1. INTRODUCTION 3 2. LITERARY BACKGROUND. 5 2.1 ATTENTION AND PROXIES USED 5 2.2 EFFECT OF THE INTERNET ON INVESTOR ATTENTION 7 2.3 SEARCH VOLUME AND INVESTOR ATTENTION 8 2.3 ATTENTION AND STOCK LIQUIDITY. 9 3. METHODOLOGICAL APPLICATION. 11
3.1 GOOGLE SEARCH VOLUME DATA 11
3.2 COMMON STOCK LIQUIDITY DATA 15 3.3 CONTROL VARIABLES AND TRANSFORMING TO ENSURE STATIONARITY 17 3.4 CORRELATION AND CAUSALITY 19 3.5 BASIC MODEL 20 3.6 MODEL EXPANDED WITH INTRODUCTION OF LAGS 23 4. CONCLUSION 25
1. Introduction
Imagine a world where all the information that has ever existed is at your fingertips. A world where we can predict the future by pooling together the collective actions of millions of people. Where we know when the next outbreak of flu will happen because we have already seen an increased interest by the collective hive-mind in anti-flu remedies. A world where someone without a medical education can go into a doctors’ office having a reasonable diagnosis set him/herself. That world is the world we live in today. Unprecedented access to information and access to the thoughts of people are a matter of fact due to the rise of the Internet.
Another consequence of the Internet lies not in how information is available to everyone, but how the Internet creates information that can be used on its own. The fact that people online specifically search for certain information can be seen as a new source of information in itself entirely. Da, Engelberg, and Gao (2011) give an example of how this can be used in the real world. The US Centre For Disease Prevention has been using an increase in specific search terms (Flu Medicine, Headache, and Cough for example) in order to map upcoming flu epidemics weeks before their old forecast measures could map these. The fact that the Internet requires targeted searching in order to find the information any individual wants creates a dataset in itself on which research can be done. It can be hypothesized that this dataset can be seen as a proxy for attention. After all, I would only search online for something if it for some reason has grabbed my attention. As Bank, Larch, and Peter (2011) state: “an Internet user will only actively Google a specific keyword if he or she is interested in the object underlying the search term”. The insight this gives into the minds of people gives rise to the conclusion that if this dataset shows where the attention of people lies, then can we use this dataset in order to see in which companies investors are interested.
By using data collected from Google Search Trends of the search habits of people this thesis creates such a dataset, which acts as a proxy for investor attention. By investigating the relationship between attention and the liquidity of stocks this thesis uses behavioral finance to expand the literature with a relatively new dataset, and aims to broaden the research by adding in new markets with different levels of financial development and Internet adoption. The research conducted in this thesis aims to find out what the relationship is between investor attention and stock liquidity, as well as find out which way the causality of this relationship goes, and if the effects of attention might persist over a larger time period.
the methodological framework which will be used, 3) correlation and causality analysis, models and results, and lastly 4) concluding remarks.
2. Literary Background.
A large amount of research has already established the connection between the grabbing of attention and the consequent investment by investors in the stock that has grabbed his/her attention. What this basically means is that since the capabilities of the investor are limited, he/she reduces his/her set of possible investment opportunities to a smaller set than the several thousands out which he could potentially pick. Odean (1998), for example, demonstrates the existence of bounded rationality and the scarcity of cognitive capacity in individuals. In his 1998 paper: “Do investors trade too much” he finds that investors limit their search for investment opportunities to stocks that have recently grabbed their attention. Having established the importance of where investor attention lies in terms of investor decisions in where to invest, Odean then, together with Barber (Barber & Odean, 2008), look at how we can measure the concept of investor attention. As a proxy for investor attention they look at newspaper headlines, unusual trading volumes, and extreme returns. They find that within these attention-grabbing factors investors are even willing to let the preferences they have with regard to what they want to invest in slide, in favor of these stocks that have grabbed their attention. Using “paper media” to measure where investor attention might lie proves to be a topic of interest to the scientific community. Fang and Peres use, in similar style, newspaper articles published as a proxy for which stocks have grabbed the attention of the investor. Furthermore, Tetlock (2007) uses a well-read popular financial column published by the financial times. In his research he quantifies the sentiment in these columns and creates what he calls a pessimism index to quantify the sentiment among investors.
that firms with greater advertising expenditure, ceteris paribus, have a larger number of both individual and institutional investors. Moreover they find this therefore leads to a better liquidity of their common stock.
2.2 Effect of the Internet on investor attention
attention of the investor. For their proxy for attention they use the frequency of edits of Wikipedia pages. Wikipedia, being the crowd-sourced online encyclopedia, has the unique quality of allowing everyone to create, edit, and contribute to its pages. Rubin and Rubin (2010) look at the intensity of editing of the pages of certain publically traded firms to uncover a pattern of firms that have the attention of investors and find significant results in explaining movement on the stock market using this proxy for attention.
2.3 Search volume and investor attention
with but different from existing proxies for investor attention, captures investor attention in a more timely fashion, [and] likely measures the attention of retail investors”. They further state that “existing measures of investor attention […] are indirect proxies, Google search volume is a direct proxy for investor attention.” One of their findings is that an increase in the SVI (Search Volume Index) leads to higher stock prices in the next two weeks and eventually leading to a price reversal. Lastly, and to this authors’ knowledge the sole instance where research using a search volume dataset has been conducted in Europe, is the paper “Google search volume, and its influence on liquidity and returns of German stock.” This paper by Bank, Larch, and Peter (2011) finds that an increase in search volume is associated with a rise in trading activity and stock liquidity. Moreover, they find that an increase in search queries is related to a temporary increase in stock returns. This finding is similar to what Da, Engelbert, and Gao (2011) find in their study of American stocks.
One interesting observation, which has to be noted and is made in connection to the previous research, is that a few of these studies note that the effect of the Internet might primarily better capture the attention of individual investors as opposed to institutional investors. This observation also makes sense intuitively since institutional investors after all have access to a larger amount of resources, such as an increased cognitive capacity and reduced bounded rationality due to more human resources and access to more sophisticated databases and algorithms. Nonetheless the research shows that datasets pooled from the Internet have proven to be reliable when it comes to mapping investor attention.
2.3 Attention and stock liquidity.
rate is relatively easy to obtain which allows for the examination of liquidity effects across a large number of stocks over a longer period of time.
3. Methodological application.
The previous discussion has made a compelling argument as to the importance of knowing where the attention of investors lie, and the application of new Internet based datasets on finding investor attention. Combining this interest in the subject of investment attention with the model of Datar, Naik, and Radcliffe (1998) to prove the presence of investor attention and show the effect on common stock liquidity has led to the following research question to be answered in this thesis: “Can attention, as measured by Google Search Volume, be used as an adequate explanatory variable in variations of common stock liquidity.” Moreover, by using a dataset covering 4 different European countries, namely: 1) Spain, 2) Germany, 3) United Kingdom, 4) Poland, this research expands on existing research by introducing a European perspective, which leads to a perspective in terms of cultural differences, difference in internet adoption rates, and robustness across countries. The latter is especially important since, with the exception of the paper by Bank, Larch, and Peter (2011) all research using search volume datasets has been conducted in the United States.
3.1 Google Search Volume data
The first step in answering this question is the retrieval of the dataset to be used in this research. To measure attention I will use a dataset of Google Search Volume retrieved from the publically accessible database Google Trends1. This
Markellos and ensures sufficient search volume in order to prevent terms with insignificant amounts of observations in the dataset. Moreover, these companies and countries chosen represent a wide variety of both industries, as well as cultural differences, and differences in the home countries’ Internet adoption rate. The time period used is October 2005 – January 2015 giving a total dataset of just shy of 20,000 observations. After deletion of companies that have incomplete observations I am left with a dataset of 33 companies, with complete observations for the entire time period. The description of the firms can be seen in table 1.
All the search volume has been geo-located to the specific country on which market the firm is traded. This means, for example, in the case of Siemens only search queries from within Germany are taken. This allows for the data to be used in more than one way, for example in making trans-national comparisons, as well as preventing ambiguity by eliminating the chance of the company name having a different meaning in another language. In this I follow most previous research, most notably the geo location of Google Search Volume data by Bank, Peter, and Larch (2011). Furthermore, Google provides a so-called relative popularity index. This makes the data better suited to be compared on a weekly basis. Google describes the data transformation they apply to the search volume as follows: “each data point is divided by the total searches of the geography and the time range it represents, to compare relative popularity. The resulting numbers are then scaled to a range of 0 to 100”2. The reason this data
Company Abbreviation Country of Origin Industry Company Abbreviation Country of Origin Industry Acciona ACC Spain Construction Iberdrola IBE Spain Electric Utility Asseco ASO Poland Information Technology Inditex IDX Spain Retail Barclays BAR United Kingdom Banking Indra Sistemas IND Spain Information Technology British
American
Tobacco BATO
United
Kingdom Tobacco KGHM KGHM Poland Mining Bayer BAY Germany Pharmaceuticals LPP LPP Poland Retail BBVA BBVA Spain Banking Mapfre MAP Spain Insurance Beiersdorf BEI Germany Personal Care Mbank MBK Poland Banking Boryszew BOR Poland Automotive/Chemicals Merck MRK Germany Pharmaceuticals BP BP United Kingdom Oil and Gas Repsol REP Repsol Oil and Gas Commerzbank COM Germany Banking Rio Tinto RT United Kingdom Mining Continental CON Germany Automotive Sacyr SAC Spain Construction Daimler DAIM Germany Automotive Siemens SMS Germany Engineering Diageo DIA United Kingdome Beverages Synthos SYN Poland Oil and Gas Deutsche
Telekom DT Germany Telekom Telefonica TEL Spain Telecom Eurocash EC Poland Retail ThyssenKrupp TK Germany Engineering GlaxoSmithKline GSK United Kingdom Pharmaceutical Vodafone VODA United Kingdom Telecom HSBC HSBC United Kingdom Banking
Table 1. Description of companies used, home countries, and industries.
standard deviation of the Google Search Volume data, ranging from 7.02 for Sacyr to 25.77 for Eurocash. Furthermore, we can see most of the Jaque Bera statistic being significant at the 1 percent level indicating non-normality of the dataset. Because the following research will make use of correlation analysis and linear regression, neither of which require normality I do not find any issue with this.
3.2 Common stock liquidity data
The dependent variable in the model, which will follow later, will be based on the turnover rate as measure of liquidity as proposed by Datar, Naik, and Radcliffe. The reasons for using this method over the bid – ask spread must have become clear from the discussion in the literature review, but essentially comes down the ease of obtaining correct information, and the proven correlation between liquidity and trading frequency. The data needed has all been extracted from Thomson Reuters Datastream3. Some transformations had to be made, however,
to create the turnover variable as described by Datar, Naik, and Radcliffe. Firstly, I returned Turnover by Volume for my updated dataset of 33 companies. In order to transform this I then returned Total Shares outstanding and Percentage of free floating. Turnover by volume gives an absolute number of the shares traded for the week. Total shares outstanding returns the total shares each firm has outstanding in a particular week. And lastly, Percentage of free floating represents “the total amount of shares available to ordinary investors, expressed as a percentage of total number of shares outstanding.” In order to transform this to the turnover rate the following formula has been used: 𝑇𝑅 = 𝑇𝑢𝑟𝑛𝑜𝑣𝑒𝑟 𝑏𝑦 𝑉𝑜𝑙𝑢𝑚𝑒 𝑇𝑜𝑡𝑎𝑙 𝑆ℎ𝑎𝑟𝑒𝑠 𝑂𝑢𝑡𝑠𝑡𝑎𝑛𝑑𝑖𝑛𝑔 ∗ % 𝑜𝑓 𝐹𝑟𝑒𝑒 𝐹𝑙𝑜𝑎𝑡𝑖𝑛𝑔 The descriptive statistics for the turnover rate variable are available in table 3. The standard deviation of the turnover rate also varies greatly ranging from 0.0074 for GlaxoSmithKline to 0.3504 for Synthos. Furthermore, we observe the entire dataset being non-normal by the rejection of the Jaque Bera hypothesis in every case. Similar to the Google volume data this will not lead to issues due to normality not being required in the research performed. Lastly the correlation between the unmodified dataset can be found in appendix 1.
exacerbated by the increased market share of Google as a search provider over the time period of the dataset. In order to ensure non-stationarity will not be an issue in this research I test for stationarity in the variables for each of the 33 companies researched. An Augmented Dickey Fuller test has been employed to do this research, the results of which can be found in appendix 2. In the Augmented Dickey Fuller test evidence is found for non-stationarity in both of the aforementioned variables. This leads to the transformation of the variables as discussed in the next section. 3.3 Control variables and transforming to ensure stationarity A linear regression will be run in order to determine the relation between GSV and TR. Dependent in the model will be the TR. In order to ensure robustness of the model several control variables have been devised which are known to significantly affect stock liquidity. Firstly, a variable to control for week-to-week variation in the trading price of stock is created by the following formula: 𝑅! = 𝑆ℎ𝑎𝑟𝑒𝑝𝑟𝑖𝑐𝑒! 𝑆ℎ𝑎𝑟𝑒𝑝𝑟𝑖𝑐𝑒!!! − 1
Where 𝑅! is the increase or decrease in share price from one week to another. The share prices, on a weekly interval, are returned from Datastream. Furthermore, another ratio that significantly influences liquidity is the market to book ratio, which has been used as a control variable for stock liquidity either in the form of market to book, or book to market, in a plethora of studies on the topic. The market to book ratio measures the value of the accounting balance equity to the market value of equity. In essence a market to book (MTB) ratio over 1 means a company is relatively overvalued by the market compared to the book value of equity, a MTB ratio of less than 1 means a company is relatively undervalued by the market compared to the book value of equity. The following equation therefore has been used to create the variable:
𝑀𝑇𝐵! = (
𝑀𝑎𝑟𝑘𝑒𝑡 𝑉𝑎𝑙𝑢𝑒!
𝑇𝑜𝑡𝑎𝑙 𝐴𝑠𝑠𝑒𝑡𝑠!− 𝑇𝑜𝑡𝑎𝑙 𝐿𝑖𝑎𝑏𝑖𝑙𝑖𝑡𝑖𝑒𝑠!)
Where 𝑀𝑇𝐵! is the Market to Book ratio in a specific week based on the DataStream data for Market Value, Total Assets, and Total Liability values for that specific week. Lastly the model controls for the company’s earnings in a specific week. In order to ensure weekly earnings I use the EPS (earnings per share) variable from DataStream, which is a time-series created by taking the earnings in the year prior and up to the current week, then divided by the shares outstanding.
3.4 Correlation and causality A first step towards proving the relation between attention, as measured by GSV, and liquidity, as measured by TR, is by creating a correlation table with the new stationary variables, the results of which can be found in table 4. The values in the correlation matrix range from -0.0658 for Vodafone to 0.5283 for Merck. As expected to majority of the values are positive, with 31 out of potentially 33, which would indicate the amount of attention a company receives has a positive effect on the liquidity of that company. Creating a correlation for the entire dataset leads to an overall correlation between GVI and TRI of just over 10%.
Company Correlation Company Correlation ACC 0.1835 IBE 0.2439 ASO 0.0582 IDX 0.1837 BAR 0.1994 IND 0.0289 BATO 0.1394 KGHM 0.3255 BAY 0.2668 LPP 0.0213 BBVA 0.1054 MAP 0.1806 BEI 0.2621 MBK 0.1799 BOR 0.0157 MRK 0.5283 BP 0.5154 REP 0.0929 COM 0.122 RT 0.4457 CON 0.2606 SAC 0.4315 DAIM 0.4922 SMS 0.3658 DIA 0.2813 SYN 0.1395 DT 0.3637 TEL 0.1762 EC -0.012 TK 0.4443 GSK 0.1008 VODA -0.0658 HSBC 0.1462 OVERALL 0.1002 Table 4. Correlation matrix between GVI and TRI.
attention, before engaging in potential purchase. To proof the direction of causality between GVI and TRI I undertake a Granger Causality test. Since I theorize the attention towards a certain company is relatively short–term and therefore the effects which attention has on the liquidity of a stock is relatively short-term as well I have opted to use a relatively short-term lag of 4 weeks for the Granger Causality analysis. From the results of the causality analysis, which can be found in appendix 3, it becomes clear that causality does not just flow one direction. Of the 33 companies, at the 10% significance level, 7 have causality run from TRI to GVI (21%), 10 have causality run both ways (30%), and 9 have causality run from GVI to TRI (27%). Overall this means just over 78% of the companies in the dataset have causality running one way or the other between GVI and TRI. In general this causality analysis shows there is indeed statistical support for the theory that investors tend to buy stocks after an increased amount of attention, it is, however, hard to make solid inferences from this data. The reason for the bidirectional causality and the causality from TRI to GVI might lie in the attention gathered by abnormal amounts of market activity as found in the paper by Barber and Odean, where they use abnormal trading volume as one of the proxies for attention. Since abnormal turnover has been used as a proxy for attention, it seems like a plausible explanation that attention as measured by GVI and liquidity as measured by TRY interact with each-other in a bidirectional way. Furthermore, it seems plausible that, albeit on a smaller scale, investors might want to find information after the purchase of a stock to make inferences for future returns.
3.5 Basic model
upon in a later section with the addition of several lags to the GVI variable, it seems natural to include all control variables that might have an effect on stock liquidity. As mentioned these control variables are, Market to Book ratio, Earnings per share, and week-to-week price variations. Incorporating these control variables into the equation with GVI and TRI gives the following model on which regressions are run:
𝑇𝑅𝐼 = 𝛼 + 𝛽!𝐺𝑉𝐼 + 𝛽!𝑅 + 𝛽!𝑀𝑇𝐵𝐼 + 𝛽!𝐸𝑃𝑆𝐼 + 𝜖
The models are run with all variables present in each of the 33 company regressions. The output of the model can be seen in table 5. Overall the statistical relevance of the models as measured by the adjusted R^2 varies significantly from 0% to about 34%. It is interesting to note that the models with the lowest explanatory value seem to correlate with the models in which the GVI has the least amount of significance. Overall the GVI is significant at the 1% level for 22 out of 33 companies (66%), at the 5% level for 5 out of 33 companies (15%), and at the 10% level for 1 out of 33 companies (3%) meaning for any relevant level of statistical significance it scores 84% of the companies. With regards to the GVI it is also interesting to note that all 3 of the negative observations have insignificant coefficients, therefore, all observations that are significant are positive. This is in line with the theorized concept of investors’ attention leading to an increase in liquidity. Moreover, the observation that all significant coefficients are positive leads to the conclusion that investors are more likely to buy stock in case of a company having grabbed their attention. Furthermore, for the control variables R is significant at at least the 10% level for 19 out of 33 companies (58%), MTBI for 17 out of 33 companies (52%), and EPSI for 14 out of 33 companies (42%). An interesting observation that therefore can be made is that GVI overall is the variable with the most significance in the models explanatory power of the dependent variable TVI. Although there is great variability within the explanatory power of the models the economic conclusions that can be drawn from these models are fairly significant.
However, the model seems to have some variation in the amount of explanatory power it has with regard to the market on which the firm is traded. It seems that in particular the Polish companies seem to have the least amount in both explanatory power by the model, as well as the least amount of significance coming from the attention proxy GVI. This might be related to the relatively lower developed status of the financial market in Poland, as well as the lower level of Internet adoption in Poland. However, any well-structured theorization for exactly why this model seems to have less significance for the Polish firms is beyond the scope of this thesis and is recommended to be a topic for future research.
3.6 Model expanded with introduction of lags
of the explanatory value of the models based solely on GVI lag. Hence, the formula for the lagged model is as follows:
𝑇𝑅𝐼 = 𝛼 + 𝛽!𝐺𝑖 −1 + 𝛽!𝐺𝑖 −2 + 𝛽!𝐺𝑖 −3 + 𝛽!𝐺𝑖 −4 + 𝜖
significance, and the low explanatory value of the lagged models indicates that this observation is unfortunately of little real world use.
4. Conclusion
In this thesis I set out to find the effect attention has on stock liquidity. To create the models this research has made use of a relatively new dataset to use as a proxy for attention. In doing this, and finding significant results, this research has lent credibility to the use of Google Search Volume data as a proxy for investor attention. This thesis has expanded research already conducted within this area with evidence from a new dataset spanning more industries and a more varied amount of geographic locations, as well as a larger dataset than research on this topic has done before. In the concluding section of this thesis I will look back on the results and the inferences that can be made from the models. 4.1 Basic model, causality, and correlation Both the correlation matrix and the basic model have shown the positive relation between investor attention, as measured by Google Search Volume, and common stock liquidity, as measured by Datar, Naik, and Radcliffe’s turnover rate. Although it is hard to make proper inferences about the direction of causality due to the finding of bi-causality I believe the finding of at least some level of causality flowing from Google Search Volume to liquidity, supported by the findings of the basic model, gives sufficient proof that at least some level of causality can be established. By adjusting the variables to ensure stationarity this research has provided statistically significant and robust results to prove Google Search Volume can be an interesting measure for both institutional as well as individual investors looking to incorporate a new measure of investor attention into their liquidity models. It provides a basis on which both researchers and investors alike can expand by using this Google Search Volume attention proxy in their forecasting models.
4.2 Lagged model and final remarks
The incorporation of several lags into the model has shown an interesting reversal effect of investor attention on stock liquidity, which can plausibly be explained by the high significance of the positive effect attention has to stock liquidity in the same trading week, as shown by the basic model. Meaning investors have already acted upon the attention previously. Unfortunately, the explanatory value of the model was, although statistically significant, not sufficient to make any valuable economic conclusions.
5. Limitations and recommendations for future research
This final section will look at the limitations of this thesis as well as provide guidance for future research expanding on this new dataset on investor attention
5.1 limitations
6. References
Amihud, Y., 2002. Illiquidity and stock returns: cross-section and time-series effects. Journal of Financial Markets, 5(1), 31-56.
Amihud, Y. & Mendelson, H., 1996. Asset pricing and the bid-ask spread. Journal
of Financial Economics, 17(2), 223-249.
Antweiler, W. & Frank, M., 2004, Is all that talk just noise? The information content of Internet stock message boards. The Journal of Finance, 59(3), 1259-1294.
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(3), 239-264.
Barber, B. & Odean, T., 2001, The internet and the investor. The Journal of
Economic Perspectives, 15(1), 41-54 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(2), 758-818.
Bollen, J., Mao, H., & Zeng, X., 2011. Twitter mood predicts the stock market.
Journal of Computational Science, 2(1), 1-8
Chordia, T., Roll, R., & Subrahmanyam, A., 2001. Market liquidity and trading activity. The Journal of Finance, 56(2), 501-530
Da, Z., Engelberg, J., & Gaeo, P., 2011. In search of attention. The Journal of
Finance, 66(5), 1461-1499.
Datar, V., Naik, N., & Radcliffe, R., 1998. Liquidity and stock returns: An alternative test. Journal of Financial Markets, 1(2), 203-219.
Fama, E., 1965. The behavior of stock-market prices. The Journal of Business, 38(1), 34-105. Fama, E., & French, K., 1988. Dividend yields and expected stock returns. Journal of Financial Economics, 22(1), 3-25. Fang, L., & Peress, J., 2009. Media coverage and the cross-section of stock returns. The Journal of Finance, 64(5), 2023-2052. Grullon, G., Kanatas, G., & Weston, J., 2004. Advertising , breadth of ownership, and liquidity. Review of Financial Studies, 17(2), 439-461.
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(4), 1116-1127.
Kristoufek, L., 2013. Can Google Trends search queries contribute to risk diversification?, Scientific Reports, 3.
Merton, R., 1987. A simple model of capital market equilibrium with incomplete information. The Journal of Finance, 42(3), 483-510.
Rubin, A., & Rubin, E., 2010. Informed investors and the Internet. Journal of
Business Finance & Accounting, 37(7-8), 841-865.
7. Appendices
Company
Granger
