Predicting daily stock returns using retail investors’ attention
Rick Gerritsen
1Master thesis Finance
Supervisor: Dr. M.M. Kramer
July 11, 2014
Abstract
This paper studies the predictive power of the search volume on European stock returns. The number of searches on the ticker of a stock is used as a measure of retail investors’ attention. Besides finding positive abnormal returns after an abnormal increase of the search volume on weekly frequency data, the same result is found on daily data. Companies with a relative low market value, a low dividend yield or a high P/E ratio are more susceptible for investors’ attention.
JEL classification: G14 – G17
Keywords: Google search volume – Predicting stock returns – investors’ attention
2
1. Introduction
The efficient market theory (Fama, 1970) leaves no room for the scarce cognitive resources (Kahneman, 1973) investors face when studying investment opportunities. In this paper the effect of investors’ attention on stock price returns is studied. In particular the search frequency in Google on a company’s ticker is used as a direct measure of attention towards a stock. The paper of Da, Engelberg and Gao (2011) shows that the search frequency in Google is likely to capture the attention of retail investors. Considering the attention hypotheses of Barber and Odean (2008), an upward stock price movement is expected when the attention from retail investors towards a stock is increased.
That the attention of especially retail investors is measured using the search frequency in Google is not surprising. Retail investors typically apply heuristics to overcome cognitive constraints like limited memory, limited attention and limited processing power (Nofsinger, 2010). The use of a search engine fits in this description; it is easy to use, it doesn’t take much time to look at the results and the results are increasingly of higher quality (Brin and Page, 1998). Institutional investors have more resources at their disposal to limit their cognitive constraint in terms of finance and time. They don’t apply a shortcut using a search engine to collect information, but can use financial data terminals, like a Bloomberg terminal or Thomson Reuters, to collect all information necessary to base their investment decision on. The use of these terminals are too expensive for the average individual investor. This paper’s thesis is that the process of using a search engine captures the attention of retail investors by looking at the specific keyword the retail investor uses to search for stock information.
3 Figure 1 depicts the search volume index on "Vakantiegeld" for the period January 2004 until May 2014.
Recent research on capturing the attention of investors (Barber and Odean, 2008) uses three criteria to capture the attention of the investors for a stock i.e. stocks with high abnormal returns, stocks with high abnormal trading volume and stocks covered in the news (Dow Jones news feed on the ticker of the company). However, these measures of attention are all based on the assumption that investors do actually pay attention on these events. Da et al. (2011) qualify these measures of attention as indirect measures, Barber and Odean (2008) do show however that retail investors are displaying attention-driven investment choices based on these three indirect measures of attention.
Da et al. (2011) provided a prominent paper which is elaborately used throughout this whole paper. Their theoretical framework and research techniques to create a model where investors’ attention is studied, will be followed there where possible. Da et al. (2011) find that the search frequency captures the attention of retail investors towards an individual stock. This measure of attention captures other attention than the measures of attention proposed by Barber and Odean (2008).
The main contribution of this paper is the use of data on a daily frequency, there where Da et al. (2011) use data on a weekly frequency. Measuring the predicting power of the daily search frequency on daily stock returns could, in potential, be more useful compared to weekly search frequency. It would enable investors which use an investment strategy based on the changes in the search frequency on a daily basis, to act in a shorter timeframe compared to investors which strategy is based on weekly frequency search volume. The second contribution is the use of European stock market data instead of United States stock market data Da et al. (2011) use.
4 In my analyses I find that an increase in the search volume on the ticker of a company predicts positive abnormal returns on both a weekly sample and a daily sample. The sample consists of stocks included in the S&P 350 Europe during June 2011 until December 2013 for the weekly data sample and between January 2014 until March 2014 for the daily data sample. These results increase the robustness of the findings of Da et al. (2011) that search volume captures investors’ attention. I find that the predictive power of the search volume is highest for stocks which have a high propensity to speculate.
5
2. Literature review
Although the internet has been around for decennia, behavioural research using individuals search frequency or micro blogs (e.g. Twitter) is relatively scarce. This section of the paper will discuss a selection of papers which have used the search frequency to capture attention. However, at first the current state of the literature concerning investor attention is address with a selection of papers. The paper of Da et al. (2011) is discussed in more detail because of the importance of their findings and research methodology for this paper.
2.1 Investor attention
One of the underlying assumptions of the efficient market hypothesis is the rationality of the actors participating in the market, the agents. The rationality of the agents will make sure that the valuation of a company will be set as such that it always converges to the fundamental value (Fama, 1970). Even when not all agents are making ration valuations, a deviation from the fundamental will be offset by arbitrageurs. Arbitrage is a riskless and costless investment strategy where the only outcome is a profit (Shleifer and Vishny, 1997). Shleifer and Summers (1990) do argue that such a clean arbitrage opportunities doesn’t exist, the stock price can diverge further from the fundamental value because of noisy traders or market sentiment.
Barber and Odean (2008) find that retail investors’ are showing attention driven buying behaviour. They argue that the retail investor has in most cases only one active decision he can make: buying. The retail investor hardly ever is in the position to sell after his attention is drawn to a stock, because he only holds a limited amount of stocks in his portfolio which he is able to sell. In the dataset of Barber and Odean (2008) households only have a mean short position of 0.29 percent. Institutional investors won’t face this constraint, their portfolio will consist of more stocks and short selling is more prominent. In the paper of Barber and Odean (2008) absolute abnormal returns, extreme trading volume and news headline in the Dow Jones news feed are used as measures of attention. Chemmanur and Yan (2009) add advertising expenses divided by sales as a measure of retail investor’ attention.
6 period sentiment. When the sentiment is high at the beginning of period, the returns on speculative stocks are relatively low. The reverse effect is observed when sentiment measures are relatively low. Baker and Wugler (2006) specify speculative stocks as stocks which are, young, small, experience extreme growth and distressed stocks as hard to arbitrage which makes them speculative.
2.2 Search volume as a measure
The use of search volume as an indicator for attention of individuals received a large amount of attention when Ginsberg et al. (2009) published a paper where they claimed it was possible to detect influenza activity earlier using the Search Volume Index (SVI ) than conventional measures. Using the SVI of the 50 million most common search queries as input, Ginsberg et al. (2009) looked for the search query which yielded the best fit with influenza activity reported by the CDC. Out of these 50 million most common search queries, Ginsberg et al. (2009) use the 45 best fitting search queries to measure influenza activity. This paper seems to have used a more extensive database compared to Google Trends, this research method can’t be performed without this database.
Choi and Varion (2009, 2012) create a guide how to predict the present with the search volume index (SVI) of Google by showing examples for economic indicators. Their research method make it possible to create economic indicators which lead official numbers from government agencies. Choi and Varion (2009, 2012) show an analyses of four different indicators, motor vehicles and parts, initial claims for unemployment benefit, travel to Hong Kong and consumer confidence. Not using one keyword as the search frequency, but one Google Trends category’s total search frequency is used to create a best fitting model. Choi and Varion (2009, 2012) don’t use an economical argument for using a category.
7 Da et al. (2011) studied the use of the SVI of Google as a measurement of attention on a fundamental level. As the keyword to capture the attention of investors they use the ticker of a company, whenever a Google search is performed on the ticker of the company this is included in the search frequency. Da et al. (2011) argue that whenever a Google search is performed on the ticker of the company, the person which makes this search is truly paying attention to the stock of the company. When using the name of the company, instead of the ticker, this will cause many noise in the measure of attention. Searching on the name of a company doesn’t necessarily implicate that an investor is interested in the stock of the company, it could also mean that the individual is interested in the company’s products, career possibilities or other information not directly linked to investing decisions (Da et al., 2011).
The first result in the paper of Da et al. (2011) is that the SVI is leading the measures of attention proposed by Barber and Odean (2008) significantly by a week. After establishing the lead advantage the SVI has, the predicting power of the SVI is studied. Controlled for several other measures of attention, Da et al. (2011) find significant positive abnormal returns the first 2 weeks after an abnormal positive change in the SVI this is followed up by a price reversal in the year after. This result is in line with attention driven buying behaviour of Barber and Odean (2008). Da et al. (2011) add Dash-5 reports, a measure of retail investors trading, in their regression. They find that the SVI likely measures the impact of retail investors. The significant negative coefficient of the interaction variable market capitalization with the Abnormal Search Volume Index (ASVI, suggests that especially companies with a relative small market capitalization are more prone to be susceptible for changes in SVI, this is line with the work of Baker and Wurgler (2006).
8
3. Methodology and data
This section of the paper discusses the methodology and data sample used to study investors’ attention. At first the hypotheses are formulated which will be answered in this paper, secondly the variables used to study investors’ attention are defined and the descriptive statistics of these variables are presented. The last part of this section presents the methodology.
3.1 Hypotheses
At first I study how the SVI is related to other measures to capture the attention of investors. Da et al. (2011) find that the SVI leads the measures of attention absolute abnormal returns and abnormal trading volume. For the data sample used in this paper the same analyzes are performed to test the findings of Da et al. (2011) and analyze the relation on daily frequency data.
H1 : The search volume index leads the measures of attention absolute abnormal returns and
abnormal trading volume.
After establishing the value of the SVI as a measure of investors’ attention compared to other measures of attention, the impact of the SVI on abnormal returns will be studied. Where Da et al. (2011) performed this analyses on data with a weekly frequency, I will study whether the same results hold for daily frequency data. Though, in order the test the robustness of Da et al. (2011) results I test the predictive power on weekly frequency data first.
H2: The weekly search volume index has a predictive power on future abnormal weekly returns
To the best of my knowledge an analysis with daily frequency SVI data hasn’t been studied in other papers, however, if the SVI has a significant impact on future returns on a weekly basis, this result should also be observed on a daily basis. The following hypotheses is formulated to study this predictive power:
H3 : The daily search volume index has a predictive power on future abnormal daily stock returns.
Da et al. (2011) find that especially the stock returns of companies with a relatively small market capitalization can be partially predicted with investor attention. Research by Baker and Wurgler (2006) indicate that investor sentiment is especially of influence on companies which are harder to value. To study the sensitivity of the results across different characteristics of companies, the data sample will be partitioned on market value, price/earnings ratio and dividend yield to study the different impact of the search frequency on firm characteristics.
H4a: Stocks with a market value ratio are more susceptible for changes in the SVI.
H4b: Stocks with a higher price/earnings ratio are more susceptible for changes in the SVI.
9
3.2 Data sample construction
Table 1 presents an overview of the variables used in this study. The first panel shows the variables that capture the search frequency, during this study the search volume index (SVI) from google is used. Google Search has a market share hovering around 70 percent of the worldwide searches, this makes it the most attractive and representative database to use. The search frequency in Google is “broad matched” when it is collected. This means that when the term INGA is within a search query, like [INGA stock price] this is included in the SVI of INGA (Choi and Varian, 2012).
As discussed in the literature review, Da et al. (2011) use the ticker of a company to collect the SVI from, because a search on the ticker of the company is clearly a sign of attention towards the stock. The same reasoning is applied in this paper to use the ticker of the company, also the SVI on the name of the company is included as a robustness test to test that the SVI on the ticker is indeed a better measure of attention compared to the SVI on the name of the company.
Table 1: Definition of variables
Variable Definition
Panel A: Variables collected from Google Trends. All variables are collected on both a weekly frequency as a daily frequency.
SVI_ticker The search volume index which contains the aggregated search frequency based on the ticker.
SVI_name The search volume index which contains the aggregated search
frequency based on the name.
ASVI_ticker The log of the search volume index on the ticker minus the log of the median of the search volume from the past 8 observations. ASVI_name The log of the search volume index on the name minus the log of
the median of the search volume from the past 8 observations.
Panel B: Variables related to investors' attention. All variables are collected on both a weekly frequency as a daily frequency.
Stock return The return of the stock
Abnormal return The stock return adjusted for market movements as in Brown and Warner (1985).
Absolute abnormal return The absolute stock return adjusted for market movements as in Brown and Warner (1985).
Abnormal trading volume The trading volume divided by the average 52 past observations as in Barber and Odean (2008).
Advertising expenses/Sales Advertising expenses divided by total sales as in Chemmanur and Yan (2009).
10 The database to extract the SVI from is called Google Trends. This database is not providing the actual number of search queries on a specific keyword, but instate provides a normalized SVI from 0 to 100. Whenever the search volume for a specific keyword is too low, Google Trends only returns zero’s. In this case the SVI is not included on the sample. Also tickers with an ambiguous meaning are excluded from the sample too reduce noise. Examples of these ambiguous ticker names are AIR, PC, TEL, MAN and GAS.
Google trends offers the option to include or exclude particular regions in world when obtaining the SVI for a keyword. In this study no regions are excluded, the SVI is based on worldwide searches. This criteria is chosen based on the idea that not only European investors with a possible attention towards a stock have the possibility to buy or sell the stock, but all investors are able to make transactions on the European stock market.
The Thomson one ticker code is collected for all companies included in the S&P 350 Europe from 2011 to 2014. The S&P 350 Europe is an index that contains Europe’s largest market capitalized companies out of 17 European markets. The median market capitalization of the constituents included is 11,599 million euro, with the largest company with a market capitalization of 176,165 million euro and the smallest 1,170 million euro. The index seems well diversified in 10 different sectors with Financials (21.9%), Health care (13.3%) and Consumer Staples (13.2%) the largest sectors included.
Table 2: Data sample of S&P 350 Europe stocks
The following table shows how the data sample used in this paper is constructed for each of the search volume index. The second column shows the amount of stocks which yield a valid search volume index out of Google Trends. The third column provides information on the amount of stocks excluded in the data sample to minimize noisy ticker names. The fourth column shows the amount of time periods are included for the given dependent variable. The fifth column gives the total number of observations included for each SVI.
Search Volume Index
Number of stocks with a valid search volume index
Number of stocks with an ambiguous ticker name
Number of time periods included
Number of observations
Ticker weekly 319 17 137 43,703
Name weekly 329 N/A 137 45,073
Ticker daily 288 17 45 12,960
Name daily 331 N/A 43 14,233
11
Table 3: Descriptive statistics for the search volume index
The following table shows the distribution of the Search Volume Index. *, **and *** represent a 10%, 5%
and 1% significance level, respectively. The sample period is between July 2011 and December 2013 for the weekly frequency SVI and for the daily data set January 2014 until March 2014.
Search Volume Index
Mean Median Standard
deviation
Skewness Kurtosis Jarque-Bera
Ticker weekly 63.54 67 21.06 -0.75 3.25 4493***
Name weekly 60.21 64 21.58 -0.57 2.71 2818***
Ticker daily 67.54 72 36.32 -0.92 3.57 3096***
Name daily 60.98 64 34.41 -0.45 2.44 1065***
The descriptive statistics of the SVI are presented in table 3, the SVI has a non-normal distribution as pictured on the Jarque-Bera (Jarque and Bera, 1980) test statistic. Due to the sufficiently large data set, the analysis performed on non-normal distributed data is assumed to be inconsequential (Brooks, 2008). As the numbers are normalized using the total search volume in Google, the mean and the median can’t be interpreted as absolute numbers. Rather the differences over time are of interest, which are assumed to be an increased or decreased attention of retail investors towards a stock in this thesis. The standard deviation shows there are large shifts of search volume on both the name as the ticker on daily and weekly data frequency.
The return indexes of companies quoted on the S&P 350 Europe are collected both on daily as weekly frequency data out of Thomson Datastream. To calculate the return at time t, the following formula is used:
𝑅𝑖,𝑡 =
𝑃𝑖,𝑡−𝑃𝑖.𝑡−1
𝑃𝑖,𝑡−1 (1)
Where 𝑅𝑖,𝑡 is the return at period t for stock i, 𝑃𝑖,𝑡 the stock price at time t for stock I and 𝑃𝑖,𝑡−1 the stock price at time t -1 for stock i.
To correct for overall market movements in the stock returns, the market model of Brown and Warner (1985) is used. The S&P 350 Europe is used to estimate the β and the α for company i.
𝐴𝑅𝑖,𝑡= 𝑅𝑖,𝑡− (𝛼𝑖+ 𝛽𝑖∗ 𝑅𝑚,𝑡) (2)
12 The search volume index is adjusted for random movements by subtracting the normal level of searches (Da et al., 2011). This prevents the search volume index for capturing noise movements instead of abnormal movements. Though, as a robustness check an analysis is also performed without this adjustment, simply using the SVI to capture investors’ attention.
𝐴𝑆𝑉𝐼𝑖,𝑡= log(𝑆𝑉𝐼 𝑖,𝑡) − log [𝑀𝑒𝑑 (𝑆𝑉𝐼𝑡−1, … , 𝑆𝑉𝐼𝑡−8)] (3) Where 𝐴𝑆𝑉𝐼𝑖,𝑡 is the abnormal search volume index at time t, log(𝑆𝑉𝐼𝑖,𝑡) is the natural logarithm of the search volume index at time t and log [𝑀𝑒𝑑 (𝑆𝑉𝐼𝑡−1, … , 𝑆𝑉𝐼𝑡−8)] is the natural logarithm of the median of the past eight observations.
In the analysis the ASVI is controlled for three other measures of attention: abnormal trading volume and absolute abnormal return (proposed by Barber and Odean, 2008) and advertising expenses/sales (proposed by Chemmanur and Yan, 2009).
To reduce noise on abnormal trading volume, the trading volume is divided by the average 52 past time periods (Barber and Odean, 2008).
𝐴𝑉𝑖,𝑡 = 𝑉𝑖,𝑡
𝑉𝑖,𝑡 (4)
Where 𝐴𝑉𝑖,𝑡 is the abnormal volume traded on time period t for stock i, 𝑉𝑖,𝑡 is the volume traded at time period t for stock i and 𝑉𝑖,𝑡 is the average traded volume based on the past 52 past time periods for stock i at time period t.
For the absolute abnormal return the market model of Brown and Warner (1985) is used. Note that this is the absolute difference between the return and predicted return indicated by the brackets.
𝐴𝑏𝑠𝑜𝑙𝑢𝑡𝑒 𝐴𝑅𝑖,𝑡= ⃓ 𝑅𝑖,𝑡− (𝛼𝑖+ 𝛽𝑖∗ 𝑅𝑚,𝑡)⃓ (5) Where 𝐴𝑏𝑠𝑜𝑙𝑢𝑡𝑒 𝐴𝑅𝑖,𝑡 is the absolute abnormal return at time t for stock I, 𝑅𝑖,𝑡 the return that is observed at time t for stock i and (𝛼𝑖,𝑡+ 𝛽𝑖∗ 𝑅𝑚,𝑡) is the return that is expected from the market after estimating the β and the α. The 𝛽 and 𝛼 are estimated in the two years prior to the sample period both for the weekly sample as the daily sample.
13
Table 4: Descriptive statistics for controlling measures of attention
The following table shows how the control variables are distributed depicted in the first column. *, **and ***
represent a 10%, 5% and 1% significance level, respectively. The sample period is between July 2011 and December 2013.
Control variable Mean Median Standard
deviation
Skewness Kurtosis Jarque-Bera
Abnormal trading volume 0.97 0.89 0.46 6.46 169.79 48128206***
Absolute abnormal returns 1.18 1.05 0.64 4.60 47.20 25481***
Advertising expenses / Sales 2.09 1.76 0.96 1.23 1.53 102***
Panel B: Daily data set January 2014 until March 2014.
Control variable Mean Median Standard
deviation
Skewness Kurtosis Jarque-Bera
Abnormal trading volume 1.14 1.01 0.65 5.32 59.96 1480062***
Absolute abnormal returns 0.01 0.008 0.012 4.668 60.77 1504622***
Advertising expenses / Sales 2.71 1.94 0.66 9.19 113.24 5062233***
Table 4 shows the descriptive statistics for the investors’ attention measures used to control for when analyzing the impact of the ASVI on abnormal returns. Also here the sample is non-normal distributed showed in the Jarque-Bera test statistic (Jarque and Bera, 1980) for normality. For these measures of attention a lower standard deviation is found compared to the SVI.
Table 5: Correlation between measures of attention
Panel A: The following table shows the correlation between the measures of attention on a weekly frequency The sample period is between July 2011 and December 2013.
ASVI_Ticker ASVI_Name Abnormal trading volume Absolute abnormal return ASVI_Ticker 1.000 ASVI_Name 0.124 1.000
Abnormal trading volume 0.042 0.003 1.000
Absolute abnormal returns -0.001 0.006 0.168 1.000
Panel B: The following table shows the correlation between the independent variables which are observable on a daily frequency. The sample period is between January 2014 and March 2014.
ASVI_Ticker ASVI_Name Abnormal trading volume Absolute abnormal return ASVI_Ticker 1.000 ASVI_Name 0.000 1.000
Abnormal trading volume -0.002 0.022 1.000
Absolute abnormal returns 0.002 0.007 0.591 1.000
14 frequency data this is rather high. A correlation between stock returns and trading volume is found in a number of studies for example in the paper by Jain and Joh (1988).
The correlation between the ASVI on the ticker of a company and the ASVI on the name of the company is at the low end for weekly frequency search volume with a correlation of 0.12. For daily frequency search volume there is no correlation between these two. This could be caused due to the limited amount of daily observations included in the daily frequency sample.
3.3 Methodology
The first hypothesis that studies the relationship between measures of investor attention, will be analyzed using a vector autoregressive model. This model tests each of the measures of attention as a dependent variable explained by the other measures of attention. The advantage of using a vector autoregressive model over an OLS to test the relationship is that it’s not sure which variable leads and which variables lags. The optimal lag structure is based on the lowest Akaike information criteria (Akaike, 1981), for the weekly data these are nine lags and for the daily data these are seven lags. A constant and a time trend is used in running the vector autoregressive model. Four different measures of attention that can be observed on a weekly basis are tested: the SVI on the ticker of the company, the SVI on the name of the company, absolute abnormal returns and abnormal trading volume. The variable advertising expenses/sales is not used in this analysis because this measure of attention is only available on a yearly basis. Da et al. (2011) also include Chunky news reported in the Dow Jones news archive as a measure of attention, though this would be an interesting variable to include in the analyses, due to data constraints this variable is not included.
The second, third and fourth hypotheses that test the predictive power of the ASVI, are studied by estimating the regression coefficient of the ASVI after controlling for other measures of attention i.e. abnormal trading volume, abnormal returns and advertising expenses/sales. This is done by a pooled ordinary least squares (Pooled OLS) estimate. Da et al. (2011) also include the interaction variable ASVI * log market capitalization to capture the influence of company size on the results. The market capitalization is included to control for risk (Da et al., 2011). The following multivariate regression model will estimate the impact of the ASVI on the abnormal return measured with 𝛽1.
15 The use of pooled panel data doesn’t distinguish between companies, though there could be specific effects. To test whether the effects are random or fixed a Hausman test (Hausman and Tayler, 1981) is performed on the random effect model. On both the weekly as the daily regression the null hypothesis is rejected, therefore the fixed effect model is used to estimate the coefficient of the regression (5).
16
4. Results
This section of the paper present the results on the analysis of the SVI as a measure of retail investors’ attention. At first the results of the lead-lag relation between the SVI and other measures of attention is presented. Secondly the predictive power of the SVI on a weekly frequency is analyzed to test whether this is in line with the results of Da et al. (2011). After this the same analyzes is performed on a daily frequency where the second hypothesis can be answered. Lastly, the results of the predictive power on daily frequency is tested on company characteristics.
4.1 Does the SVI lead other attention measures?
Table 5 present the results of a vector autoregressive model where four measures of attention are included in the analyzes on both a weekly frequency and daily frequency. The first panel shows the measures of attention on a weekly frequency, an important result is that the SVI on the ticker is a leading the search volume on the name of the company. This result is in line with the work of Da et al. (2011) and indicates that the SVI on the ticker is a better measure of attention compared to the SVI on the name of the company. The SVI on the ticker is also leading abnormal trading volume significantly when lagged a week. Absolute abnormal returns does lead the SVI on the ticker, indicating that this measure of attention is more timely in capturing investors’ attention.
For the daily frequency sample the results on the weekly frequency don’t hold. The SVI on the ticker is not leading the other measures of attention though the coefficient are slightly positive. These results don’t fully support the first hypothesis, therefore the first hypothesis is rejected.
Table 6: Vector autoregression model drivers of investor attention
Panel A: The following table compares the lead-lag relation between the search volume on the ticker and name, abnormal trading volume and absolute abnormal returns using a vector autoregression (VAR) model. Panel A shows the results on a weekly frequency for a sample period from June 2011 until December 2013 Both a constant and a time trend are included when running the VAR. *, **and *** represent a 10%, 5% and 1% significance level, respectively.
Lagged 1 Week SVI_Ticker SVI_Name Abnormal trading volume Absolute abnormal return R2 SVI_Ticker 0.7177*** -0.0067 -0.0183* 0.3478** 0.48 SVI_Name 0.0089*** 0.7028*** -0.0038 -0.0156 0.50
Abnormal trading volume 0.0025** 0.0042*** 0.0033*** 0.1925*** 0.25
Absolute abnormal return -0.0001 0.0000 0.3638*** 0.3638*** 0.05
17
Panel B: This panel shows the VAR result on a daily frequency for a sample period from January 2014 until March 2014. . *, **and *** represent a 10%, 5% and 1% significance level, respectively.
Lagged 1 Day SVI_Ticker SVI_Name Abnormal trading volume Absolute abnormal return R2 SVI_Ticker 0.7106*** -0.0008 -0.0021 -0.8763 0.51 SVI_Name 0.0013 0.7118*** -0.0096 1.3405 0.49
Abnormal trading volume 0.0019 -0.0019 0.0006*** -0.8266* 0.12
Absolute abnormal return 0.0000 -0.0000 0.3617*** 0.1258*** 0.02
4.2 Predicting weekly stock returns using retail investors’ attention
Before looking at the predictive power of the ASVI on a daily frequency, the weekly frequency ASVI is used to test the robustness of the work of Da et al. (2011) with European stocks. After controlling for the alternative measures of attention, abnormal trading volume, absolute abnormal returns and advertising expenses, the ASVI has a significant predicting power on abnormal stock returns for the first three weeks after an abnormal increase of the ASVI. Most predictive power of the ASVI is within the first week after an abnormal increase with a coefficient of 0.505 basis point.
Table 7: ASVI_ticker and S&P 350 Europe stock returns on a weekly frequency
Panel A: This table shows the pooled OLS regression result with the dependent variable being the abnormal return in basis points during the first 4 weeks after an increase of the ASVI. The independent variables are ASVI_ticker, abnormal trading volume, advertising expenses and the market capitalization. [ ] are the standard errors. *, **and *** represent a 10%, 5% and 1% significance level, respectively. The Sample period is from June 2011 until December 2013.
Time period in weeks Week 1 Week 2 Week 3 Week 4
ASVI_ticker weekly 0.505* 0.018* 0.027** 0.010
[0.325] [0.011] [0.013] [0.010]
Abnormal Trading volume -1.126*** -0.110*** -0.119*** -0.115***
[0.378] [0.038] [0.038] [0.038]
Absolute abnormal return 1.712*** -2.801*** 0.925 -1.538***
[0.563] [0.614] [0.581] [0.582]
Log Market Cap -0.195 -2.695*** -0.168*** -2.585***
[0.033] [0.195] [0.033] [0.199]
Log Market Cap x ASVI_ticker -0.025* -0.005 -0.001** -0.001
[0.019] [0.004] [0.000] [0.000]
Advertising expenses / Sales 0.001 0.001 0.001 0.001
[0.000] [0.000] [0.000] [0.000]
Fixed effect yes yes yes yes
Observations 39304 39015 38726 38437
R2 0.024 0.027 0.025 0.025
18 Panel B: Excluding interaction term log Market Cap x ASVI_ticker and excluding attention measures.
Time period in weeks Week 1 Week 2 Week 3 Week 4
ASVI_ticker weekly 0.008 0.006 0.008 0.006
[0.040] [0.055] [0.052] [0.059]
Fixed effect yes yes yes yes
Observations 40936 40635 40334 40033
R2 0.019 0.020 0.020 0.021
The interaction term Log market cap x ASVI is significant at the first and the third week with a negative coefficient. This indicates that a change in the ASVI for smaller companies has a larger effect on abnormal returns compared to a change in the ASVI of larger companies. Panel B shows an analyzes without the interaction term and without the control variables. This analyzes shows that the ASVI doesn’t hold its significant predictive power when it’s not controlled for other measures of attention and the interaction term.
The measures of attention absolute abnormal return and abnormal trading volume proposed by Barber and Odean (2008) do indeed have a predictive power on abnormal stock returns. An increased abnormal trading volume predicts significant negative abnormal returns for the first 4 weeks. An increase in the absolute abnormal returns predicts an increase in the abnormal return a week after followed by significant negative returns in the second, third and fourth week. Advertising expenses as a control variable seems not to capture investors’ attention at all indicated by the non-significant low coefficients.
4.3 Predicting daily stock returns using measures of attention
After establishing that the ASVI is indeed able to capture investors’ attention on weekly frequency data, now daily frequency data is used to answer the research question whether daily frequency SVI has a predictive power on stock returns. Especially with the result that an increased ASVI yield high positive significant abnormal in the first week, significant returns for the first five days can be expected. Table 9 shows that the ASVI is indeed also able to predict stock returns on a daily frequency.
19 The market value of the stocks is of large influence on the predictive power of the ASVI indicated by the negative significant interaction term Market cap x ASVI_ticker. Companies with a relative small market capitalization are more under the influence of investors’ attention compared to larger companies in the daily frequency sample analyzed, this effect is also observed in the weekly frequency data. When the interaction term is withheld from the main equation (5), the coefficients and significance disappears of the ASVI depicted in panel B from table 9.
Table 8: ASVI_ticker and S&P 350 Europe stock returns on a daily basis
Panel A: This table shows the pooled OLS regression results with the dependent variable being the abnormal return during the first 5 days after an increase of the ASVI. The independent variables are the ASVI_ticker ,abnormal trading volume, absolute abnormal return and advertising expenses. [ ] are the standard errors. *, **and *** represent a 10%, 5% and 1% significance level, respectively. The sample period is from January 2014 until March 2014.
Time period in days Day 1 Day 2 Day 3 Day 4 Day 5
ASVI_ticker daily 0.125** 0.173*** 0.191*** 0.228*** 0.181**
[0.064] [0.064] [0.066] [0.069] [0.071]
Abnormal Trading volume 0.003 0.029 -0.032 -0.051* 0.017
[0.026] [0.026] [0.026] [0.027] [0.027]
Absolute abnormal return -0.920 1.652 0.692 1.021 -5.483***
[1.341] [1.355] [1.374] [1.393] [1.438]
Log Market Cap -0.076*** -0.060** -0.047* -0.043 -0.063**
[0.027] [0.027] [0.027] [0.028] [0.028]
Log Market Cap x ASVI_ticker -0.026 -0.037** -0.043*** -0.051*** -0.039**
[0.015] [0.015] [0.015] [0.016] [0.017]
Advertising expenses / Sales 0.001 0.001** 0.001** 0.001 0.001
[0.001] [0.000] [0.000] [0.000] [0.000]
Fixed effect yes yes yes yes yes
Observations 11990 11712 11432 11152 10871
R2 0.0517 0.0488 0.0435 0.0404 0.0406
Panel B: excluding interaction term log Market Cap x ASVI_ticker.
Time period in days Day 1 Day 2 Day 3 Day 4 Day 5
ASVI_Ticker daily 0.012 0.015* 0.006 0.011 0.015
[0.009] [0.009] [0.009] [0.009] [0.009]
Abnormal Trading volume 0.003 0.029 -0.033 -0.052 0.016
[0.026] [0.026] [0.026] [0.027] [0.027]
Absolute abnormal return -0.916 1.655 0.707 1.032 -5.464***
[1.341] [1.356] [1.375] [1.393] [1.438]
Log Market Cap -0.076*** -0.060** -0.046* -0.041 -0.062**
[0.027] [0.027] [0.027] [0.028] [0.028]
Advertising expenses / Sales 0.001 0.001** 0.001** 0.001 0.001
[0.001] [0.000] [0.000] [0.000] [0.000]
Fixed effect yes yes yes yes yes
Observations 11990 11712 11432 11152 10871
R2 0.0516 0.0486 0.04315 0.0398 0.0403
20
Panel C: excluding interaction term log Market Cap x ASVI_ticker and excluding attention measures.
Time period in days Day 1 Day 2 Day 3 Day 4 Day 5
ASVI_ticker daily 0.012 0.016* 0.006 0.011 0.016*
[0.009] [0.009] [0.009] [0.009] [0.010]
Fixed effect yes yes yes yes yes
Observations 12035 11756 11475 11194 10912
R2 0.02863 0.02841 0.0282 0.0283 0.0296
Without any control variables included in the pooled regression, the ASVI still has a significant predictive power on the second and fifth day after an abnormal increase of the SVI, though on a 10% significance level. Also the coefficient are decreased significantly when not controlling for other measures of attention shown in panel C of table 8.
In the main equation (5) the SVI is adjusted for noise be deducting the median of the past 8 days (3). Table nine reports the outcomes of the regression without this adjustment to the SVI as a robustness test. This change lowers the coefficients and significance considerable, only the second day after an increased SVI a significant positive return is found. This in contrast with the ASVI where all 5 days yielded in significant positive returns.
Table 9: SVI_ticker and S&P 350 Europe stock returns on a daily basis
Panel A: This table shows the pooled OLS regression result with the dependent variable being the abnormal return during the first 5 days after an increase of the SVI. The independent variables are the ASV_ticker ,abnormal trading volume, absolute abnormal return and advertising expenses. [ ] are the standard errors. *, **and *** represent a 10%, 5% and 1% significance level, respectively. Sample period is from January 2014 until March 2014.
Time period in days Day 1 Day 2 Day 3 Day 4 Day 5
SVI_ticker daily 0.005 0.025*** 0.011 0.012 -0.012
[0.008] [0.008] [0.008] [0.008] [0.008]
Abnormal Trading volume 0.006 0.029 0.031 -0.051 0.017
[0.026] [0.025] [0.026] [0.027] [0.027]
Absolute abnormal return -0.937 1.666 0.626 0.961 -5.471***
[1.345] [1.356] [1.375] [1.394] [1.439]
Log Market Cap -0.073*** -0.063** -0.049* -0.045 -0.059**
[0.027] [0.027] [0.027] [0.028] [0.028]
Log Market Cap x SVI_ticker -0.000 -0.001*** -0.003 -0.003 0.003
[0.000] [0.002] [0.002] [0.002] [0.002]
Advertising expenses/ Sales 0.001 0.000 0.000 0.000 0.000
[0.000] [0.000] [0.000] [0.000] [0.000]
Fixed effect yes yes yes yes yes
Observations 11864 11712 11432 11152 10870
R2 0.042 0.048 0.039 0.036 0.028
21
Panel B: excluding interaction term log Market Cap x ASVI_ticker.
Time period in days Day 1 Day 2 Day 3 Day 4 Day 5
SVI_Ticker daily 0.001 0.006 -0.001 0 0
[0.001] [0.009] [0.001] [0.000] [0.000]
Abnormal Trading volume 0.002 0.029 -0.031 -0.051* 0.017
[0.026] [0.026] [0.026] [0.027] [0.027]
Absolute abnormal return -0.886 1.654 0.628 0.964 -5.479***
[1.342] [1.357] [1.376] [1.395] [1.439]
Log Market Cap -0.075*** -0.059 -0.047 -0.042 -0.062**
[0.027] [0.027] [0.027] [0.028] [0.028]
Advertising expenses / Sales 0.001 0 0 0 0
[0.000] [0.000] [0.000] [0.000] [0.000]
Fixed effect Yes yes yes yes yes
Observations 11990 11712 11432 11152 10871
R2 0.038 0.043 0.034 0.031 0.026
Panel C: excluding interaction term log Market Cap x ASVI_ticker and excluding attention measures.
Time period in days Day 1 Day 2 Day 3 Day 4 Day 5
SVI_ticker daily 0.001 0.0001 -0.001 0 0
[0.001] [0.001] [0.001] [0.000] [0.000]
Fixed effect Yes yes yes yes yes
Observations 12035 11756 11475 11194 10912
R2 0.028 0.028 0.028 0.028 0.025
After excluding the interaction term from the regression all predictive power of the SVI is disappeared as shown in both panel B an panel C. When capturing retail investors’ attention it is of importance to adjust the SVI to ASVI and to control for both market capitalization and other measures of attention.
4.4 Sensitivity analysis predictive power of ASVI
The last hypothesis in this paper is whether companies with a higher propensity to speculate, will be more susceptible for ASVI than companies which aren’t. In table 10 the companies are first partitioned in five sub-samples on their market value, price/earnings ratio and dividend yield, before analyzing the predictive power of the ASVI. Only the coefficient are reported in this table, though these are controlled for abnormal trading volume, absolute abnormal returns and advertising expenses.
22 in these tables for the interaction variable ASVI * market capitalization, indicates that low market value companies are more susceptible for ASVI changes than high market value companies.
Table 10: Sensitivity analysis ASVI_Ticker on abnormal returns on a daily basis
Panel A: The following table shows the coefficients of the abnormal returns in basis points for the first 5 days after an increase of the ASVI by 1. The impact is controlled for abnormal trading volume, absolute abnormal returns and advertising expenses also the interaction variable ASVI x market cap is included. The sample is partitioned in 5 sub-samples based on the market value. 1 represent the quintile with the 20% highest market value companies and 5 represent the 20% lowest market value companies. *, **and *** represent a 10%, 5% and 1% significance level, respectively. Sample period is from January 2014 until March 2014.
Day 1 Day 2 Day 3 Day 4 Day 5
Quintile 1 (High) -0.101* -0.155* -0.054 0.037* 0.313 2 -5.226*** 2.880*** 3.845*** -1.129** 0.567* 3 1.393* -5.838** -1.723** -7.602*** 2.872*** 4 2.315** 3.098** -1.104 -3.088*** -0.547* 5 (low) 1.457** 1.611* 0.649* 1.458** 0.509** High – low -1.56 -1.77 -0.70 -1.42 -0.20
Controlled for measures of attention yes yes yes yes yes
Fixed effect yes yes yes yes yes
Panel B: Partitioned on Price/Earnings ratio
Day 1 Day 2 Day 3 Day 4 Day 5
Quintile 1 (high) 1.241** 0.972* -0.077* 0.164* -0.760*** 2 -0.115* -0.442* -0.467** 0.143** -0.013 3 -0.031** -0.034 0.064* 0.057* 0.080 4 0.074* 0.102** 0.183*** 0.233*** 0.176** 5 (low) 0.142** 0.151** 0.942*** -0.452*** -2.090*** High – low 1.10 0.82 -1.02 0.62 1.33
Controlled for measures of attention yes yes yes yes yes
Fixed effect Yes yes yes yes yes
Panel C: Partitioned on dividend yield
Day 1 Day 2 Day 3 Day 4 Day 5
Quintile 1 (high) 0.461** 0.159** -0.306** -0.298* -0.539** 2 0.052* 0.187* 0.229* 0.321** 0.197* 3 -0.194** -0.317** -1.112* -1.195** -0.961*** 4 1.577*** 2.383*** 1.445** 1.967*** 0.252* 5 (low) 4.373*** 3.329*** 0.152* -1.871* 0.518*** High – low -3.912 -3.170 -0.458 1.573 -1.057
Controlled for measures of attention Yes yes yes yes yes
Fixed effect Yes yes yes yes yes
23 ASVI, for the fourth and the fifth day still a difference is observed where the impact on low market value companies is larger than on high market value companies.
As a measure for growth stocks, the price/earnings ratio is used to partition the sample in stocks with high growth expectations and low growth expectations. Baker and Wurgler (2006) state that stocks with a higher grow expectation are more susceptible for investor sentiment. This result also holds for investor attention as depicted in panel B of table ten for the first two days after an increased ASVI.
Lastly the company sample is partitioned on dividend yield, where we would expect companies with a low dividend yield to be susceptible for investors’ attention compared companies with a high dividend yield (Baker and Wurgler, 2006). This result is exactly found in the analyzes in panel C of table ten, the first three days and the fifth day there is a large difference in the coefficients of these two partitions.
5. Conclusion
This paper studies the predictive power of the Google search volume index on abnormal stock returns and its relation with other measures of attention. Based on the idea that investors first gather information about a firm using a search engine before trading a stock, and that these investors are mainly retail investors (Da et al., 2011), a positive abnormal return is expected. This result would be in line with the reference study of Da et al. (2011) in their paper “In Search of
Attention”. To control for other measures of attention, absolute abnormal returns, abnormal trading
volume and marketing expenses are included in the pooled OLS regression. In order to contribute to this new, direct, measure of attention, an analyses is also performed on a daily basis where the paper of Da et al. (2011) only studies weekly data. This paper uses the S&P 350 Europe stocks between 2011 and 2013 for the weekly data and January 2014 until February 2014 for the daily data.
The first result found is that the search volume on tickers is leading the search volume on the name and abnormal trading volume. The absolute abnormal returns is however leading the search volume on the ticker. This result only holds for the weekly data sample, on a daily sample no lead-lag relation is found for any measure of attention. The search volume on the ticker is leading the search volume on the name of the company, which makes using the ticker the best search volume to predict stock returns.
24 When looking at daily data frequency, an abnormal increase in the search volume does predict higher returns the following 5 days. This effect is mostly observed at the smaller market value range of the sample, companies with a high price/earnings ratio and with a low dividend yield. This result is in line with work of Baker and Wurgler (2006) who found that stocks which are more speculative are more susceptible for investor sentiment. The use of daily search volume helps to capture the attention of investors in a more timely fashion than weekly search volume.
Limitations and further research
Due to data constraints the daily data sample used in this study consists of 43 days for each stock included, this a rather limited amount. To increase the robustness of the results, a larger dataset could be used, this would also enable further research looking for a reversal. On a weekly frequency data this reversal is found in the work of Da et al. (2011), on daily frequency data this results is still missing.
25
References
Akaike, Hirotugu, 1981. Likelihood of a model and information criteria,Journal of Econometrics16, 3-14.
Baker, Malcolm, and Jeffrey Wurgler, 2006. Investor sentiment and the cross‐section of stock returns, The Journal of Finance 61, 1645-1680.
Barber, Brad M., and Terrance Odean, 2008. All that glitters: The effect of attention and news on the buying behaviour of individual and institutional investors,Review of Financial Studies21, 785-818.
Barber, Brad M., and Terrance Odean, 2000. Trading is hazardous to your wealth: The common stock investment performance of individual investors, The Journal of Finance 55, 773-806. Brin, Sergey, and Lawrence Page, 1998. The anatomy of a large-scale hypertextual Web search
engine.Computer networks and ISDN systems30, 107-117.
Brooks, C., 2008. Introductory Econometrics for Finance, Cambridge University, Cambridge Brown, Stephen J., and Jerold B. Warner, 1985. Using daily stock returns: The case of event
studies. Journal of financial economics 14, 3-31.
Chemmanur, Thomas, 2009.Advertising, attention, and stock returns. Diss. School of Business Administration, Fordham University, Working paper.
Choi, Hyunyoung, and Hal Varian, 2012. Predicting the present with google trends, Economic
Record 88, 2-9.
Choi, Hyunyoung, and Hal Varian, 2009. Predicting initial claims for unemployment benefits,
Google Inc.
Da, Zhi, Joseph Engelberg and Pengjie Gao, 2011. In search of attention, Journal of Finance 66, 1461-1497
Fama, Eugene F., 1970. Efficient capital markets: A review of theory and empirical work, Journal of
Finance 25, 383-417.
Fama, Eugene F., and James D. MacBeth. 1973, Risk return and equilibrium: Empirical tests,
Journal of Political Economy 81, 607–636.
26
Hausman, Jerry A., and William E. Taylor, 1981. Panel data and unobservable individual effects. Econometrica: Journal of the Econometric Society, 1377-1398.
Jaffe, Jeffrey F., and James M. Mahoney, 1999. The performance of investment newsletters, Journal of Financial Economics 53, 289-307.
Jain, Prem C., and Gun-Ho Joh, 1988. The dependence between hourly prices and trading volume, Journal of Financial and Quantitative Analysis 23, 269-283.
Jarque, Carlos M., and Anil K. Bera. 1980. Efficient tests for normality, homoscedasticity and serial independence of regression residuals, Economics Letters 6, 255-259.
Kahneman, Daniel, Attention and effort, 1973.
Kahneman, Daniel, and Amos Tversky, 1979. Prospect theory: An analysis of decision under risk. Econometrica: Journal of the Econometric Society, 263-291.
Nofsinger, John R., 2010. Psychology of Investing, 4th edition
Preis, Tobias, Helen Susannah Moat, and Eugene H. Stanley, 2013. Quantifying trading behaviour in financial markets using Google Trends, Scientific reports 3.
Seasholes, Mark S., and Guojun Wu, 2007. Predictable behaviour, profits, and attention, Journal of
Empirical Finance 14, 590–610.
Shleifer, Andrei, and Robert W. Vishny, 1997. The limits of arbitrage. The Journal of Finance 52, 35-55.
Shleifer, A., Summers, L., 1990. The noise trader approach to finance. Journal of Economic
