1
Master Finance Thesis
Google’s Search Volume Index: Does it Pay to Pay Attention?
by Daniël Haamke
Abstract
This paper builds on recent studies that demonstrate that individual investors are net buyers of attention grabbing stocks. The analysis shows that Google’s search volume index has predictive power in explaining the behavior of stock returns for Dutch listed firms. An increase in the search volume index is associated with higher short term stock returns and a price reversal in the long run.
Keywords: Behavioral finance, Google search volume index, investor’s attention JEL classification: G02, G10, G14, G17
Date and location: 7 January 2016, Groningen
Author: Daniël Haamke
Email: dhaamke@gmail.com
Student number: 2140918
2 1. Introduction
This paper examines the effect of investor attention, captured by using Google’s search frequency tool, on stock returns of Dutch listed firms. Merton (1987) suggests that financial models based on frictionless markets and complete information are often inadequate for capturing the entire diversity and complexity of real capital markets. Instead, he argues that investor attention could be relevant to the behavior of stock market returns. Similarly, Barber and Odean (2008) posit that investor attention affects stock prices, as individual investors are net buyers of attention grabbing stocks. Google Trends provides a search frequency tool for various search queries entered into the search engine. When people are searching online for something, they are indubitably paying attention. Therefore, Google’s search volume index indicates the amount of attention given to various search queries over time. According to Da, Engelberg and Goa (2011), it is likely that Google’s search frequency captures retail attention. The authors (Da et al. 2011) also find that an increase in investor attention at a given time, captured by an increase in search frequency, predicts higher stock prices for the following two weeks.
The crux of this paper is based on the attention-induced price pressure hypothesis of Barber and Odean (2008). In their study, Barber and Odean (2008) show that individual investors are more likely to buy stocks that receive their attention, rather than to sell these stocks. Buying permits individual investors to choose stocks from the whole investment universe, while selling stocks, in contrast, does not because investors can only sell those stocks they already hold in their portfolio. The reason that individual investors can only sell stocks they already own is due to the common inability to short stocks. Furthermore, individual investors do not have the possibility to analyze all possible investment opportunities (Kahneman, 1973). Therefore, investor attention determines the choice set for possible investments. Consequently, the findings of Barber and Odean (2008) predict that an increase in attention for a particular stock results in an increase in demand for a particular stock, which in turn should lead to temporarily higher stock prices.
At the same time, Simon (1971) contends that the availability of a large amount of information can lead to a lack of attention. Kahneman (1973) supports this theory by adding that attention is a limited cognitive resource. This argument could be considered even more pertinent in current times. The introduction of the Internet has led to an increase in available information as well as easier access to such information. The overabundance of information provided by the Internet, in combination with the fact that attention is a limited cognitive resource, renders measuring and analyzing all available material an almost impossible task. Therefore, when investors have access to an overwhelming amount of information for possible investment alternatives, as is currently the case, they may be even more prone to attention buying behavior.
3 Indicating Google’s search volume index as a proxy for investor attention is reasonable for two reasons. First of all, the search volume index represents real attention because if users search for something, they must be paying attention to what they are searching for. Second, information available online and the use of this information have increased significantly in the past two decades. Recognizing online search frequency as a measure of attention is therefore consistent with the fact that investors are gathering more and more information online (Barber and Odean, 2001).
Several scientific papers support the idea that Google’s search volume index is a proxy for investor attention. Researchers have found relationships between Google’s search volume index and both stock returns and stock market activity. The relationship between search frequency and stock market prices is confirmed by Da et al. (2011), who analyze stocks in the United States (U.S.), and by Bank et al. (2011), who assess a similar hypothesis for German stocks. The key finding of both papers is that an increase in search frequency leads to higher short term stock returns. Vlastakis and Markellos (2012) find that Google’s search volume index is significantly related to stock trading volume and historical conditional volatility.
This study investigates Dutch AEX stocks between 2010 and 2014 using the search volume index of firm names as the attention proxy. The purpose of this analysis is to examine whether Google’s search volume index for Dutch firm names captures investor attention, and if it has an impact on the behavior of stock returns. The main contribution of this paper is adding a zero-investment strategy – that longs a portfolio that contains stocks with the largest increase in search volume index and shorts a portfolio that contains stocks with the largest decrease in the search volume index – and evaluating its performance based on the weekly search volume index, instead of monthly data, as used by Bank et al (2011). Analyzing the performance on weekly data unfolds the possible short term effect of variation in attention on the performance of the zero investment strategy. Furthermore, the trading strategy is examined up to ten weeks after its formation, while Bank et al. (2011) only test for one month after the portfolio formation and do not distinguish between the different weeks within a month. Another contribution is the use of Dutch listed stocks and a more recent dataset.
This study further demonstrates that it is likely that a search volume index for firm names captures investor attention. It also shows that an increase in search volume index results in higher short term stock prices. The zero-investment strategy reveals a positive alpha for the first week after formation and a negative alpha for weeks five to 10. This indicates that the short term price increase is corrected in the long run.
4 2. Literature overview
Several studies analyze the relationship between Google’s search volume indexes and the behavior of stock prices. The first section describes the history and embedding of investor attention in the academic finance literature. The second section summarizes relevant studies that use Google’s search volume index as a measure of attention.
2.1 Financial markets and investor attention
“Although I must confess to a traditional view on the central role of rational behavior in finance, I also believe that financial models based on frictionless markets and complete information are often inadequate to capture the complexity of rationality in action.”
Robert Merton (1987)
Merton (1987) suggests that financial models based on frictionless markets and complete information, such as the Capital Asset Pricing Model (CAPM) of Sharpe-Lintner-Mossing, are often unable to entirely capture the diversity and complexity of real capital markets. In his paper, Merton (1987) suggests that there is a possibility that investor attention is relevant to the behavior of stock returns, even when the provided information investors receive has been previously announced.
5 individual or retail investors; institutional investors are less prone to attention-driven buying behavior due to several available resources, such as time, knowledge and a professional research infrastructure. This enables institutional investors to make more informed investment decisions.
An increase in attention for a particular stock causes an upward price pressure because individual investors are net buyers of attention grabbing stocks. This concept is also referred to as the price pressure hypothesis of Barber and Odean (2008), which indicates that an increase in attention for a particular stock is associated with a subsequent upward price pressure for the same stock. The price pressure hypothesis is based on the fact that individual investors are more likely to buy stocks that receive their attention, rather than to sell those stocks (Barber and Odean, 2008).
2.2 Google’s search volume index as a proxy for investor attention
Although most of the earlier mentioned indirect attention proxies are intuitively coherent, they are not a guarantee of investor attention. These proxies cannot ensure investor attention, as they only measure events that may or may not cause investors to pay attention. Information, such as news headlines, being available does not automatically imply that an investor will read an article or pay attention to the provided information. The fact that indirect measures of attention cannot definitively determine which specific information is consumed lowers the accuracy and the information value of these proxies as measures for attention.
Kahneman’s (1973) argument that attention is a cognitive limited resource for investors could be even more applicable to today’s investment practices. The introduction of the Internet has led to an increase in available information and has made access to existing investment information more easy. This development has tremendously changed the way investors gather information (Barber and Odean, 2001). Consequently, the abundance of information provided by the Internet requires more attention. The overwhelming increase in information makes measuring and analyzing all available material an almost impossible task.
In May 2006, Google Inc. launched Google Trends, a public web facility that provides normalized search volume indexes for various search-queries entered into Google. This search volume index effectively captures the variation in interest for search-keywords over time. By providing this information, a direct attention proxy became publically accessible. This because in case you search for information, you are surely paying attention (Da et al, 2011).
6 index. Moreover, Choi and Varian (2012) provide some basis research methods for predicting the future with Google’s search volume index. In their paper, they provide some comparable results for economic data and activity (Choi and Varian, 2012). According to Choi and Varian (2012), most investors follow monthly government data reports, which are normally only available with a lag. They claim that Google Trends data can help predict the future. Their study further provides some useful insights into analyzing search volume data in relation to other variables. Choi and Varian (2012) find, for example, that a monthly 1% increase in the search volume for the word ‘Ford’ is associated with a roughly 0.5% increase in the monthly sales of Ford vehicles.
Furthermore, Kulkarni, Haynes, Stough, and Paelinck (2009) find that Google’s search volume index has a predictive power on future housing prices. The authors identify a causal relationship between the search volume of combinations of words related to buying real estate, and the house price index for twenty U.S. cities (Kulkarni et al., 2009).
Together the findings of these studies indicate that Google’s search volume index contains relevant information. Recently, the search volume index has also been applied in the field of finance. Da, Engelberg, and Gao (2011) extend the research of Barber and Odean (2008) concerning investor attention by using weekly data from Google’s search volume index as a direct attention proxy. The authors research the relationship between investor attention and stock returns. They find that an increase in investor attention regarding a particular stock, measured by the weekly search frequency for the corresponding stock ticker symbols, predicts higher short term stock prices and an eventual price reversal within the year.
Similarly, Bank et al. (2011) find that the monthly search volume index for German firm names is positively related to stock trading activity and future stock returns. Bank et al. (2011) further find that an increase in the search volume index is associated with greater trading activity. Furthermore, an increase in the search volume index predicts higher short term stock returns (Bank et al. 2011). Interestingly, Bank et al. (2011) use firm names instead of stock ticker symbols. In this way, a broader measure of attention is used. One of the reasons why Bank et al. (2011) use firm names is that, unlike in the U.S., the use of ticker symbols among investors is rather uncommon. A disadvantage of using firm names, mentioned by Da et al. (2011), is that search-queries for firm names are less appropriate for measuring investor attention due to several other reasons people may have for searching for a firm name.
7 predictive proxy for future stock returns.
More recently, Ding and Hou (2015) explore the predictive power of Google Trends data and find that the majority of the variation in search volume index cannot be explained by passive indirect attention measures. Their findings are in line with the previous results of Da et al. (2011). Furthermore, Ding and Hou (2015) find that investor attention measured by changes in search volume index also has a positive effect on stock liquidity.
3. Methodology and data
3.1 Research questions and hypotheses
The main question of this paper is twofold. The first question concerns to what extent Google’s search volume index precedes the other proxies for investor attention, while the second question asks if Google’s search volume index has predictive powers regarding future stock returns.
First, I investigate whether the findings of Da et al. (2011) hold for attention measure based on the search volume of firm names, instead of the search volume index for stock ticker symbols. Therefore, the first two hypotheses and the used methodology are similar to those of Da et al. (2011). The reason why firm names are employed as an alternative for stock ticker symbols is explained in the following section (Data and descriptive statistics). Secondly, the potential investors perspective of the attention-induced price pressure hypothesis of Barber and Odean (2008) are investigated by evaluating the performance of a zero-investment strategy, which longs a portfolio that contains stocks with the largest increase in search volume index and shorts a portfolio that contains stocks with the largest decrease in the search volume index.
Da et al. (2011) find that the search volume index for stock ticker symbols precedes the alternative attention measures abnormal trading volume and absolute abnormal returns. Therefore, I hypothesize that this result holds for attention measured by Google’s search volume index for firm names.
H1: The weekly search volume index for firm names precedes the alternative measures of attention
abnormal trading volume and absolute abnormal return.
8
H2: An increase in the weekly search volume index for firm names predicts higher short term stock prices.
After analyzing whether the findings of Da et al. (2011) hold, a more practical approach is used to investigate the relevance of the presented findings for investors and, hence, is more appropriate for determining economic significance. This section thus analyzes if the attention-induced price pressure hypothesis of Barber and Odean (2008) is relevant from an investor perspective, by using a trading strategy based on Google’s search volume index of firm names. Bank at al. (2011) find that a zero-investment strategy based on a monthly search volume index that longs a portfolio that contains stocks with the largest increase in search volume index and shorts a portfolio that contains stocks with the largest decrease in search volume index, demonstrates a positive alpha for the first month after the portfolio formation. As Da et al. (2011) find that an increase in search volume index predicts higher stock prices in the following two weeks, this paper analyzes the zero-investment strategy based on the weekly search volume index. The alpha is assessed up to 10 weeks after formation.
Based on previous literature, it is expected that a portfolio that contains stocks with a large increase in the search volume index will outperform a portfolio that contains stocks with a large decrease in search volume index in the short run. Therefore, an attention-induced return premium, which is not explained by recognized risk factors, is expected for the zero investment strategy that longs a portfolio that contains stocks with a large increase in the search volume and shorts a portfolio that contains stocks with a large decrease in search volume index. In other words, a positive significant alpha is expected for the zero-investment strategy in the short run. In case the return differences between the portfolios are entirely explained by the risk factors, the estimated alpha cannot be significant.
H3: A portfolio that contains stocks with a large increase in Google’s search volume index outperforms a
portfolio that contains stocks with a large decrease in search volume index in the short run.
Furthermore, Da et al. (2011) find that after five weeks following an increase in search volume index, a price reversal occurs. The zero-investment portfolio returns are analyzed for performances after five weeks, to test whether the price reversal applies to the portfolios that contain stocks with a large increase in search volume. A price reversal is expected because the attention-induced price increase is irrational. The earlier price increase is unrelated to the fundamental value of the stock and is therefore bound to be corrected by market forces.
H4: A portfolio that contains stocks with a large decrease in Google’s search volume index outperforms a
9 3.2 Data and descriptive statistics
The public web facility Google Trends offers information on how often a specific keyword is entered on the Google website relative to the total search volume or total searches in a specific time period. The search volume provided by Google Trends is a value relative to the total number of searches in the corresponding time interval. Google Trends’ search data is normalized and scaled to a range between 0 and 100. For periods when the search frequency for a particular search term is too small to pass the selected threshold, the index results in a value of 0. The search volume index demonstrates a value of 100 for periods when the largest search volume index for a particular search term is observed. Consequently, since the data is normalized, only the relative variation in the search volume index per search item can be used. Comparing the search volume index between companies does not offer much information because the index does not provide the absolute volume of total searches. This also means that the median and mean cannot be interpreted as absolute numbers.
Table 1 provides a summary of all the variables and their corresponding definitions used in this paper. Panel A presents variables based on Google’s search volume index and Panel B shows other variables related to investor attention.
For this paper the weekly search volume index of firms that were listed on the AEX between January 3rd, 2010 and January 3rd, 2015 is used. The AEX index is a common barometer for the Dutch equity market and contains the largest Dutch listed firms. The AEX index is recognized as a good representation of the Dutch stock market. Google Trends allows filtering the search volume index for particular countries so that only queries within a specified region are obtained. For this study, no filters are applied. The main reason for not adjusting the selection process to only Dutch queries is that foreign
Variable Definition
Panel A. Variables from Google Trends
SVI Aggregrate search volume frequency from Google Trends based on firm name
logSVI The log of SVI during the week
ASVI The log of SVI during the week minus the log of the median of the SVI during the
previous 8 weeks Panel B. Other variables related to investors' attention
Return The stock or portfolio return
Abnormal return The actual return minus the normal stock return expected for a stock
based on the market model as in Brown and Warner (1980; 1985)
Absolute abnormal return The absolute value of the actual return minus the normal stock return expected
for a stock based on the market model as in Brown and Warner (1980; 1985)
Abnormal trading volume The trading volume during the week divided by the average trading volume of
the past 52 weeks as in Barber and Odean (2008)
Log market cap The log of the market capitalization of the firm
10 investors with attention on particular Dutch stocks are able to trade these stocks, and it is plausible that they use the same search query for a firm.
However, Google Trends only offers limited data for Dutch stock ticker symbols, and for some ticker symbols no data is provided. Possible explanations for this could be that investors do not make use of Dutch stock ticker symbols to gather information, or the total search volume is too low to pass the threshold that Google employs to calculate the normalized search index. That there is less data available for Dutch stocker symbols accords with the fact that it is uncommon to use ticker symbols in the Netherlands and other European countries. Therefore, the search volume of firm names is used as a proxy for investor attention for this study. This approach differs from that of Da et al. (2011), which uses stock ticker symbols. The search volume of firm names is a broader proxy of attention compared to stock ticker symbols because it also captures, for example, attention for a firm’s product or services. The pairwise correlation for the weekly search volume index of firm names and their corresponding ticker symbols between 2004 and 2014 for the American stocks Apple, Cisco Systems, and Intel are respectively 0.781, 0.423, and 0.535. This illustrates that both attention measures are related and that it is arguable that the search volume index for firm names could also capture investors’ attention.
Many other studies have used the search volume index of firm names to capture investor attention. Vlastakis and Markellos (2012) use the search volume index for a firm name because they consider it a broader measure to capture the information demand for a specific stock. The authors argue that investors are more likely to use a firm name as a keyword when they are demanding information related to a firm rather than the ticker symbol, which solely provides information directly related to the stock. In other words, when an investor demands relevant information about a certain firm, they might be interested in more general information, such as news about the firm that is not necessarily directly related to the stock ticker symbol (Vlastakis and Markellos, 2012). However, Vlastakis and Markellos (2012) agree with Da et al. (2011) that the search volume index for a firm name does include some disruption by people who demand other information about the firm, but use the firm name as their search query. They assume that the demand for other information about the firm is random and should not influence the variable in a systematic manner (Vlastakis and Markellos, 2012). Bank at al. (2011) also use the firm name as a proxy for investor attention and argue that it is a much broader measure of recognition and therefore captures investor attention in a more comprehensive way compared to stock symbols. Furthermore, Bank at al. (2011) argue that it is unlikely that an average investor who searches for firm information on Google uses a stock symbol, ISIN number, or another stock specific code.
11 data at all, as well as firms that attended the AEX index during the sample period for less than two years. Due to this restriction, Aperam, DE Master Blenders, Koninklijke BAM Groep, Wereldhave, Gemalto, Imtech, and Ziggo are omitted from the sample. Table 2 displays the sample construction and the search query used for the 24 selected Dutch listed firms. Table 3 provides the descriptive statistics of the variables used in this paper.
The key variable of interest in this paper is the variable based on the search frequency on Google or Google’s search volume index. Following Da et al. (2011), the search volume index is adjusted for a normal level of attention.
𝐴𝑆𝑉𝐼𝑖,𝑡 = 𝑙𝑜𝑔 (𝑆𝑉𝐼𝑖,𝑡) – 𝑙𝑜𝑔[𝑀𝑒𝑑 (𝑆𝑉𝐼𝑖,𝑡−1, … , 𝑆𝑉𝐼𝑖,𝑡−8)] (1)
Where 𝐴𝑆𝑉𝐼𝑖,𝑡 is the abnormal search volume index for firm i at time t, 𝑙𝑜𝑔 (𝑆𝑉𝐼𝑖,𝑡) is the logarithm of search volume index of firm i during week t, and 𝑙𝑜𝑔[𝑀𝑒𝑑 (𝑆𝑉𝐼𝑖,𝑡−1, … , 𝑆𝑉𝐼𝑖,𝑡−8] is the logarithm of the median value of the search volume index during the preceding eight weeks. Adjusting the search volume index for the median value has the advantage that the index is corrected for a normal level of attention and that time movements and other low-periodicity events are removed (Da et al., 2011).
The stock return indexes are gathered from Thomson Reuters Datastream. The weekly return of a stock is calculated by using the following equation:
Firm Search Query Firm Search Query
AEGON "aegon" Kon. KPN "kpn"
Kon. Ahold "ahold" Kon. Philips "philips"
Air France - KLM "air france klm" PostNL "postnl"
Akzo Nobel "akzo nobel" Randstad "randstad"
ArcelorMittal "arcelormittal" Reed Elsevier "reed elsevier"
ASML Holding "asml" Royal Ducht Shell "shell"
Corio "corio" SBM Offshore "sbm offshore"
Kon. Boskalis Westminster "boskalis" TNT "tnt"
Kon. DSM "dsm" TomTom "tomtom"
Fugro "fugro" Unibail Rodamco "unibail rodamco"
Heineken Holding "heineken" Unilever "unilever"
ING Group "ing" Wolter Kluwer "wolters kluwer"
Sample Construction Table 2
12 𝑅𝑖,𝑡 = 𝑙𝑜𝑔 ( 𝑅𝐼𝑖,𝑡
𝑅𝐼𝑖,𝑡−1) (2)
Here 𝑅𝑖,𝑡 is the return at period t for stock i, 𝑅𝐼𝑖,𝑡 is the return index value at time t for stock i, and
𝑅𝐼𝑖,𝑡−1 is the return index value at time t with one week lag for stock i. The reason for selecting the return
index instead of the stock price is that the return index is adjusted for dividend payments.
According to Brown and Warner (1980; 1985) and MacKinlay (1997), the abnormal return of a particular stock is the actual return minus the normal return of the stock. The normal return is defined as the expected return for a particular stock, based on for example a market model (MacKinlay, 1997). The methodology for calculating the abnormal returns is similar to the methodology I used for my bachelor thesis (Haamke, 2014). For stock i at time t the abnormal return is:
ARi,t = Ri,t −E(Ri,t) (3)
Where ARi,t, ,is the abnormal return, Ri,t,is the actual return, and E(Ri,t) is the expected or normal
return for stock i at time t. According to MacKinlay (1997), the constant mean return model and the market model are two common methods for calculating the normal return for a stock. However, the market model is generally preferred because the variance of the abnormal return is reduced by adjusting the share of the return that is associated with the variation in the market return (MacKinlay, 1997). The Brown and Warner (1980; 1985) market model for a stock is:
Rit = αi + βi Rmt + εit (4)
Here Rit is return for stock i the time t and Rm is the market portfolio return at time t. αi and βi, are respectively the alpha and beta or systematic risk of stock i, and εi the error term. According to
MacKinlay (1997) the expected error term equals 0. The following equation is used to calculate the estimated alpha (α̂i):
𝛼̂𝑖 = 𝑢̂𝑖 – (𝛽̂𝑖∗ 𝑢̂𝑚) (5)
Where 𝑢̂𝑖 is the mean of the return of stock i, 𝑢̂𝑚 is the mean of the market, and 𝛽̂𝑖 is the estimated beta or estimated systematic risk parameter of stock i during the estimation window. The estimation window used for calculating the parameters 𝛼̂𝑖 and 𝛽̂𝑖, is defined as the 2 years prior to the sample period. The market portfolio used for this analysis is the AEX index. The following equation is used to calculate the estimated beta (𝛽̂𝑖):
𝛽̂𝑖 = 𝐶𝑜𝑣(𝑅𝑖,𝑅𝑚)
13
Where Ri is the return of stock i, and Rm is the market return during the estimation. The following equation is used to calculate the weekly abnormal returns:
𝐴𝑅𝑖,𝑡= 𝑅𝑖,𝑡 – (𝛼̂𝑖+ 𝛽̂𝑖𝑅𝑚,𝑡) (7)
The abnormal return (𝐴𝑅𝑖,𝑡) at time t for security i is calculated by subtracting the estimated normal return (𝛼̂𝑖+ 𝛽̂𝑖𝑅𝑚,𝑡) from the actual return (𝑅𝑖,𝑡) for the for stock i during time t. The estimated normal return is calculated by taking the estimated alpha of stock i (𝛼̂𝑖) plus the beta of stock i (𝛽̂𝑖) times the market return (𝑅𝑚,𝑡) during time t.
When applying search volume index, I control for other alternative measures of attention as in Da et al. (2011), namely, absolute abnormal return and abnormal trading volume; both measures of attention are used by Barber and Odean (2008). The absolute returns are calculated by taking the absolute value of the abnormal returns gathered with formula 7. The abnormal trading volume is calculated as in Barber and Odean (2008):
𝐴𝑇𝑉𝑖,𝑡 = 𝑉𝑜𝑙𝑖,𝑡 ∑52𝑗=1𝑉𝑜𝑙𝑖,𝑡−𝑗
52
(8)
Here 𝐴𝑉𝑖,𝑡 is the abnormal trading volume on time t for stock i, 𝑉𝑜𝑙𝑖,𝑡 is the trading volume at time t for stock i, and ∑ 𝑉𝑜𝑙𝑖,𝑡−𝑗
52 𝑗=1
52 is the average trading volume of the past 52 time periods for stock i. The absolute abnormal return is calculated as in formula 7. The difference between the abnormal return described by formula 7 and the absolute abnormal return is that the latter is the absolute difference between the expected and actual return and, hence, is always positive.
Variables Mean Median Standard Skewness Kurtosis
Deviation
Search volume index (SVI) 71.75 75.00 18.10 -0.79 3.45
Log Search Volume Index (logSVI) 4.24 4.32 0.28 -1.30 4.75
Abnormal search volume index (ASVI) -0.02 0.00 0.88 -4.13 152.11
Abnormal trading volume (ATV) 0.99 0.90 0.62 12.85 350.29
Abnormal return (AR) -0.14 -0.09 3.50 -1.02 37.44
Absolute abnormal return (AAR) 2.27 1.60 2.67 5.99 86.37
Obervations 6003 6003 6003 6003 6003
Table 3 Descriptive Statistics
14 Table 3 presents the descriptive statistics of the variables used in this paper. The search volume index demonstrates a higher standard deviation compared to the other attention measures because it is directly calculated from the raw data. The standard deviations of the abnormal search volume index, abnormal trading volume, and absolute abnormal return, which are all used in the panel regression, are more comparable to each other. Furthermore, the table shows that the samples of all variables are abnormally distributed. However, according to Brooks (2014), this is not problematic because the violation of the normality assumption is practically inconsequential for sample sizes that are sufficiently large.
Moreover, Table 4 presents the correlations among the variables of interest to indicate whether and how strongly the pairs of used variables are related to each other. The highest contemporaneous correlation, besides the correlations between the different search volume index variables, is found between the abnormal trading volume and absolute abnormal returns. This is comparable to the correlation between absolute abnormal return and abnormal trading volume found by Da et al. (2011), which is 0.311.
The zero-investment portfolios are constructed by sorting the sample of stocks by the abnormal search volume index in the previous week. Stocks that have the largest positive abnormal search volume index are the stocks with the largest increase in the search volume index and stocks with the smallest or most negative abnormal search volume index exhibit the largest decrease in search volume index. The zero-investment strategy longs a portfolio that contains stocks with the largest increase in search volume index and shorts a portfolio that contains stocks with the largest decrease in the search volume index. The portfolio that contains stocks with the largest increase in search volume index contains the five stocks that received the largest positive change in Google’s search volume index. In other words, these stocks represent the five stocks that received the greatest increase in attention. On the other hand, the portfolio that contains stocks with the largest decrease in the search volume index contains five stocks that received
Search volume Log search Abn search Abn trading Absolute abn
index volume index volume index volume return
(SVI) (logSVI) (ASVI) (AV) (AAR)
Search volume index (SVI) 1.000
Log Search volume index (logSVI) 0.983 1.000
Abn search volume index ASVI) 0.094 0.099 1.000
Abn trading volume (ATV) 0.091 0.087 0.065 1.000
Absolute abn return (AAR) 0.010 0.006 0.040 0.353 1.000
Table 4 Correlations
15 the most significant decrease in the search volume index. Consequently, this portfolio contains the five stocks that received the largest decrease in attention. The neutral portfolio contains the remaining stocks. The reason for choosing the five stocks that receive the largest increase and decrease in the relative change in attention is that dividing the sample into three quantiles of equal size, following Bank et al. (2011), results in random interferences in the high and low attention portfolios. By dividing the sample into three quantiles, the portfolio, that should contain only stocks with a large increase in search volume index, includes an occasional basis stock with a decrease in search volume index and vice versa. This is rooted in the fact that the weekly distribution of stocks with an increase or decrease in attention varies greatly. During some weeks, only several stocks receive an increase or decrease in attention. Selecting the five stocks that receive the largest increase and decrease in attention solves this impurity. This selection process is therefore not completely similar to the procedure used by Bank at al. (2011).
For the sake of convenience, the portfolio that contains stocks with a large increase in search volume index is called the portfolio with high attention grabbing stock; the portfolio that contains stocks with a large decrease in search volume index is called the portfolio with low attention grabbing stocks. The portfolio with neutral attention grabbing stocks contains the remaining stocks, with no particularly large increase or decrease in the search volume index.
The time series returns for each portfolio are constructed, following Bank at al. (2011), by taking the equally weighted average returns of the weeks following portfolio formation. The portfolios are constructed for each of the 260 weeks between January 3rd, 2010 and January 3rd, 2015. For each portfolio, the returns are calculated for the 10 weeks following the portfolio formation. The weekly returns of the portfolios are calculated as follows:
𝑅𝑡𝐻𝑃 = 𝑁1 𝐻𝑃𝑡∑ 𝑅𝑖,𝑡 𝑁 𝑖=1 (9a) 𝑅𝑡𝐿𝑃 = 𝑁1 𝐿𝑃𝑡∑ 𝑅𝑖,𝑡 𝑁 𝑖=1 (9b) 𝑅𝑡𝑁𝑃 = 1 𝑁𝑁𝑃𝑡∑ 𝑅𝑖,𝑡 𝑁 𝑖=1 (9c)
16 3.3 Research method
The first hypothesis tests if the weekly search volume index for firm names precedes alternative measures of attention as abnormal trading volume and absolute abnormal return. To test the first hypothesis, the analysis applies a Vector Auto Regression (VAR) model, as in Da et al. (2011), which use the VAR model for a similar hypothesis. The VAR model tests if each used variable is a linear function of the past lags of the other used variables or the variable itself. For the VAR, I use the variables absolute abnormal return, abnormal trading volume and the log of the Google Trends search volume index (logSVI) for firm names. Da et al. (2011) use a VAR for each stock and then average the VAR coefficients across all stocks. A slightly different method is applied here by running the VAR for the summarized data simultaneously for all stocks. The following equation presents the VAR model:
𝑌𝑡 = 𝜇 + ∑𝑃𝑖=1𝜙𝑖𝑌𝑡−𝑖+𝜀𝑡 (10)
Where 𝑌𝑡 is the column vector containing the variables absolute abnormal return, abnormal trading volume, and the log of the search volume index, 𝜇 is a constant vector, 𝜙𝑖 is the vector coefficient, 𝜀𝑡 is a vector of random error terms, and p is the optimal number of lags.
Moreover, the second hypothesis examines if an increase in the weekly Google’s search volume index for firm names predicts higher short term stock prices. To test the second hypothesis, the test uses a panel regression model to estimate a regression coefficient for the search frequency. When applying the panel regression, I control for other alternative measures of attention as in Da et al (2011), namely, absolute abnormal return and abnormal trading volume; both measures of attention are used by Barber and Odean (2008). Furthermore, the control variable market capitalization is included to control for size, as in Da et al. (2011). The coefficients are determined using a panel regression model. This analysis is comparable with that of Da et al. (2011), although they conduct a Fama-MacBeth (1973) cross-sectional regression to adjust for time specific, market wide shocks. While there could be firm specific effects, the regular panel regression used in this paper does not recognize those effects. Depending on the data, either a fixed effects model or a random effects model is used. To test whether a random or fixed effect model should be employed, a Hausman (1978) test is performed. The fixed effect model is used because the null hypothesis is rejected. This leads to the following regression model:
𝐴𝑅𝑖,𝑡= 𝛽1𝐴𝐴𝑅𝑖,𝑡−1+ 𝛽2𝐴𝑆𝑉𝐼𝑖,𝑡−1 + 𝛽2𝐴𝑇𝑉𝑖,𝑡−1 + 𝛽3Log 𝑀𝑎𝑟𝑘𝑒𝑡 𝐶𝑎𝑝𝑖,𝑡−1 + 𝑢𝑖,𝑡+ 𝑣𝑖,𝑡 (11)
17 median search volume index during the previous eight weeks, 𝐴𝑇𝑉𝑖,𝑡−1 is the abnormal traded volume for stock 1 at time t with one week lag, Log 𝑀𝑎𝑟𝑘𝑒𝑡 𝐶𝑎𝑝𝑖,𝑡−1 is the log of the market capitalization of firm i at time t, 𝑢𝑖,𝑡 represents the firm specific fixed effects, and 𝑣𝑖,𝑡 is the idiosyncratic disturbance term. The reason to apply a one week lag to independent variables is to analyze the predictive power of each variable.
The third and fourth hypotheses test the performance differences of a portfolio which contains stocks with a large increase in the search volume index, versus a portfolio with stocks with a large decrease in search volume index. To analyze the risk adjusted performance differences between the portfolio that contains stocks with the largest increase in search volume index and the portfolio that contains stocks with the largest decrease in search volume index, the analysis first calculates the alphas of a long-short portfolio. This zero-investment strategy goes long in the portfolio that contains stocks with the largest increase in attention, based on the relative positive change in the search index, and goes short in the portfolio that contains stocks with the largest decrease in attention, based on the relative small or negative change in the volume index. The time series of the equally weighted average returns of the zero investment strategy are regressed using two different factor models as in Bank et al. (2011), namely the CAPM and the Fama and French (1993) three-factor model, in order to correct for market risk and other style tilts in the long-short portfolio. The differences between the long and short portfolio may be driven by exposure to recognized risk factors as the market or style tilt. The Fama and French (1993) three-factor model corrects, besides the market factor, for size and value effects. Using both models is valuable in this analysis because differences in returns of high and low attention grabbing stocks may be due to exposure of the long-short portfolio to the mentioned risk factors. The following model is used to calculate the alpha corrected for market risk:
𝑅𝑡𝐻𝑃− 𝑅
𝑡𝐿𝑃 = 𝛼𝑗+ 𝛽1𝑗(𝑅𝑚,𝑡− 𝑅𝑓,𝑡) + 𝜀𝑗,𝑡 (12)
Where 𝑅𝑡𝐻𝑃 is the equally weighted average return in week t for the portfolio that contains stocks with the largest increase in attention stocks, based on the relative largest increase in the search volume index, 𝑅𝑡𝐿𝑃 is the equally weighted average return in week t for the portfolio that contains stocks with the largest decrease in attention, based on the relative largest decrease in the search volume index, and 𝑅𝑚,𝑡− 𝑅𝑓,𝑡 is the market return, in this case the AEX, which is one of the most common barometers for the Dutch equity index, in week t minus the risk free rate. I use the three-month Euribor as the risk free rate, which is a frequently used risk free rate, as used, for example, by Kramer (2012). The following model is used to calculate the alpha corrected for the Fama and French factors.
18 𝑅𝑡𝐻𝑃− 𝑅
𝑡𝐿𝑃= 𝛼𝑗+ 𝛽1𝑗(𝑅𝑚,𝑡− 𝑅𝑓,𝑡) + 𝛽2𝑗𝑆𝑀𝐵𝑡 + 𝛽3𝑗𝐻𝑀𝐿𝑡 + 𝜀𝑗,𝑡 (13)
Where 𝑆𝑀𝐵𝑡 is calculated, following Fama and French (1993), as equally weighted, the average return on three small portfolios minus the average returns on three large portfolios, calculated in the equation below:
𝑆𝑀𝐵𝑡 = 1
3(𝑆𝑚𝑎𝑙𝑙 𝑣𝑎𝑙𝑢𝑒𝑡+ 𝑆𝑚𝑎𝑙𝑙 𝑛𝑒𝑢𝑡𝑟𝑎𝑙𝑡+ 𝑆𝑚𝑎𝑙𝑙 𝑔𝑟𝑜𝑤𝑡ℎ𝑡) (14a) − 13(𝐵𝑖𝑔 𝑣𝑎𝑙𝑢𝑒𝑡+ 𝐵𝑖𝑔 𝑛𝑒𝑢𝑡𝑟𝑎𝑙𝑡+ 𝐵𝑖𝑔 𝑔𝑟𝑜𝑤𝑡ℎ𝑡)
Where 𝑆𝑀𝐵𝑡 is the return of a zero investment factor for size at time
t, 𝑆𝑚𝑎𝑙𝑙 𝑣𝑎𝑙𝑢𝑒𝑡, 𝑆𝑚𝑎𝑙𝑙 𝑛𝑒𝑢𝑡𝑟𝑎𝑙𝑡, and 𝑆𝑚𝑎𝑙𝑙 𝑔𝑟𝑜𝑤𝑡ℎ𝑡 are the returns for the index for the MSCI Netherlands Small Value index, the MSCI Netherlands Small Neutral index, and the MSCI Netherlands Small Growth index in week t. 𝐵𝑖𝑔 𝑣𝑎𝑙𝑢𝑒𝑡, 𝐵𝑖𝑔 𝑛𝑒𝑢𝑡𝑟𝑎𝑙𝑡 and 𝐵𝑖𝑔 𝑔𝑟𝑜𝑤𝑡ℎ𝑡 are the returns for the index for the MSCI Netherlands Big Value index, MSCI Netherlands Big Neutral index, and the MSCI Netherlands Big Growth index in week t. 𝐻𝑀𝐿𝑡 is calculated, following Fama and French (1993), as the average return on two value portfolios minus the average returns on two growth portfolios, calculated as in the equation below:
𝐻𝑀𝐿𝑡 = 1
2(𝑆𝑚𝑎𝑙𝑙 𝑣𝑎𝑙𝑢𝑒𝑡+ 𝐵𝑖𝑔 𝑣𝑎𝑙𝑢𝑒𝑡) (14b) − 12(𝑆𝑚𝑎𝑙𝑙 𝑔𝑟𝑜𝑤𝑡ℎ𝑡+ 𝐵𝑖𝑔 𝑔𝑟𝑜𝑤𝑡ℎ𝑡)
Here 𝐻𝑀𝐿𝑡 is the return of a zero-investment factor for size at time t. 𝑆𝑚𝑎𝑙𝑙 𝑣𝑎𝑙𝑢𝑒𝑡, and 𝐵𝑖𝑔 𝑣𝑎𝑙𝑢𝑒𝑡 are the returns for the MSCI Netherlands Small Value index and for the MSCI Netherlands Big Value index in week t. 𝑆𝑚𝑎𝑙𝑙 𝑔𝑟𝑜𝑤𝑡ℎ𝑡 and 𝐵𝑖𝑔 𝑔𝑟𝑜𝑤𝑡ℎ𝑡 are the returns for the MSCI Netherlands Small Growth index and for the MSCI Netherlands Big Growth index in week t.
4. Results
19 different timeslots in order to investigate the performance differences of stocks with the largest increase and decrease in search volume index during several weeks after portfolio formation. This section concludes with a robustness check.
4.1 The lead-lag relation among measures of attention
The lead-lag relation among measures of attention is examined by using a VAR. There are three variables used in the VAR model, namely the log of the search volume index, the abnormal trading volume, and the absolute abnormal return. The VAR coefficients are reported in Table 5. According to the results, the coefficient on the lagged search volume index is positive and significant for the current week’s abnormal trading volume (Column 1 Table 5). Therefore, one can infer that search volume index leads the attention proxy abnormal trading volume. This result is consistent with the results found by Da et al. (2011). The lagged search volume index does not lead the attention measure absolute abnormal return. However, the abnormal trading volume does lead the absolute abnormal return (Column 3 Table 5). Hence, it may be that the search volume index lagged for two weeks leads abnormal trading volume. Although the results partly support the hypotheses that the weekly search volume index for firm names leads the alternative measures of attention abnormal trading volume and absolute return, there is not enough evidence to fully accept the hypothesis. The first hypothesis is therefore rejected.
Log search Absolute abn Abn trading R2
volume index return volume
(logSVI) (AAR) (AV)
(1) (2) (3) (4)
Log Search volume index (logSVI) 0.7058*** -0.0009 -0.0108*** 0.772
(0.01) (0.15) (0.01)
Absolute abn return (AAR) -0.0920 0.1565*** 0.2005*** 0.065
(0.40) (0.01) (0.01)
Abn trading volume (ATV) 0.2157*** 0.0221*** 0.2180*** 0.105
(0.01) (0.01) (0.01)
Lagged 1 Week
Table 5
Vector Autoregression model of Attention Measures
20 4.2 Google’s search volume index and price pressure
21 trading volume has a significant, positive predictive power regarding future abnormal returns. The positive effect of abnormal trading volume in week 1 is corrected in week 3 by a significantly larger negative effect. Da et al. (2011) do not find a significant negative effect in the subsequent weeks. Therefore, it is possible that the positive effect for Dutch stocks is corrected in a more timely fashion by market forces compared to the sample of Da et al. (2011). Furthermore, the control variable log market capitalization has a strong negative significant influence on future abnormal returns. This is in line with earlier results that firms with a smaller market capitalization have higher returns, on average, than firms with a large market capitalization (Fama and French, 1993).
4.3 Google search activity and short run portfolio returns
In this section, I test the third hypothesis, which states that a portfolio that contains stocks with a large increase in Google’s search volume index outperforms a portfolio that contains stocks with a large decrease in search volume index in the short run. Despite the insignificant findings of Google’s search volume index in the previous section, there could still be an effect, as insignificant results may be driven by a lack of power of the used sample, and the sign of the coefficients for the search volume index in Table 6 are in line with the attention-induced price pressure hypothesis of Barber and Odean (2008). This analysis differs from the previous section’s results, as the portfolios are constructed with stocks that receive the largest increase and decrease in search volume index. By selecting only such stocks, a stronger effect is expected.
Week 1 Week 2 Week 3 Week 4 Weeks 5
(1) (2) (3) (4) (5)
Abnormal search volume index (ASVI) 0.040 0.016 -0.043 -0.012 -0.020
(0.052) (0.053) (0.053) (0.053) (0.053)
Abnormal trading volume (ATV) 0.125* -0.105 -0.182** 0.002 -0.057
(0.079) (0.079) (0.079) (0.079) (0.080)
Absolute abnormal return (AAR) 0.031* 0.043** -0.024 -0.056*** 0.017
(0.019) (0.019) (0.019) (0.019) (0.019)
Log market capitalization -0.411*** -0.444*** -0.507*** -0.536*** -0.472***
(0.155) (0.156) (0.156) (0.157) (0.158)
Observations per week 5980 5957 5934 5911 5888
R2 0.017 0.017 0.017 0.017 0.016
Table 6
ASVI and AEX Stock Returns
22 The hypothesis is tested by analyzing the alpha of a zero investment strategy that longs a portfolio that contains stocks with a large increase in Google’s search volume index and shorts a portfolio that contains stocks with a large decrease in search volume index. The time return series are constructed as in Bank et al. (2011) by taking the equally weighted average returns of the weeks following the portfolio formation. For the sake of convenience, the portfolio that contains stocks with a large increase in search volume index is labeled the portfolio with high attention grabbing stock; the portfolio that contains stocks with a large decrease in search volume index is termed the portfolio with low attention grabbing stocks. The portfolio with neutral attention grabbing stocks contains the remaining stocks. For each of the 522 high and low attention portfolios, the returns are calculated for the 10 weeks following portfolio formation. The time series returns of the long-short portfolio are regressed on risk factors by employing two different factor models, namely the CAPM and the Fama and French three-factor model (1993), in order to correct for market risk and other style tilts in the long-short portfolio.
Table 7 presents the results of the zero-investment strategy. The results show that the alpha for the long-short strategy is positive and significant for the first week for both the market model or CAPM, and the Fama and French (1993) three-factor model (Columns 1 and 2 in Panel A, Table 8). The alpha for the long-short strategy, however, becomes insignificant for the following weeks. The results also reveal that the alpha decreases and even becomes negative after the second week. Panels B and C of Table 8 present the alphas for the two attention portfolios. The alpha for the high attention portfolio is positive but not significant, while the alpha for the low attention stock is negative and significant for the market model (Columns 1 & 2 in Panels B & C, Table 8). Therefore, both portfolios contribute to the short term positive alpha for the long-short strategy. According to the results, it appears that stocks with a large increase in the search volume index outperform stocks with a large decrease in search volume index. The results demonstrate an attention induced return premium in the first week after the change in the search volume index, which is not captured by any risk factors. This short term upwards price pressure for stocks with an increase in search volume index is consistent with the hypothesis and with the attention-induced price pressure hypothesis introduced by Barber and Odean (2008) and the finding of Da et al. (2011).
4.4 Google search activity and long run portfolio returns
23 three-factor model (Columns 9 & 10 in Panel A, Table 7). The alpha for weeks 5 to 10 is low compared to the first two weeks. Panels B and C of Table 8 present the alphas for the high and low attention portfolios. The alpha for the high attention portfolio is negative and significant, as is the alpha for the portfolio with low attention stock (Columns 9 & 10 in Panels B & C, Table 7). The alpha for the high attention portfolio is more than twice as negative as the alpha for the low attention portfolio, indicating that the high attention stocks receive a large price correction. This price reversal is consistent with the hypothesis and findings of Da et al. (2011).
25 4.5 Statistical versus economic significance
The provided test results show that stocks that receive a large increase in attention measured by the search volume index are expected to have higher short term stock prices and an eventual price reversal. In contrast with the stocks that receive a large decrease in attention, which are associated with lower short term stock prices. The significant results of the zero-investment strategy suggest that an attention based investment strategy executed by participating in a real equity market is profitable. However, there is a difference between statistical and economic significance (Hoover and Siegler, 2008). Statistical significance, based on, for example, t-tests and p-values, does not guarantee economic significance. For instance, effects could be significant, but too small for economic significance. Although the differences in returns of the portfolio are large and significant, they do not indicate whether the zero-investment strategy is profitable, including transaction costs associated with participating and trading in real equity markets.
26
Figure 1. Performance of different attention portfolios. This figure shows the performance of the long-short portfolio, the portfolio that contains stocks with the largest increase in search volume index (named the portfolio with high attention stocks), the portfolio that contains stocks with the largest decrease in search volume index (named the portfolio with low attention stocks), and the market portfolio or benchmark. The portfolios are constructed based on the change in search volume index of the firm names in the previous week. The long-short portfolio longs the portfolio with high attention stocks and shorts the portfolio with low attention stocks. The portfolios are constructed every Monday, based on the abnormal search volume index of the previous week. On Friday, the stocks of both portfolios are sold, so that the investment process can be repeated on the subsequent Monday morning. As such, the portfolios are rebalanced each week. This strategy is repeated every week, starting from January 15th 2010 until January 3rd 2015.
27 (Keim and Madhaven, 1997). The commission fee for executing an order depends on the broker. Some brokers use fixed fees while other use a percentage of the traded value or a combination of both. For example, the Dutch low-cost broker De Giro asks for 2 Euros plus 0.02% of the total transaction as commission (gathered from https://www.degiro.nl on December the 14th, 2015). The bid-ask spreads for stock listed on the AEX index are relatively low, due to the liquidity of these stocks. The bid-ask spread for the stock Royal Dutch Shell, listed on the AEX-index on December 14th 2015, is 0.04% (gathered from http://www.lse.co.uk/SharePrice.asp?share- price=RDSA on December the 14th 2015). The zero investment strategy, which cannot be executed by individual investors who cannot short stocks, could have been interesting for intuitional investors. As long as the weekly transaction costs were lower than the average weekly alpha of the strategy, which is 0.303% for this sample (Column 1 Table 7). Assuming the earlier mentioned transaction costs are representative, then 0.06% constitutes the variable transaction costs plus a fixed fee of 2 Euros. The average turnover of the portfolios with high and low attention stocks, which both contain five stocks, is 50%. This results in an average of five transactions per week for the long-short portfolio, which equals 0.30% variable costs (0.06% * 5). In this case, excluding the fixed costs of 2 Euros, which are relatively negligible for large transactions, the weekly transaction costs are more or less equal to the weekly alpha. Consequently, assuming that the used transaction costs are representative, this strategy is on average not profitable for the used sample. However, in case lower transaction costs apply, the investment strategy may be profitable.
4.6 Robustness check
This section provides the robustness check, which is performed by repeating the regression used for testing hypotheses 3 and 4 for two subsamples. The subsamples are obtained by dividing the total sample into one sampling period from the 3rd of January 2010 until the 13th of July 2012 and one sampling period from the 13th of July 2012 until the 3rd of January 2015. This check is similar to the robustness check used by Da et al. (2011). The robustness check tests only the main results. Therefore, the long-short portfolio is analyzed for both samples after weeks 1 and 5 until ten weeks after the portfolio formation.
28 According to the results of the robustness check, there is limited evidence that stocks with a large increase in the search volume index outperform stocks with a large decrease in search volume index for both subsamples. The results of this robustness check are therefore in line with the earlier found results and support the third hypothesis. On the other hand, the alphas for the long run are less comparable with previous results shown in Table 8. Only the first subsample demonstrates negative significant alphas, which are also more negative than the earlier observed long term alphas (Columns 3 & 4 in Panel A, Table 8). In contrast with earlier results, the long term alphas for the second subsample reveal (Columns 7 & 8 in Panel A, Table 8) positive alphas for the market model (CAPM, and the Fama and French (1993) three-factor model. This result is caused by the large negative alphas of the low attention portfolios of the second subsample (Columns 7 & 8 in Panel C, Table 8). The alphas for the high attention portfolios of the second subsample are also negative, but less negative then the alphas for the low attention portfolios. The results of the robustness check are therefore not fully in line with earlier found results in this paper.
CAPM FF CAPM FF CAPM FF CAPM FF
(1) (2) (3) (4) (5) (6) (7) (8)
Panel A: Coefficients the for long-short portfolio
Market -0.083 -0.183* 0.129 0.082 0.007 0.113 0.155 0.187 (0.075) (0.095) (0.083) (0.098) (0.106) (0.114) (0.110) (0.115) SMB -0.052 0.197 0.131 0.273 (0.121) (0.099) (0.136) (0.153) HML 0.232 0.102 -0.304 -0.197 (0.146) (0.121) (0.125) (0.114) Alpha 0.314 0.354* -1.427*** -1.180** 0.274 0.230 0.369 0.451 (0.217) (0.218) (0.523) (0.539) (0.205) (0.205) (0.497) (0.530) Observations 131 131 131 131 130 130 130 130 R2 0.009 0.031 0.018 0.055 0.000 0.045 0.015 0.044
Panel B: Alphas for the high-attention portfolio
Alpha 0.046 0.132 -1.204*** -0.752* 0.159 0.217 -0.717* -0.253
(0.171) (0.159) (0.448) (0.392) (0.159) (0.153) (0.366) (0.362)
Panel C: Alphas for the low-attention portfolio
Alpha -0.268* -0.222 0.223 0.428 -0.116 -0.012 -1.087*** -0.704**
(0.163) (0.153) (0.390) (0.378) (0.149) (0.142) (0.328) (0.342)
Week 1 Weeks 5-10 Week 1 Weeks 5-10
Table 8
29 5. Conclusion
5.1 Summary and key findings
This paper examined to what extent Google’s search volume index of firm names precedes the alternative measures of attention abnormal trading volume and absolute abnormal returns, while it further analyzed the predictive power of the search frequency on future stock returns for Dutch listed firms on the AEX between 2010 and 2014. Indicating the search volume index as a proxy for investor attention is reasonable, as when people are searching online for information about a company they are inadvertently paying attention to the subject of their investigation. The idea that investor attention is relevant for stock prices is based on a hypothesis posed by Barber and Odean (2008) contending that individual investors are net buyers of attention grabbing stock and that it is likely that many individual investors consider buying only those stocks that initially catch their attention. According to this hypothesis, an increase in attention for a particular stock results in a temporary upward price pressure for the same stock. This attention-induced price pressure hypothesis is confirmed by the findings of Da et al. (2011) and Bank et al. (2011) by using the search volume index as a direct proxy for investor attention. As in Vlastakis and Markellos (2012) and Bank et al. (2011), this paper also applied Google’s search volume index of firm names as attention proxy.
The analysis yielded significant results indicating that the search volume index leads to the attention proxy abnormal trading volume. However, it did not find significant results demonstrating that the search volume index leads to the attention proxy absolute abnormal returns. Furthermore, this paper revealed a positive but insignificant predictive relationship between an increase in the search volume index for firm names, positive abnormal returns during the first two weeks, and negative abnormal returns after the second week, which is in line with the findings of Da et al. (2011). The reason for the insignificant results may be rooted in a lack of power of the used sample. The sample size used for this paper is relatively small compared to the sample used by Da et al. (2011). Another reason could be that by using the search volume index for firm names, rather than stock ticker symbols as in Da et al. (2011), arbitrary inconsistencies are included in the search data. This is due to the fact that the search volume of firm names is a broader proxy of attention compared to stock ticker symbols, as it also captures, for example, attention for a firm’s product or services.
30 Additionally, this study demonstrates significant results indicating that, in the long term, a portfolio, which contains stocks with a large decrease in the search volume index, outperforms a portfolio of stocks with a large increase in search volume index in the long run.
In general, this study supports the concept that Google’s search volume index captures investor attention for Dutch stocks and finds that an increase in this attention results in an attention-induced price pressure, introduced by Barber and Odean (2008) and supported by the empirical studies of Da et al. (2011) and Bank et al. (2011).
5.2 Limitations of the study and directions for future research
The main limitation of this analysis is that a relatively small sample was used to test for statistical significance. Because the construction of the sample is limited, the sample has a lack of power and results might be biased. Furthermore, this decreased the robustness of several test results in this paper. Therefore, it could be interesting to examine whether a larger sample would yield stronger results for the predicative power of Google’s search volume index of firm names regarding stocks returns.
As such, this study is a relevant starting point for additional academic research. During crises, individual investors may change their trading behavior due to more volatile markets and a higher perceived risk. This could affect individual investors’ trading decisions by, for example, reducing their participation in the stock market. The results further indicate that the long-short portfolio performed better after the small bear market in 2011. A potential reason could be that although people used Google to gather investment information, they did not participate in the stock market because the market was more volatile and investors perceived a higher risk. As such, it could be interesting for future researchers to analyze the attention-induced short term price in more detail during a crisis.
31 References
Bank, M., Larch, M., Peter, G., 2011. Google search volume and its influence on liquidity and returns of German stocks. Financial Markets and Portfolio Management 25, 239–264.
Barber, B., Odean, T., 2008. All that glitters: The effect of attention and news on the buying behavior of individual and institutional investors. Review of Financial Studies 21, 785–818.
Barber, B., Odean, T., 2001. The internet and the investor. The Journal of Economic Perspectives 15, 41–54. Brooks, C., 2014. Introductory Econometrics for Finance. Cambridge University Press, Cambridge.
Brown, S. J., Warner, J. B., 1980. Measuring security price performance. Journal of Financial Economics 8, 205–258.
Brown, S. J., Warner, J. B., 1985. Using daily stock returns: The case of event studies. Journal of Financial Economics 14, 3–13.
Chemmanur, T. J., Yan, A., 2009. Advertising, attention, and stock returns, Working paper, Boston College and Fordham, University.
Choi, H., Varian, H., 2009. Predicting the Present with Google Trends. Working Paper, Google Inc. Da, Z., Engelberg, J., Goa, P., 2011. In search of attention. Journal of Finance 66, 1461–1499.
Ding, R., Hou, W., 2015. Retail investor attention and stock liquidity. Journal of International Financial Market, Institutions and Money 37, 12–26.
Fama, E. F., French K. R., 1993. Common risk factors in the returns on bonds and stocks. Journal of Financial Economics 33, 3–53.
Fama, E. F., MacBeth J. D., 1973. Risk return and equilibrium: empirical tests. Journal of Political Economy 81, 607–636.
Gervais, S., Kaniel, R., Mingelgrin, D., 2001. The high-volume return premium. Journal of Finance 56, 877–919.
Ginsberg, J., Mohebbi, M., Patel, R., Brammer, L., Smolinski, M., Brilliant, L., 2009. Detecting influenza epidemics using search engine query data. Nature 457, 1012–1014.
32 Grullon, G., Kanatas, G., Weston, J.P., 2004. Advertising, breath of ownership, and liquidity. Review of Financial Studies 17, 439–461.
Haamke, D., 2014. What is the effect of dividend change announcements on share prices after the financial crisis in the United States. University of Groningen.
Hausman, J, 1978. Specification Tests in Econometrics. Econometrica 46, 1251–1272.
Hausman, J. A., Taylor, W.E., 1981. Panel data and unobservable individual effects. Econometrica 49, 1377–1398.
Hirshleifer, D., Hong Teoh, S., 2003. Limited attention, information disclosure, and financial reporting. Journal of Accounting and Economics 36, 337–386.
Hoover, K.D., Siegler, M. V., 2008. The rhetoric of ‘signifying nothing’: a rejoinder to Ziliack and McCloskey. Journal of Economic Methodology 15, 57–68.
Ibbotson, R., Jaffe, J., 1975. “Hot issue” markets. Journal of Finance 30, 1027–1042.
Jain, P. C., Joh, G., 1988. The dependence between hourly prices and trading volume. Journal of Financial Quantitative Analysis 23, 269–283.
Joseph, K., Wintoki, M., Zhang, Z., 2011. Forecasting abnormal stock returns and trading volume using investor sentiment: Evidence from online search. International Journal of Forecasting 27, 1116–1127. Kahneman, D., 1973. Attention and effort. Prentice-Hall, Englewood Cliffs.
Keim, D. B., Madhavan, A., 1997. Transactions costs and investment style: an inter-exchange analysis of institutional equity trades. Journal of Financial Economics 463, 265–292.
Kramer, M. M., 2012. Financial advice and individual investor portfolio performance. Financial Management 41, 395–428.
Kulkarni, R., Haynes, K.E., Stough, R.R., and Paelinck, J.H.P., 2009. Forecasting housing prices with google econometrics. George Mason University School of Public Policy Research Paper No. 2009–10. MacKinlay, C., 1997. Event studies in economics and finance. Journal of Economic Literature 35, 13–39. Merton, R., 1987. A simple model of capital market equilibrium with incomplete information. Journal of Finance 42, 483–520.
33 Slovic, P., 1972. Psychological study of human judgement: Implications for investment decision making. Journal of Finance 27, 779–799.
Tetlock, P. C., 2007. Giving content to investor sentiment: The role of media in the stock market. Journal of Finance 62, 1139–1168.
Vlastakis, N., Markellos, R. N., 2012. Information demand and stock market volatility. Journal of Banking and Finance 36, 1808–1821.
