1. Introduction
In this paper, we use Google search volume (Da, Engelberg and Gao (2011)) as proxy of retail investors’ attention to study the dynamic relationship with stock market volatility, test whether it can add more information for modeling volatility, examine if it can help to forecasting volatility in- and out-of-sample per country and compare these phenomenon in different markets.
In stock markets, huge movements catch investors’ eyes. The model of Lux and Marchesi (1999) implies that volatility triggers search activity. And Merton (1987) establishes that investor attention may be relevant for stock pricing and stock liquidity.
When the attention of investors increases, this may indicate trading activity increases.
Many paper document that retail trades can make stock prices move (Kumar and Lee (2006), Dorn, Huberman and Sengmueller (2008), Kaniel, Sear and Titman (2008), Hvidkjaer (2008)). And Foucault, Sraer and Thesmar (2011) prove that trading activities made by retail investors are positively related to volatility, which can be regarded as behaviors of noisy traders. They estimate that volatility is driven by retail investors about 23% except fundamentals. Therefore, abnormal volatility attracts retail investors’ attention and then causes retail investors invest in, which in turn makes volatility.
Nevertheless, measuring retail investors’ attention is a hard work since it cannot be observed directly. In empirical studies, many proxies for attention have been employed, like published news announcements and headlines (Mitchell and Mulherin (1994), Berry and Howe (1994), Frieder and Subrahmanyam (2005), Barber and Odean (2008) and Yuan (2008), Fang and Peress (2009)), trading volume (Gervais, Kaniel, and Mingelgrin (2001), Barber and Odean (2008), Hou, Peng, and Xiong (2008)), advertisement expense (Grullon, Kanatas, and Weston (2004), Lou (2008),
2
Chemmanur and Yan (2009)), price limits (Seasholes and Wu (2007)) and extreme returns (Barber and Odean (2008)). But using these proxies need critical assumption that if stock’s name is mentioned or its return or turnover is extreme, then that indicate retail investors must pay attention to it, while this assumption cannot be guarantee in practice.
Internet search volume is proved to be a more direct and easier method to measure retail attention by Da, Engelberg and Gao (2011), who use search volume of stock tickers from Google Trends. They show that this method is timelier than other well-established attention proxies and mainly captures the retail investor attention.
This method seems to be adequate for two reasons. First is that, nowadays, internet has become a popular way to search for information for individual investors. Since the usage of internet increased steadily worldwide during the recent decades, World Wide Web became accessible by nearly everyone and everywhere. And it is the largest pool which supply available information, freely or costly. Internet user usually choose search engine to seek information when needed. Second, an internet user will actively search a specific word only if he or she has interest in or demand for information about the object underlying the keyword.
Google search volume is the most popular proxy since Google search engine has the largest worldwide market shares, accounted for about 90.7% of all search engine on 2011.1 Google search volume has significantly positive effects on trading activity, trading volume, stock liquidity and return volatility, both historical and implied (Vlastakis and Markellos (2010), Bank, Larch and Peter (2011)).
In addition, there are several evidences that internet search volume has power to forecast, such as unemployment rates, home sales, automotive sales (Choi and Varian
1 Source: StatCounter Global Stats (http://gs.statcounter.com/)
3
(2009)) and influenza (Ginsberg et al. (2009)). In the financial field, Google search volume is documented to predict earnings (Da et al (2010a), Drake, Roulstone and Thornock (2011)), abnormal returns and trading volumes (Joseph, Wintoki, Zhang (2011)). Da et al. (2011) report that an increase in Google search volume predict higher stock prices in the short-run and reversals in the long run, which is consist with the attention theory of Barber and Odean (2008).
Nowadays is the era of globalization. Investors use international portfolio to diversify risks and increase profits. Besides to invest in developed markets and emerging markets, more and more investors like to invest in frontier markets. It is proved that when portfolio contains equities of frontier markets, both portfolio risk and returns can be improved (Jayasuriya and Shambora (2009)). In this point of view, we want to comprehensively explore not only the relationship between retail investors’
attention measured by internet search volume and the stock market volatility but also the predict power of search volume for volatility forecasting in different markets with different development levels.
From the view of international portfolio, we’d like to use the same internet search engine, Google search volume index, measuring worldwide attentions of retail investors for three different MSCI indices, developed, emerging and frontier markets (DM, EM, FM) index, to test if search volume can increase different index volatility forecasting power. However, this method cannot be used since Datastream doesn’t provide these three MSCI indices’ intraday high and low prices, which are needed for realized volatility, and Google Trends also have not enough search volume data of MSCI frontier markets index.
Moreover, in stock markets, the main retail investors usually are local residents not foreigners. Instead, we focus on leading indices of each country, which belongs to
4
markets of various development levels according to MSCI Market Classification.2 And we should choose the internet search volume whose search engine has the highest market shares in its home country to measure local attentions of retail investors. In almost countries, Google is the most popular search engine. Baidu and Naver is leading search engine in China and South Korea respectively. But there are problems that they either have no English version or do not provide detail search volume data to be downloaded. Therefore, we use Google search volume to measure local attentions for each country’s leading index and then test if search volume can improve volatility forecasting in different markets.
At first, we estimate a VAR model for every stock index to capture the dynamic relationship between Google search volume and stock index volatility. And then examine if past volatility can significantly influence present search volume (Granger (1969) and Sims (1972)) by Granger causality tests, see how volatility reacts over time to shock of search volume, and vice versa, by impulse response function and test how much of volatility can be explained by internet search volume by long-run variance decomposition under the VAR model. Next, we use three other regression models, AR(1), HAR and EGARCH, to rule out whether search volume has additional information for modeling volatility. Last, we compare the forecasting ability of the volatility models with and without lagged search volume in- and out-of-sample by using the mean squared error (MSE), the quasi-likelihood loss function (QL) and the R2 of regression of the actual realized volatilities on their prediction.
We find past search volume is useful to predict future volatility generally and half of countries’ Granger causality is bi-directional: high search activities follow high volatility, and high volatility follows high search activities. But, when there is a
2 Source: http://www.msci.com/products/indices/market_classification.html
5
positive shock of search volume, volatility wouldn’t react immediately but have positive movement later while volatility can affect search volume immediately.
Throughout all countries, movement of volatility is contributed by search volume is ranging from 0.11% to 20.53%. As consistent with Foucault, Sraer and Thesmar (2011), search volume adds valuable information for modeling volatility and influences future volatility positively. Search volume also can improve volatility forecasting in- and out-of-sample. But it becomes much more insignificantly in out-of-sample forecast evaluation.
As the developed level of markets is lower, the phenomenon that search volume can help to forecast volatility becomes less obvious. Besides the developed level of markets, there are some possible reasons of why this phenomenon can’t be seen from our tests and models in some countries. The proper reasons are lower frequency of data, less univocal search terms, lower market shares of Google, location of countries, smaller penetration rate of internet users and lesser market shares of retail investors.
The remainder of this paper is organized as follows. In Section 2, we describe the search volume data, data set of realized volatility and the statistics. Section 3 explains the method, models and tests that we use where section 3.1 studies the dynamic relationship between Google search volume and stock index volatility, section 3.2 examine whether the search volume can add valuable information to different volatility models and section 3.3 evaluates in- and out-of-sample volatility forecasts to examine if search volume can help to forecast future volatility. Section 4 is the results of tests and modeling. Finally, section 5 concludes.
6