Internet Search Intensity and Its Relation with Trading Activity and Stock Returns*

I INTRODUCTION

The efficient market hypothesis assumes that stock prices should instantaneously reflect new information as it arrives (Fama 1970). This assumption requires investors to be aware of the full set of information available in the market. However, psychological evidence suggests that investors can only handle a certain amount of information at any point in time and thus, as Kahneman (1973) has argued, attention is a scarce cognitive resource. The study by Barber and Odean (2008) establishes the theory in investor attention. They argue that investors face a search problem when selecting which stocks to buy. However, the search problem is not as formidable when investors sell stocks, since they can only sell the stocks that they already own.¹ Thus, when facing a search problem, investors are more likely to buy rather than sell stocks that attract their attention. This prediction leads to short-term positive price pressure at an attention-grabbing event and this price pressure dissipates as attention wears off over time.

The purpose of this paper is to examine the role attention plays in financial markets. We first hypothesize that attention helps explain the cross-sectional variation in trading activity. Chordia et al. (2007) argue that trading activity is a function of liquidity trading, the number of informed agents, the degree of learning by investors on how asset returns are generated, and the dispersion of agents’ information signals. The authors further point out that investors’ liquidity needs are realized only in stocks that are more visible to investors. Chordia et al. (2007) find that proxies for visibility such as size, firm age, the book-to-market ratio, and price levels are important determinants of cross-sectional variation in trading activity. Since attention is a necessary condition for a firm to be publicly recognized, this should lead to a positive relation between attention and trading activity in the cross section when combined with the view that the buying behavior of investors is influenced by attention-grabbing events.

In classical finance theories with a no-arbitrage argument, profitable opportunities from trading securities are exploited by arbitragers and their arbitrage activities should correct mispriced securities. This argument is built on the assumption that information is readily available to investors. However, arbitrage opportunities could become less obvious when information is sparse or even ignored (e.g., Zhang 2006). Trade barriers such as transaction costs and short-sale constraints could further hinder arbitrage activities. If the attention effect is due to ignorance of price relevant information or to constraints in allocating attention, the effect should be more pronounced in the presence of severe limits to arbitrage. This is because investors who have a greater capacity to obtain information and exploit mispricing will be more likely to avoid stocks where arbitrage is difficult.

This paper employs monthly share price data drawn from the Australian market from 2004 to 2015. Our proxy for investor attention is the Google search volume index (SVI) following Da et al. (2011). The use of the Australian market is motivated by two facts. First, how attention can be directly captured remains an open question and many alternative proxies have been employed in empirical analysis. With the development and convenience of the Internet, online resources are rapidly becoming one of the most important information sources for investors. However, compared to other alternative proxies, the use of Internet search as a proxy for attention is fairly limited in markets outside the United States.² In Australia, Google is the dominant search engine, with over 90% of all searches run, greater than the 72% reported in the US market.³ The second most popular search engine in Australia is Bing, but it comprises only about 4% of the market. Thus, Google search engine is arguably an important information source for Australian investors. Second, the Australian market is of interest in its own right. As noted in Chiah et al. (2016), about 70% of the listed stocks on the Australian Securities Exchange (ASX) are priced below AU$1. Thus, the distribution of share prices is highly skewed and companies with low share prices are common in this market. Dyl and Elliott (2006) point out that small and less well-known firms tend to have low share prices in order to broaden their investor base. Accordingly, attention is expected to play a nontrivial role in this market.

We find that abnormal Google search volume (ASVI), which is defined as the difference between the current SVI and the mean SVI over the past 12 months, scaled by the standard deviation of the SVI over the past 12 months, is positively related to turnover, order imbalance (OIB) and liquidity in the cross section. These relationships remain robust when other important determinants of trading activity are included in the analysis. The positive relation between ASVI and the OIB between buy and sell orders is consistent with the view that attention attracts more buying than selling. We also find that high investor attention shocks lead to an increase in trading volume, OIB and stock liquidity in the subsequent month.

Consistent with the literature, the existence of the attention effect is unambiguous. At the portfolio level, the return difference between stocks experiencing a high ASVI and stocks experiencing a low ASVI is positive in the short run and this return difference falls considerably as the holding period is extended. This return pattern is also confirmed at the individual stock level after controlling for other firm characteristics, which are known to affect stock returns. Our results are robust to a number of sensitivity checks, including (i) employing different search terms, (ii) examining whether the relation between ASVI and stock returns is caused by price-sensitive announcements, (iii) considering seasonality in returns, (iv) testing the existence of the attention effect after controlling for turnover and liquidity shocks, and (v) using weekly rather than monthly stock return data. Our results provide strong out-of-sample evidence on the relation between search intensity and stock returns. The explanatory power of ASVI on trading activity and stock returns is consistent with the theoretical prediction of Barber and Odean (2008) that the proportion of stocks being purchased relative to stocks being sold will be greatest following periods of high attention.

Having confirmed the existence of the attention effect, we formally test how limits to arbitrage contribute to the effect. An important methodological consideration is how limits to arbitrage are measured. The most common measure of arbitrage risk is idiosyncratic stock volatility (e.g., Pontiff 1996; Wurgler and Zhuravskaya 2002). The other common measure is information uncertainty, which is often proxied by analyst coverage (Hong et al. 2000) and dispersion in earnings forecasts (e.g., Diether et al. 2002; Zhang 2006). Trading costs are also linked to limits to arbitrage and proxies include stock prices, firm size, and trading volume (e.g., Bhushan 1994; Stoll 2000; Nagel 2005). In this paper, we construct a limits-to-arbitrage index based on five popular limits-to-arbitrage measures following a recent methodological innovation by Stambaugh et al. (2015). The measures include idiosyncratic volatility, analyst coverage, share price, firm size, and share turnover. Our tests involve an examination of portfolios double-sorted on ASVI and the limits-to-arbitrage index and cross-sectional regressions performed for stocks with different levels of arbitrage risk. Consistent with our conjecture, we find that the attention effect is most pronounced in stocks with high arbitrage costs. Specifically, these stocks exhibit stronger short-run price pressure and subsequent reversal.

Our paper contributes to the literature in two major ways. First, an investigation of the attention effect in Australian equity returns not only provides a comparison with other markets, but also establishes evidence of common components of attention in stock returns through similar findings. Our results suggest that Google search volumes contain useful information in generating asset returns. Second, the paper contributes to the literature on investor attention in asset pricing dynamics by introducing the role limits to arbitrage play. In particular, we show that attention shocks lead to strong stock price reactions, especially in those stocks that are informationally constrained.

The remainder of this study is organized as follows. Section II describes the data and the research design. Section III reports the findings on the relation between attention and trading activity. Section IV examines the relation between attention and stock returns. Section V explores the role limits to arbitrage play in explaining the attention effect. Finally, Section VI concludes the paper.

II DATA AND METHODS

B Attention shocks and trading activity

This paper measures abnormal change in the SVI as a proxy for an attention shock in the spirit of Bali et al. (2014) and Chen et al. (2001):

(1)

where AVGSVI _{i∣t − 12,t − 1} and SDSVI _{i∣t − 12,t − 1} are the mean and standard deviation of the SVI for stock i over the past 12 months, respectively. A stock with an increase (decrease) in ASVI relative to its past 12-month average is regarded as having more (less) attention.⁷

We run Fama and MacBeth (1973) regressions to test the relation between ASVI and trading activity. The model is specified as follows:

(2)

Chordia et al. (2007) show that trading activity measures such as turnover and OIB depend on the degree of liquidity trading, informed trading, and the level of uncertainty and disagreement about fundamental values. Accordingly, we examine the extent to which ASVI explains cross-sectional variation in turnover and OIB. Moreover, when individual investors acquire useful information relevant to a stock via Google searches, the information asymmetry problem for the stock is mitigated. Noise trading may also increase following retail investor attention. As a result, liquidity should improve for stocks with better attention. In fact, Ding and Hou (2015) find a positive relation between investor attention and stock liquidity for similar reasons. We therefore include liquidity as a trading activity measure.

Turnover is monthly trading volume scaled by the number of shares outstanding. OIB in a month is the average of daily OIB defined as the difference between total buy orders and total sell orders divided by total daily orders. Following Barber and Odean (2008), OIB is calculated using value rather than number of trades. The order flow data is not available for all stocks in our sample (about half of our sample) and the data starts in 2006. Analyses related to OIB are thus limited to a subset of stocks with available data. The Amihud (2002) illiquidity ratio is used as a proxy for liquidity, and it is defined as the average ratio of daily absolute stock returns to daily trading volume in a month. Since the Amihud ratio is highly right skewed and a higher value corresponds to a lower level of liquidity, following Edmans et al. (2013), we log-transform the measure first and then, for ease of interpretation, multiply the measure by −1. Since trading activity variables exhibit high autocorrelations, we adjust the variables following the method outlined in Chordia et al. (2007) to eliminate nonstationarity.

According to Chordia et al. (2007), trading activity can be explained by stock visibility, portfolio rebalancing needs, differences of opinion, and uncertainty about firm fundamental values. The control variables in equation () therefore follow those used by Chordia et al. (2007). We use firm size, age, the book-to-market ratio, and stock price to measure stock visibility. Age is the number of listing months since the initial public offering. For rebalancing needs, we separate the returns of the past month into POS and NEG, where POS (NEG) is defined as the monthly return of an individual stock if it is positive (negative) and zero otherwise. Finally, we use leverage to measure differences of opinion and beta to measure uncertainty about firm fundamental values. Leverage is computed by using the book value of debt divided by total assets. Beta is estimated using the market model over the past 252 trading days. Following Chordia et al. (2007), both size and price are logged and adjusted for nonstationarity.

Since high trading activity could be due to overreaction from investors, we also include continuing overreaction (CO) as a control variable, where CO is the weighted average of daily signed volumes within a month and is computed in the spirit of Byun et al. (2016):

(3)

where VOL _i,d,t is the trading volume for stock i on day d in month t,D is the number of trading days in month t, w _i,d,t is a weight that takes a value of d (dth trading day of month t), and SV is the signed volume. The signed volume in day d for stock i is defined as

(4)

where r _i,d,t is the stock return in day d for stock i in month t.

To strengthen the regression analysis, we use portfolio sorts to further test the relation between ASVI and trading activity. In each month over the sample period, stocks are first sorted by their ASVI into 10 portfolios. Decile 1 contains stocks with the lowest ASVI values and decile 10 contains stocks with the highest. The average trading activity is then calculated for each decile over the subsequent months. This approach reveals a simple picture of how trading activity varies across the spectrum of attention shocks.

C Asset pricing tests

To test the relation between ASVI and stock returns, we first establish the evidence using portfolio sorts. Specifically, at the end of each month, stocks are ranked and sorted into deciles according to their ASVI. Equal-weighted buy-and-hold portfolio returns over the next 1, 3, 6, and 12 months are calculated for each decile portfolio. The return difference between the highest- and lowest-ASVI portfolios represents the return premium associated with investor attention. The portfolios are updated monthly. Risk adjustment is carried out using the Fama and French (2015) five-factor model:

(5)

where R _p,t is the return for portfolio p in month t; R _f,t is the risk-free rate in month t; MKT is the market excess return; SMB, HML, PMU, and LMH are the factor-mimicking portfolios for size, the book-to-market ratio, profitability, and investments, respectively. The factors are constructed following the procedure outlined by Chiah et al. (2016).

The portfolio-level analysis establishes preliminary evidence on the relation between ASVI and stock returns. We then analyze the relation at the individual stock level, running the following cross-sectional regressions each month:

(6)

where the dependent variable is the return over the subsequent one, two to six, and seven to 12 months after month t. Controls include variables that are potential determinants of stock returns, including size, the book-to-market ratio, momentum, idiosyncratic volatility, beta, the demand for extreme returns (MAX), short-term reversals, and the overreaction measure CO. Following Bali et al. (2011), MAX for a stock is defined as its maximum daily return in a month. The monthly idiosyncratic volatility of stock i is calculated as the standard deviation of the daily residuals from the Fama and French (1993) three-factor model:

(7)

where R _i,t is the return for stock i and R _f,t is the risk-free rate at month t, MKT is the value-weighted market excess return, and SMB and HML are the size and book-to-market factors, respectively. Following Ang et al. (2006), we estimate idiosyncratic volatility only for stocks that traded at least 65 days over the past 252 trading days.

III ATTENTION AND CROSS-SECTIONAL VARIATIONS IN TRADING ACTIVITY

Table 2 presents the regression results. Panels A and B report the time-series average of the monthly cross-sectional regression estimates using turnover, OIB and liquidity as the measure of trading activity, respectively. For the baseline model without the control variables (regression (1)), the coefficients of ASVI are all positive and significant at the 1% level. This finding suggests a significant positive relation between trading activity and ASVI. Regressions (2) and (3) include the control variables. Given the high correlation between size and share price, we separately include size (regression (2)) and share price (regression (3)) in the analysis. Across the trading activity measures, the impact of the control variables seems rather mixed. For example, when liquidity is used as the dependent variable, the coefficients of the book-to-market ratio and leverage are negative. Past negative returns also exhibit somewhat different patterns in predicting trading activity variables. This may be because these variables capture different aspects of trading activity. Nevertheless, the main takeaway in Table 2 is that the relation between ASVI and trading activity is robust after we control for important determinants of trading activity.

To further explore the relation between ASVI and trading activity, we form decile portfolios on ASVI at time t and calculate the trading activity for each portfolio at times t, t + 1, and t + 2, respectively. Showing trading activities at different time periods allows us to investigate how they change over time. Table 3 presents the results. At time t, from the low-ASVI group to the high-ASVI group, trading activity exhibits an overall increasing trend. A positive relation between ASVI and OIB is also consistent with the view that attention attracts more buying than selling. The differences between the high- and low-ASVI groups are positive and significant at the 1% level for all the trading activity variables. At time t + 1, the increasing pattern in trading activity becomes more obvious. For example, turnover (liquidity) increases from 0.0417 (−1.4933) in the low-ASVI group to 0.0608 (−1.1537) in the high-ASVI group. This finding implies that stocks with a higher ASVI are traded more than stocks with a lower ASVI. The result from OIB is less apparent but the difference remains statistically significant. The difference in trading activity between the low- and high-ASVI groups continues to be significant at t + 2 except for OIB. However, compared to times t and t + 1, the trend is less obvious. As the OIB related results are based on the sample of stocks with available order flow data, the finding may be influenced by firm size.

IV ATTENTION AND STOCK RETURNS

V ATTENTION AND LIMITS TO ARBITRAGE

Our results thus far demonstrate a robust relation between ASVI and stock returns. In this section, we test the role limits to arbitrage play in the ASVI–return relation. Recently, Stambaugh et al. (2015) built a simple proxy for mispricing based on a composite rank of numerous firm characteristics known to be associated with anomalous returns. Motivated by their work and the fact that there are a few proxies for limits to arbitrage, we construct a limits-to-arbitrage index for each firm in our sample.

Our limits-to-arbitrage index is constructed based on five proxies for arbitrage costs, including idiosyncratic volatility, analyst coverage, share price, firm size, and share turnover. Among existing proxies for limits to arbitrage, idiosyncratic volatility is the most popular measure. For example, Pontiff (2006) shows that arbitrageurs prefer to hold fewer stocks with higher idiosyncratic volatility. McLean (2010), using idiosyncratic volatility as a proxy for arbitrage costs, shows that lower arbitrage costs are associated with a weaker (negative) relation between short interest and subsequent stock returns. Analyst coverage is used as a proxy for information uncertainty. More analyst coverage indicates a stronger information environment (Hong et al. 2000) and will mitigate mispricing (e.g., Zhang 2006). Share price, firm size, and share turnover are related to potential transaction costs. Share price has been found to be inversely related to the bid–ask spread and brokerage costs (Bhardwaj and Brooks 1992; Stoll 2000). Small firms are generally less liquid and have high short-sale constraints. Share turnover is a common measure for stock liquidity, capturing the time required to execute an order or to trade a large block of shares.

The limits-to-arbitrage index is constructed as follows. Each month, stocks are assigned a percentile rank based on one of the limits-to-arbitrage measures. For example, stocks with the highest (lowest) idiosyncratic volatility receive the highest (lowest) rank. In other words, stocks with the highest arbitrage costs receive the highest rank, while stocks with the lowest arbitrage costs receive the lowest rank. An exception is made with analyst coverage, since the variation across firms is low. We thus assign the highest rank (i.e., 100) to stocks without analyst coverage and the lowest rank (i.e., zero) to stocks with the highest number of analysts. The interval in ranking depends on the highest number for analyst coverage. The limits-to-arbitrage index for each stock is the simple average of the ranks across the five measures. By construction, a high index value indicates a high level of arbitrage cost.

To explore the interaction between limits to arbitrage and attention, stocks are independently double sorted into quintiles based on the constructed limits-to-arbitrage index and ASVI, resulting in 25 portfolios. The equal-weighted returns to each portfolio are calculated for the following month, after which the double sorting is repeated. The results are reported in Table 8. There are several notable findings. First, reading the results across the ASVI quintiles, there is no attention effect in stocks with the lowest arbitrage costs. However, in the rest of the limits-to-arbitrage groups, returns increase monotonically across the ASVI quintiles and the high-minus-low returns are statistically significant. The high-minus-low return also increases monotonically from the lowest to the highest limits-to-arbitrage index quintiles. This result is consistent with our conjecture that the attention effect is more pronounced in stocks with high arbitrage costs. Second, reading each column, the return difference between high- and low-arbitrage quintiles is only significant in extreme ASVI groups. Interestingly, the return is negative (albeit only marginally significant) in the lowest-ASVI quintile and is positive in the two high-ASVI quintiles. The magnitude of the return difference also increases with ASVI.

A stronger attention effect in stocks with high limits to arbitrage should result in higher short-run price pressure and subsequent reversal. To confirm this, we examined the portfolio returns in the limits-to-arbitrage groups over longer holding periods of 3, 6, and 12 months. Table 9 displays the results. To conserve space, we only show ASVI quintiles in both high and low limits-to-arbitrage groups. Reversal is observed in the high limits-to-arbitrage group. For the low limits-to-arbitrage group where the 1-month ahead hedge return is not significant there is also no reversal. This confirms that the short-run price pressure and subsequent reversal is concentrated in stocks where arbitrage is more difficult. Overall, the results confirm that attention is a scarce cognitive resource for stocks with high limits to arbitrage.

To further strengthen our findings, we run equation () in different limits-to-arbitrage groups. Specifically, each month, stocks are first sorted into three portfolios using a 30:40:30 split based on the limits-to-arbitrage index. Within each group, different specifications of equation () are estimated. Table 10 presents the results. The baseline model (regression (1)) shows that ASVI is statistically significant in all three limits-to-arbitrage groups. However, the magnitude of the ASVI coefficient is much larger in the high limits-to-arbitrage group. The results remain consistent in the presence of control variables. It is worth mentioning that we include both idiosyncratic volatility and size, which are also used to construct our limits-to-arbitrage index, in the regressions. The purpose is to investigate whether the ASVI–return relation is altered by these variables in different limits-to-arbitrage groups. Our untabulated results show that the difference in the ASVI coefficients between the high and low limits-to-arbitrage groups is 0.0074 (t = 6.92), 0.0075 (t = 7.00), 0.0073 (t = 6.92), and 0.0073 (t = 6.86) for regressions (1) to (4), respectively. These results show strong support for our conjecture that the attention effect is most pronounced in stocks with high limits to arbitrage.

VI CONCLUSION

Limited attention is a consequence of limited information processing power on the vast amount of information available in financial markets. This paper contributes to the growing literature on theories and evidence of limited attention and information processing and their implications in asset pricing. In the age of information technology, obtaining information from the Internet is easy and cheap for investors. Motivated by this reasoning, we use hand-collected Google search volume as a proxy for attention. The findings of this paper provide strong out-of-sample evidence to the attention literature.

We find that ASVI helps explain the cross-sectional variation in trading activity. An increase in ASVI leads to higher turnover, a greater OIB between buy and sell orders, and increased liquidity. Consistent with attention theory, there is a positive relation between ASVI and stock returns over a short investment horizon and then reversal over a longer period. Our results are robust to a number of sensitivity checks. Furthermore, we document that the attention effect is more pronounced in stocks with high arbitrage costs. Our empirical findings provide new insights on the cause of the attention effect.

Overall, the findings of this paper are consistent with the view that information technology has shaped the operations of financial markets. The open and public nature of the Internet facilitates information sharing and has become an important source in obtaining information. For practitioners, Google keyword search volume can be used to predict macroeconomic and company trends, which could be applied to developing advanced investment strategies.