0.01140	−0.03646 ***	−0.008361	−0.03538 ***	−0.01078	−0.04902 ***	−0.01876
	(0.0100)	(0.0086)	(0.0106)	(0.0086)	(0.0114)	(0.0134)	(0.0132)	(0.0141)
MarketToBookRatiot−1	−0.01549 ***	−0.02605 ***	−0.01308 ***	−0.02075 ***	−0.01577 ***	−0.01746 ***	−0.01661 ***	−0.02639 ***
	(0.0038)	(0.0058)	(0.0034)	(0.0057)	(0.0030)	(0.0056)	(0.0038)	(0.0055)
Leveraget−1	0.01480	0.03038	0.01000	0.02309	−0.01121	0.04720 *	−0.0008624	0.04766
	(0.0206)	(0.0206)	(0.0196)	(0.0203)	(0.0296)	(0.0241)	(0.0362)	(0.0295)
ReturnOnAssetst−1	−0.01514	0.09375	0.006347	0.01537	0.04058	−0.1874	−0.1436	−0.3702
	(0.1851)	(0.1471)	(0.1812)	(0.1643)	(0.1369)	(0.1827)	(0.1690)	(0.2283)
SG&A Ratiot−1	−0.006126	−0.1218 **	−0.02189	−0.1134 **	−0.1402 ***	−0.04715	−0.1348 ***	−0.1143
	(0.0391)	(0.0592)	(0.0395)	(0.0560)	(0.0460)	(0.0834)	(0.0476)	(0.0999)
TwoMonthLagReturn	−0.01248	−0.04076 ***	−0.009664	−0.03534 **	−0.01911 **	−0.01445	−0.02566 **	−0.03839 **
	(0.0120)	(0.0154)	(0.0128)	(0.0143)	(0.0089)	(0.0153)	(0.0111)	(0.0168)
SUE	0.001483 **	0.006346 ***	0.001230 *	0.005730 ***	−0.00001437	0.0008204 *	−0.0003286 **	−0.0002516
	(0.0006)	(0.0010)	(0.0007)	(0.0008)	(0.0001)	(0.0005)	(0.0001)	(0.0007)
Forecast Stdev	−0.01517	0.1219 ***	−0.02980	0.1456 ***	0.03860	−0.05089	0.1259	−0.01209
	(0.0526)	(0.0305)	(0.0536)	(0.0510)	(0.0989)	(0.0905)	(0.1612)	(0.0730)
NewsVolumet−1	0.00008785 ***	0.00006897	0.00009106 ***	0.00008929 **	−0.00005429	−0.00006075	−0.00001721	−0.00004096
	(0.0000)	(0.0000)	(0.0000)	(0.0000)	(0.0000)	(0.0001)	(0.0000)	(0.0001)
Constant	0.2254 ***	0.1873 **	0.2276 ***	0.1465 *	0.2845 ***	0.1303	0.3635 ***	0.2533 **
	(0.0591)	(0.0764)	(0.0620)	(0.0755)	(0.0659)	(0.1188)	(0.0739)	(0.1188)
Observations	3807	3409	3807	3409	3807	3409	3807	3409
Adjusted R-squared	0.0587	0.1104	0.0528	0.0992	0.0356	0.0201	0.0448	0.0358
Number of Firms	446	348	446	348	446	348	446	348
Firm-Fixed Effects	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Time-Fixed Effects	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes

Notes: The above table reports the results of the regressions run on small and large firms separately. The sample is divided into three bins based on size. “Small Firms” constitute the firms in the smallest-size bin while the “Large Firms” constitute the firms in the largest-size bin. Specifications 1 and 2 show the results of regression of the five-day cumulative abnormal return, centered on the earnings announcement date, on the lagged Search Index. Abnormal return is calculated by subtracting the value-weighted market return from the stock return. SVIChange is the difference between average search index from day −35 to −5 and the average search index from day −66 to −36, where day 0 is the day of earnings announcement. NewsVolume is the number of news stories that mention the firm in the (−36 to −66) time period. SUE is standardized unexpected earnings, defined as actual EPS minus the mean of analyst forecasts normalized by the standard deviation of analyst forecasts (Forecast Stdev). The following variables are based on the previous quarter’s numbers: Log Size is the natural log of sales. MarketToBook ratio is defined as the sum of total assets and market value of equity minus the book value of equity divided by total assets. Leverage is defined as long-term debt plus current liabilities divided by long-term debt plus current liabilities plus book value of common equity. ReturnOnAssets is income before extraordinary items divided by lagged assets value. SG&ARatio is the SG&A expense divided by sales. TwoMonthLagReturn is the return in the period from (−60, −120). For specifications 5 and 6, CAR is calculated from day (+5 to +30) and for specifications 7 and 8, CAR is calculated from day (+5 to +45), where day 0 is the day of the earnings announcement. All days are calendar days. Standard errors in parentheses are White standard errors clustered by firm. The sample period is from January 2004 to December 2008. ***, **, * denote 1%, 5% and 10% significance levels, respectively.

separately for small and large firms. The sample is divided into three bins based on size. “Small Firms” constitute firms in the smallest-size bin while “Large Firms” constitute firms in the largest-size bin. In Table 3, specifications 1 and 2, with the dependent variable as the 5-day CAR centered at the earnings announcement date, reveal that the coefficient for SVIChange is 0.02051 for smaller firms and is significant at the 1% level. In contrast, for larger firms, the coefficient is 0.00258 and is not significant even at the 10% level. Similarly, in specifications 3 and 4 for the 3-day CAR, the coefficient for SVIChange is 0.01873 and significant at the 1% level for smaller firms, while for larger firms it is almost 10 times smaller at 0.001499 and insignificant.

The pronounced size heterogeneity documented in Table 3 directly validates the S. J. Grossman and Stiglitz (1980) prediction that information acquisition has higher marginal value when prior public information is scarce. The coefficient for small firms (0.0205) is approximately eight times larger than for large firms (0.0026), and only the small-firm coefficient is statistically significant. This pattern is inconsistent with pure attention or sentiment explanations, which would not predict such systematic variation based on information environment. Instead, it suggests that online search functions as a substitute for traditional information intermediaries—investors use search to gather insights on firms where analyst coverage and media attention are limited. This finding aligns with conceptual frameworks predicting higher personal search benefits when costs of accessing public information exceed gains, motivating investors to privately seek obscure data (S. J. Grossman & Stiglitz, 1980). Results indicate investors specifically leverage search as a tool for discovering data on more opaque companies, thereby supporting Hypothesis 2.

As a more stringent test of the information aggregation story, we also look at the pre-earnings drift to determine if, in fact, the information has been gradually flowing to the market prior to the earnings announcement.

4.2. Pre-Earnings Announcement Drift

The pre-earnings announcement period allows us to test the hypothesis that the search index is related to the gradual flow of information to the stock market prior to the earnings announcement date. Pre-earnings announcement drift is defined as the CAR n days prior to the earnings announcement date. If a higher number of searches led to net buying behavior, we should expect that the more a stock is searched before its earnings announcement the higher will be its expected pre earnings announcement drift.

We look at the 30-calendar-day CAR from day −35 to day −5 and the 40-calendar-day CAR from day −45 to day −5, where 0 is the earnings announcement date. We exclude the last five days prior to the earnings announcement to mitigate the effects of insider trading or the leakage of earnings information to the market. Abnormal return is defined as the return of the stock minus the value-weighted market return from CRSP. The difference is summed up over each of the trading days in the 30 and 40 days prior to the announcement.

The independent variable of interest is the change in the prior month search index. Specification 1 in

Table 4. Pre-earnings announcement returns.

	CAR Pre 45 Days			CAR Pre 30 Days
	(1)	(2)	(3)	(4)	(5)	(6)
SVIChanget−1	0.006441 *	0.009448 **	0.009944 **	0.01008	0.01202	0.01103
	(0.0037)	(0.0040)	(0.0041)	(0.0067)	(0.0075)	(0.0075)
LogSizet−1	−0.03781 ***	−0.04643 ***	−0.04737 ***	−0.02668 ***	−0.03255 ***	−0.03006 ***
	(0.0059)	(0.0069)	(0.0072)	(0.0051)	(0.0062)	(0.0065)
MarketToBookRatiot−1	−0.01396 ***	−0.01604 ***	−0.01771 ***	−0.008580 ***	−0.01023 ***	−0.01102 ***
	(0.0024)	(0.0029)	(0.0031)	(0.0025)	(0.0032)	(0.0035)
Leveraget−1	−0.004134	−0.004105	−0.002902	−0.003195	0.005601	0.009294
	(0.0143)	(0.0158)	(0.0166)	(0.0125)	(0.0139)	(0.0144)
ReturnOnAssetst−1		−0.1650	−0.1517		−0.1296	−0.1104
		(0.1192)	(0.1207)		(0.0978)	(0.0989)
SG&A Ratiot−1		−0.05338	−0.06660		−0.03507	−0.01656
		(0.0398)	(0.0456)		(0.0233)	(0.0317)
TwoMonthLagReturn		−0.01840 *	−0.01942 *		−0.01097	−0.01321
		(0.0096)	(0.0099)		(0.0092)	(0.0097)
NewsVolumet−1			−0.000007982 *			−0.000007972 **
			(0.0000)			(0.0000)
Constant	0.2708 ***	0.3578 ***	0.3731 ***	0.1944 ***	0.2497 ***	0.2328 ***
	(0.0385)	(0.0510)	(0.0532)	(0.0335)	(0.0445)	(0.0477)
Observations	14,038	11,655	11,056	14,038	11,655	11,056
Adjusted R-squared	0.0204	0.0270	0.0292	0.0151	0.0210	0.0223
Number of Firms	913	775	737	913	775	737
Firm Fixed Effects	Yes	Yes	Yes	Yes	Yes	Yes
Time Fixed Effects	Yes	Yes	Yes	Yes	Yes	Yes

Notes: The above table reports the results of the regressions of 45 day and 30 day cumulative abnormal return on the lagged Search Index (SVI). Abnormal returns are calculated by subtracting the value-weighted market return from the stock return. For specifications 1 to 3, the CAR is calculated from day −45 to day −5, where day 0 is the day of earnings announcement. For specifications 1 to 3, the SVIChange is the difference between average search index from day −46 to −76 and the average search index from day −76 to −106, where day 0 is the day of earnings announcement. For specifications 4 to 6, the CAR is calculated from day −35 to day −5. For specifications 4 to 6, the SVIChange is the difference between average search index from day −36 to −66 and the average search index from day −66 to −96. NewsVolume is the number of news stories that mention the firm in the prior one-month time period. The following variables are based on the previous quarter’s numbers: Log Size is the natural log of sales. MarketToBook ratio is defined as the sum of total assets and market value of equity minus the book value of equity divided by total assets. Leverage is defined as long-term debt plus current liabilities divided by long-term debt plus current liabilities plus book value of common equity. ReturnOnAssets is income before extraordinary items divided by lagged assets value. SG&ARatio is the SG&A expense divided by sales. TwoMonthLagReturn is the return in the period from (−60, −120). All days are calendar days. Standard errors in parentheses are White standard errors clustered by firm. The sample period is from January 2004 to December 2008. ***, **, * denote 1%, 5% and 10% significance levels, respectively.

reports the result of the panel regression of CAR on the search index with firm- and time-fixed effects. A higher search index predicts a higher pre-earnings announcement drift. Since we are using firm-fixed effects, the independent variable is the change in the search index from its long-term mean. Hence, we are testing if, for a given firm, the change in the level of the search index can predict future abnormal returns. A one-unit change in the search index is associated with an increase of 0.83% in the 30-day CAR prior to the earnings announcement. Specifications 2 and 3 also control for common independent variables as well as lagged returns and news coverage.

It is possible that there is a third variable related to both searches and future returns that is responsible for the results. As Barber and Odean (2007) show that high past returns can lead to increased investor attention, and this may cause the search activity to rise in the future. Future returns can also be higher because of momentum generated by past returns. To control for such a scenario, we include the past two months’ returns in the regression. The results are robust to the inclusion of past returns up to the previous six months. It does not appear that the search index is being driven by past momentum in the returns.

The results are robust to the inclusion of the above variables and provide support for Hypothesis 3.

4.3. Trading Volume

In

Table 5. Buy volume prior to the earnings announcement.

	45 Days Prior to Earnings Announcement				30 Days Prior to Earnings Announcement
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
SVIt−1	0.01667 ***	0.01670 ***	0.4702 ***	0.4705 ***	0.01621 ***	0.01628 ***	0.02767 ***	0.02769 ***
	(0.0015)	(0.0015)	(0.0391)	(0.0391)	(0.0015)	(0.0015)	(0.0023)	(0.0023)
NewsVolumet−1	−0.00001039 **	−0.00001040 **	−0.0002347 *	−0.0002345 *	−0.00001221 ***	−0.00001241 ***	−0.00001250 **	−0.00001248 **
	(0.0000)	(0.0000)	(0.0001)	(0.0001)	(0.0000)	(0.0000)	(0.0000)	(0.0000)
LogSizet−1	−0.05418 ***	−0.05424 ***	−1.5122 ***	−1.5120 ***	−0.05476 ***	−0.05482 ***	−0.08896 ***	−0.08896 ***
	(0.0038)	(0.0039)	(0.1074)	(0.1074)	(0.0041)	(0.0041)	(0.0063)	(0.0063)
MarketToBookRatiot−1	0.002228	0.002229	0.07421 *	0.07353 *	0.002701 *	0.002726 *	0.004360 *	0.004320 *
	(0.0016)	(0.0016)	(0.0430)	(0.0430)	(0.0016)	(0.0016)	(0.0025)	(0.0025)
Leveraget−1	0.01215	0.01249	0.3097	0.3131	0.01237	0.01277	0.01821	0.01841
	(0.0091)	(0.0091)	(0.2491)	(0.2489)	(0.0091)	(0.0091)	(0.0147)	(0.0146)
ReturnOnAssetst−1	0.1368 ***	0.1366 ***	3.2656 **	3.2493 **	0.1384 **	0.1380 **	0.1923 **	0.1913 **
	(0.0500)	(0.0500)	(1.3892)	(1.3872)	(0.0539)	(0.0540)	(0.0817)	(0.0816)
SG&A Ratiot−1	−0.1017 ***	−0.1022 ***	−2.8441 ***	−2.8425 ***	−0.1061 ***	−0.1062 ***	−0.1674 ***	−0.1673 ***
	(0.0245)	(0.0244)	(0.6615)	(0.6608)	(0.0260)	(0.0260)	(0.0389)	(0.0389)
TwoMonthLagReturn	0.008646 ***	0.008474 ***	0.2951 ***	0.2934 ***	0.009025 ***	0.009068 ***	0.01736 ***	0.01726 ***
	(0.0031)	(0.0031)	(0.0855)	(0.0853)	(0.0031)	(0.0031)	(0.0050)	(0.0050)
Constant	0.8890 ***	0.8900 ***	25.292 ***	25.298 ***	0.8918 ***	0.8924 ***	1.4878 ***	1.4882 ***
	(0.0297)	(0.0297)	(0.8262)	(0.8256)	(0.0314)	(0.0316)	(0.0486)	(0.0486)
Observations	8023	8023	8022	8022	8022	8022	8021	8021
Adjusted R-squared	0.1593	0.1596	0.1797	0.1799	0.1307	0.1313	0.1797	0.1799
Number of Firms	561	561	561	561	561	561	561	561
Firm Fixed Effects	yes	yes	yes	yes	yes	yes	yes	yes

Notes: The above table reports the results of the regressions of buys volume ratios on the one-month-lagged search index (SVI). Daily buys and sells are calculated using data from the NYSE’s TAQ database. The trades for each stock are signed based on the Lee and Ready (1991) algorithm. For specifications 1 to 4, the buys and sells are considered from day −45 to day −5, where day 0 is the day of earnings announcement. For specifications 5 to 8, the buys and sells are calculated from day −35 to day −5. The dependent variable in specification (1) is defined as follows: (sum of daily buys from (−45 to −5)/{(sum of daily sells from (−45 to −5) + (sum of daily buys from (−45 to −5)}. For specification (2), the dependent variable is the (sum of dollar daily buys from (−45 to −5)/{(sum of dollar daily sells from day (−45 to −5)} + (sum of dollar daily buys from (−45 to −5)}. In specifications (3) and (4), we use the average of the buys ratio for each day to mitigate the effect of intense buying or selling days. Specifications 5 to 8 are similar to specifications 1 to 4; however, the time period is from (−35 to −5). Search volume index (SVI) is the average of the search index one month prior. NewsVolume is the number of news stories that mention the firm in the period one month prior. The following variables are based on the previous quarter’s numbers: Log Size is the natural log of sales. MarketToBook ratio is defined as the sum of total assets and market value of equity minus the book value of equity divided by total assets. Leverage is defined as long-term debt plus current liabilities divided by long-term debt plus current liabilities plus book value of common equity. ReturnOnAssets is income before extraordinary items divided by lagged assets value. SG&ARatio is the SG&A expense divided by sales. TwoMonthLagReturn is the return in the period from (−60, −120). All days are calendar days. Standard errors in parentheses are White standard errors clustered by firm. The sample period is from January 2004 to December 2008 and includes only the firms listed in NYSE for whom the TAQ data is available. ***, **, * denote 1%, 5% and 10% significance levels, respectively.

, the daily buys and sells volume data was used. Daily buys and sells are calculated using data from the NYSE’s TAQ database. The trades for each stock are signed based on the Lee and Ready (1991) algorithm. Daily buys and sells were considered for the period (−5 to −45) and (−5 to −30) calendar days before the earnings announcement. The five days prior to the earnings announcement were excluded to account for any insider trading activity prior to the earnings announcement. Sample size is reduced since we have the buys and sells data for only a subsample of S&P 1500 firms that are listed in the NYSE.

The dependent variable in the specification (1) is defined as follows: (sum of daily buys from (−45 to −5)/{(sum of daily sells from (−45 to −5) + (sum of daily buys from (−45 to −5)}. For specification (2), the dependent variable is the (sum of dollar daily buys from (−45 to −5)/{(sum of dollar daily sells from day (−45 to −5)} + (sum of dollar daily buys from (−45 to −5)}. It is possible that the results are being influenced by a heavy buying or selling day in the time period (−45 to −5). Since we are summing all the buy orders in the time period, the results may have been influenced by one or two intense buy volume days. In specifications (3) and (4), we use the average of the buys ratio for each day to mitigate the effects of intense buying or selling days.

For specification (3), the dependent variable is defined as the average of the daily buys ratio from day (−45 to −5). The daily buys ratio is defined as the ratio of daily buys divided by the sum of buys and sells for the day. The dependent variable for specification (4) is defined as the average of the dollar daily buys ratio from day (−45 to −5). The daily dollar buys ratio is defined as the ratio of dollar daily buys divided by the sum of dollar buys and dollar sells.

Table 5 shows that the increase in the prior month search index score (from day −46 to −76) is positively related to the subsequent increase in the buy trading volume. Given that we are employing firm-fixed effects, the independent variable is the deviation of the search index from its long-term mean. Consequently, we are examining whether, for a specific firm, the shift in the search index level can forecast future buying pressure. In each of the specifications, coefficients are significant and economically large. For robustness, an alternate window from day (−5 to −30) was also considered and the results remain robust to the alternate choice of the window and are shown in the specifications (5) to (8). These results provide additional evidence for the theory that heightened search activity correlates with investors’ purchasing behavior, thereby supporting Hypothesis 4.

5. Discussion

Synthesizing our results with the findings of Drake et al. (2012) yields a coherent picture of how investor search activity relates to price discovery around earnings announcements. Drake et al. (2012) established that Google search volume captures meaningful variation in investor information demand: search activity rises systematically before announcements, is associated with pre-announcement price movements that preempt earnings news, and—critically—this preemption effect is “inconsistent with the measure solely capturing the behavior of less sophisticated retail investors (or noise traders)” (p. 1002). Our study extends this framework by documenting three additional patterns: (1) search activity predicts the direction of pre-announcement drift, not merely its magnitude; (2) search is associated with directional order flow (buying pressure), providing a microstructure mechanism linking attention to prices; and (3) price effects are permanent, with no evidence of post-announcement reversals.

The findings affirm the initial hypothesis that anticipates a connection between search volume preceding earnings announcements and elevated abnormal returns in the days surrounding these releases. An earnings surprise may signal that in the preceding quarter, the company took noteworthy actions, such as introducing a new product or enhancing customer service, leading to profits surpassing market expectations. It is plausible that customers and employees directly experience these changes, prompting them to seek information about the company. The surge in the search index serves as a proxy for information-seeking behavior, aiding in the aggregation of information from diverse sources.

The second hypothesis, suggesting pronounced search impacts among lesser-known stocks, also finds support. Table 3 demonstrates a more robust predictive ability of search volume for small firms, where data environments are more restricted. This validates the theorized incentives to seek valuable insights in situations where readily available information is scarce (S. J. Grossman & Stiglitz, 1980). Smaller companies, typically characterized by greater information asymmetries, limited news coverage, a following of fewer analysts, and a minor share of institutional investors’ holdings, offer individual investors a better chance of earning fair returns for their research efforts.

Turning to the third hypothesis, search volume also forecasts pre-earnings announcement drifts. The results in Table 4 indicate that investor information gathering precedes building price moves before scheduled releases, reinforcing inferences around informed trading based on discovered knowledge. Additionally, there is a discernible link between search volume and buying pressure on the stock. This implies that, in the presence of short-selling constraints, investors’ information-seeking behavior exerts buying pressure on stocks.

Several key findings underscore predicted mechanisms linking investor information-seeking behavior to prices. Search volume is associated with future trading activity and permanent price changes (Kaniel et al., 2008; Loh, 2010), reflecting informed trading. These effects are more pronounced around earnings announcements when information asymmetry is high (Beaver, 1968; Beyer et al., 2010). Search volume is associated with post-announcement drifts and market reaction to earnings announcements (Bernard & Thomas, 1989; Gleason & Lee, 2003), especially among lesser-known firms where data is limited (Hong et al., 2000). Together, these findings provide large-sample micro-level evidence illustrating how prices respond to knowledge investors uncover through search and trading (Drake et al., 2012).

The results contribute to advancing the understanding of the real-world drivers of market efficiency. Furthermore, they underscore how greater information transparency and access may facilitate efficient pricing (Tetlock, 2010; Blankespoor et al., 2017). This highlights the genuine economic value potentially generated by platforms that broaden data access.

6. Limitations and Further Research

One limitation of this paper is the reliability of using Google search activity as a proxy for investor information acquisition. While an innovative data source, search volume may reflect other confounding factors like attention or sentiment. Having a better proxy for investor information-seeking behavior could improve the paper. Additionally, the study focuses on S&P 1500 firms; expanding the sample to include other domestic and international firms can help establish the results in a wider scope. The current paper analyzes only one corporate event—earnings announcements. Future research can look at other corporate events such as mergers and acquisitions or dividend announcements.

While we acknowledge the inherent challenges of establishing causality in observational studies of investor behavior, we note that the pattern of results documented both in our study and in Drake et al. (2012) is difficult to reconcile with simple reverse-causality explanations. If search activity were merely a response to price movements or public information arrival, searches would not systematically preempt earnings news content. The preemption effect documented by Drake et al. (2012)—whereby high search activity is associated with reduced announcement reactions because the information has already had time to impact the prices—implies that search contains forward-looking information about earnings. Our temporal structure further mitigates reverse-causality concerns: the search is measured from days −66 to −35, creating substantial separation from the announcement-period returns we examine. The combination of (1) temporal precedence, (2) preemption effects consistent with informed trading, (3) controls for contemporaneous news and momentum, and (4) the absence of post-announcement reversals collectively provides stronger identification than any single test.

A potential concern with our 2004–2008 sample period is that the 2008 financial crisis could confound results if crisis-induced volatility or behavioral changes disproportionately drive the documented relationships. We address this concern through our year-quarter time-fixed effects specification, which absorbs aggregate crisis shocks affecting all firms in a given quarter. Additionally, Da et al. (2011) explicitly conducted subperiod robustness tests using the same sample period, reporting that “the regression results are qualitatively similar in the two subsample periods” (Table VII, p. 1483). This precedent from the foundational paper in the literature provides confidence that our findings are not artifacts of crisis-period dynamics.

Regarding sample scope, our focus on S&P 1500 firms provides economically meaningful coverage—representing approximately 90% of U.S. market capitalization—but may limit generalizability to smaller stocks where search effects could be even stronger. Theory predicts information acquisition is most valuable for information-scarce firms, and our documented small-firm effects (Table 3) suggest our estimates may represent conservative lower bounds. Extending the analysis to micro-cap stocks and international markets remains an avenue for future research.

While these limitations require prudent interpretations, this large-scale analysis tracing search measures to market pricing substantially advances the comprehension of asset price formation at the nexus of information gathering, trading, and disclosure efficiency. Both documented patterns and outlined caveats highlight intellectually fertile opportunities for follow-on inquiry analyzing market mechanisms through creative use of emerging data sources.

7. Conclusion

This paper uses Google search activity as a unique proxy for investor demand for information. Search volume specifically spikes prior to earnings announcements as investors seek new insights to prepare for market-moving events. The rise in search activity is positively associated with eventual market reactions upon earnings releases, especially for lesser-known stocks where incremental information carries more value due to higher information asymmetry.

The study promotes the understanding of mechanics of how new information actually reaches and impacts financial markets. The work expands the boundaries of existing investor theory focused on how heterogeneous individuals gather and leverage insights. From a practical perspective, the research also imparts valuable guidance for regulators and firms on how greater transparency facilitates efficient markets by allowing for dispersed information aggregation through trades. Overall, this work constitutes a valuable addition to the literature on the underlying informational efficiency of capital markets.