theory to support the use of trading volume as an investors’ sentiment index and found that trading volume is a suitable proxy as sentiments. So, trading volume reflects investors’ expectation of the stock prices; hence influences the average returns of the stocks in the Taiwan stock market. Another work is by Chung and Yeh (2009), who did research on the U.S. stock market using many sentiment indexes, such as consumer confidence level1, the VXO2 (the old VIX (Volatility Index)), Baker and Wurgler's orthogonal sentiment index3 to capture consumers’ and investors’ sentiments. They reported that sentiments could be used to predict the stock returns in the U.S. stock market. Our research uses J.P. Morgan confidence indexes as investor’s interview sentiments and also found that two particular investors’ interview sentiments can help predict the average returns of TAIEX.
3. Summary
Based on the established behaviour finance research findings, we have identified several investors’ psychological factors that might have an impact on the average returns of the Taiwan stock market and will incorporate them in our empirical study.
These factors include investor’s interview sentiments, their attention, herding,
1 http://www.sca.isr.umich.edu/
2 http://www.cboe.com/VXO
3 http://people.stern.nyu.edu/jwurgler/
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
12
overconfidence and over optimism. The investors’ attention and herding are reflected on the Google SVI. The investors’ overconfidence and over optimism are reflected on the J.P. Morgan data of Taiwan Stock Price Index. Moreover, we include trading volume in our study to investigate if the past performance of a stock has an impact on its present performance.
We will explain these data and the research methods used in this empirical study in the following chapter.
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
13
Research Methods
To investigate how investors’ attention and interview sentiments influence the average returns of TAIEX, we selected eight independent variables and one dependent variable from three different data sources. Section 1 describes the data from the Taiwan Economic Journal; Section 2 explains the data from the J.P. Morgan Asset Management, and Section 3 discusses the search volume data from Google Trends. In section 4, we explain the econometric method we choose to run the regression. Section 5 summaries the whole chapter.
1. Taiwan Economic Journal Data
Taiwan Economic Journal (TEJ)4 is a database that contains historical financial data and information in the major financial markets in Asia. We downloaded weekly average opening price of TAIEX and weekly total trading volume from January 5, 2014 to November 6, 2016 for this research.
2. J. P. Morgan Asset Management Confidence Indexes
We obtained six confidence indexes form J.P. Morgan Asset Management5. J.P.
4 http://www.tej.com.tw/twsite/
5hhttps://www.jpmrich.com.tw/wps/portal/!ut/p/b0/04_Sj9CPykssy0xPLMnMz0vMAfGjzOK9AkIDjEJc DQz83XycDIxczIyd3TzcDAycTfULsh0VAXrBmJ0!/?WCM_PORTLET=PC_Z7_JPUP2TE008CAC02
‧
Morgan Asset Management started investigating Taiwan investors’ sentiments change in 2004. They interview investors from time to time by asking six questions. The interviews are designed to evaluate investors’ confidence in Taiwan economics, politics, and stock market. The results are used to compose six interview sentiment indexes.
Moreover, they set the threshold used to evaluate these indexes as 100. If an index is higher than 100, it means that investors are optimistic about the market. The higher the score is, the more optimistic the investors are. If an index is below 100, it means the investors are pessimistic about the market. The lower the score is, the more pessimistic the investors are. The six confidence indexes are defined as follows:
1. Taiwan Stock Price Index: What is the possibility that TAIEX rises in the future?
2. Taiwan Economic Situation Index: What is the possibility that Taiwan economic situation becomes better in the future?
3. Taiwan Political Index: What is the possibility that the political situation between Taiwan and China becomes more stable in the future?
4. Taiwan Investment Environment Index: What is the possibility that the investment situation in Taiwan becomes better in the future?
DL2KSHG28U1000000_WCM&WCM_GLOBAL_CONTEXT=/wps/wcm/connect/jpmrich/eportal/lb_b 2c_l01/b2c_l01p220/b2c_l01p220_01/at_b2c_l01p220_01-1_00002
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
15
5. Global Economic Index: What is the possibility that the global economic situation becomes better in the future?
6. Possibility of the Value of Your Portfolio Increases within next six months:
What is the possibility that the value of your portfolio increases in the future?
We downloaded the confidence indexes data from 2014 to 2016. The data were mostly quarterly. To work with other data in weekly format, we used frequency conversion method by Litterman (1983) to raise the frequency of these interview sentiment indexes to become weekly using the statistical software Eviews.
The method assumes that an interview sentiment index this week will affect the index next week and their residuals are correlated. These functions are described as follows:
𝑥𝑥𝑖𝑖𝑖𝑖 = 𝑥𝑥𝑖𝑖𝑖𝑖−1+ 𝜀𝜀𝑖𝑖𝑖𝑖, ε𝑖𝑖𝑖𝑖 = ρε𝑖𝑖𝑖𝑖−1+ e𝑖𝑖𝑖𝑖, where ε𝑖𝑖𝑖𝑖~𝑁𝑁(0, V) and 𝑖𝑖 = 3, 4, 5, 6, 7, 8.
We used 𝑥𝑥3𝑖𝑖~𝑥𝑥8𝑖𝑖 to represent the six interview sentiment variables. The variable
𝑥𝑥1𝑖𝑖 was reserved for SVI while the variable 𝑥𝑥2𝑖𝑖 was reserved for trading volume. The initial condition is 𝑥𝑥𝑖𝑖0 = 0. The function is an ARIMA (1,1) model. The last weekly data series was November 6, 2016.
‧
Google Trends6 is a public web facility of Google Inc. (original Google Search) that shows how often a particular search-term (keyword) is entered, relative to the total
search volume across various regions of the world, and in various languages. The SVI
values represent search interest relative to the given region and time. The highest search
number during the downloading period is given SVI value of 100. The weekly SVI is
calculated by dividing the weekly search volumes with the highest search volume
assigned SVI value 100. We can find the details in Google Trends help7.
We downloaded two sets of SVI data from Google Trends using two sets of search terms. The first set consisted of the company names of the all companies8 listed in TAIEX and the second set contained the ticker symbols of these companies. The following subsections explain the procedures to obtain these data.
3.1 Company Names
A company listed in TAIEX has a full name. However, in the stock market, investors often call these companies by their abbreviated company names. For example,
“台積電” is the abbreviated company name of “台灣積體電路製造股份有限公司”.
‧
For brevity, we will use company names instead of abbreviated company names in the rest of the thesis. We retrieved the SVI in the Taiwan region using the company name of each company listed in the TAIEX as the search term from January 5, 2014 to November 6, 2016. However, we found some small capital companies do not have any SVI information. Also, some company names are common terms that may be used to conduct Google search by non-investors. For these two kinds of company, we replaced the search results with that obtained using ticker symbols. The total number of companies whose SVI have been replaced under the process was 49.
To sum the search volume data up as a single index, we used a weighted sum approach, where the weight was the company size, represented by their relative percentage of market value on November 18, 2016, downloaded from Taiwan Futures Exchange9. This approach is based on the following assumptions:
• Each search volume is independent. Increased attention on one stock will not influence others.
• The higher a company’s market value is, the more attention the company receives and hence the higher the search volume.
• The companies that constitute TAIEX remain unchanged.
After we summed up the weighted SVI of all companies, the time series contained
9 http://www.taifex.com.tw/chinese/9/9_7_1.asp
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
18
146 weeks of data.
3.2 Ticker Symbols
A company listed in TAIEX also has a ticker symbol. For example, “2330” is the ticker symbol of “台積電”. We first used ticker symbols to retrieve their SVI from Google Trends. Next, we used the similar procedures described in the previous section to obtain the weighted SVI of all companies. The time series also had 146 weeks of data.
4. Newey-West Correction of Standard Errors
Newey-West correction of standard errors method10 is a method to estimate the coefficients of a linear regression model applied to time series data. It is used to correct autocorrelation (also called serial correlation) and heteroskedasticity in the error terms in the regression model. We used the statistical software SAS to run the regression.
Below are the variables names and their meaning.
y：average returns of TAIEX 𝑥𝑥1: Google SVI
𝑥𝑥2: trading volume of TAIEX 𝑥𝑥3: Taiwan Stock Price Index
10 Please refer to the appendix to know the details.
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
19
𝑥𝑥4: Taiwan Economic Situation Index 𝑥𝑥5: Taiwan Political Index
𝑥𝑥6: Taiwan Investment Environment Index 𝑥𝑥7: Global Economic Index
𝑥𝑥8: Possibility of the Value of Your Portfolio Increases within six months We first converted all time series data into log10, and then used the difference between adjacent weeks ( 𝑥𝑥𝑖𝑖𝑖𝑖− 𝑥𝑥𝑖𝑖𝑖𝑖−1) to run the regression. In this way, the interpretation of the regression is easier: 1 percentage change of an independent variable will change a certain percentage of the dependent variable, specified by the coefficient of the independent variable in the regression model. Moreover, we can reduce the scale difference of the variables; hence increase predicting accuracy.
The linear regression model is as follows:
y𝑖𝑖 = 𝛽𝛽0+ ∑8𝑖𝑖=1𝛽𝛽𝑖𝑖𝑥𝑥𝑖𝑖𝑖𝑖+ e𝑖𝑖, and where t = 1, 2, ... ,146
We used the regression results to estimate what variables are significant in effecting the average returns of TAIEX and how long the effect lasts. Because the residuals had heteroscedasticity and autocorrelation, the method used HAC (Heteroskedasticity and Autocorrelation Consistent) estimators to correct them. We
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
20
used Parzen kernel, an HAC estimator, to run the regression. The relationship between Newey-West correction of standard errors method and Parzen kernel is shown in the appendix.
5. Summary
In the chapter, we have explained the data sources and the methods we used to conduct our research. We downloaded the data from TEJ, Google Trends, and J.P.
Morgan Asset Management to help us test different behaviour hypothesis. After that, we used the frequency conversion method to process J.P. Morgan Asset Management data and the weighted sum approach to process Google Trends SVI data. Finally, we applied Newey-West correction of standard errors method to run the regression on the processed data. We present our results with analysis in the following chapter.
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
21
Research Results and Analysis
This chapter presents the results and provides our analysis. In Section 1, we report the regression results based on the SVI by ticker symbols, various investor interview sentiment indexes and trading volume to identify which of them have influence on the average returns of TAIEX. In Section 2, we show the regression results based on the SVIs searched by company names and by ticker symbols of companies of different ranks. Additionally, we evaluated if the SVI based on the top-50 companies’ ticker symbols helps predict the average returns of TAIEX. In Section 3, we analyzed the p-value of the SVI searched by companies with different ranks in TAIEX to obtain deeper insights about investors’ behaviour in the Taiwan stock market. In Section 4, we ran regression, using the SVI with an increased trend to test the attention hypothesis of Da, Engelberg, and Gao (2011). In Section 5, we used data in different time lag to evaluate whether SVI can predict TAIEX average returns for a long period of time or not. In section 6, we summarise the chapter.
‧
1. Variables with a Significant Impact on Average Stock Returns
We first ran regression of all 7 variables on the average TAIEX returns. Following Da, Engelberg, and Gao (2009), we used the SVI searched by all companies’ ticker symbols to run the regression. The results are given in Table 1.
Table 1
Regression of all 7 variables on the average TAIEX returns Parameter Estimates
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
23
Note. t = 146. 𝑙𝑙(𝑛𝑛) = 4. *p < 0.05. **p < 0.01. ***p < 0.001.
y ∶ average returns of TAIEX 𝑥𝑥1: Google SVI
𝑥𝑥2: trading volume of TAIEX 𝑥𝑥3: Taiwan Stock Price Index 𝑥𝑥4: Taiwan Economic Situation Index 𝑥𝑥5: Taiwan Political Index 𝑥𝑥6: Taiwan Investment Environment Index 𝑥𝑥7: Global Economic Index 𝑥𝑥8: Possibility of the Value of Your Portfolio Increases within six months
Table 1 indicates that trading volume and investors’ interview sentiment of Taiwan Stock Prices are significantly related to the average TAIEX returns, while the other six variables are not (p > 0.05). Although Taiwan Economic Situation Index isn’t shown to have an impact on the average returns, this result is likely to be caused by other confounding interview sentimental variables. To test the hypothesis, we removed other interview sentimental variables 𝑥𝑥5, 𝑥𝑥6. 𝑥𝑥7, and 𝑥𝑥8 and then ran another regression. The results are shown in Table 2.
Table 2
Regression of two interview sentiment indexes, the SVI and trading volume on average TAIEX returns
Parameter Estimates
Parameter Estimate Approx. Std Err T-value Approx. Pr > |t|
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
24
Intercept 0.000382 0.000767 0.50 0.6188
𝑥𝑥1 0.024569 0.0144 1.71 0.0895
𝑥𝑥2 0.031812 0.00949 3.35 0.001**
𝑥𝑥3 -0.00613 0.00231 -2.66 0.0088**
𝑥𝑥4 0.002872 0.00136 2.12 0.0361*
Note. t = 146. 𝑙𝑙(𝑛𝑛) = 4. *p < 0.05. **p < 0.01. ***p < 0.001.
y ∶ average returns of TAIEX 𝑥𝑥1: Google SVI
𝑥𝑥2: trading volume of TAIEX 𝑥𝑥3: Taiwan Stock Price Index 𝑥𝑥4: Taiwan Economic Situation Index
Table 2 shows that investors’ interview sentiments of both Stock Price Index (p = 0.0088) and Taiwan Economic Situation Index (p = 0.0361) affect the average returns of TAIEX. Additionally, trading volume (p = 0.001) is significantly related to the TAIEX returns. This supports the research of Lee and Swaminathan (2000) that trading volume is an indication of a stock past performance and can be used to predict the stock future performance. It also agrees with that reported in Chuang, Ouyang, and Lo (2010), that trading volume is positively related to the average returns of TAIEX.
Among the three variables that impact TAIEX, trading volume and the investors’
interview sentiment of Taiwan Economic Situation are positively correlated to the average TAIEX returns: 1 % increases in trading volume and Taiwan Economic
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
25
Situation will increase 0.031812% and 0.002872% in the average returns of TAIEX. We gauged that investors’ opinions about the economic of Taiwan in the future come from their observations of the changes in daily living, which reflect the reality. It therefore has positive correlation with the TAIEX returns.
By contrast, investors’ interview sentiment of Taiwan Stock Price is negatively correlated to the average TAIEX returns: 1% increase in Taiwan Stock Price Index will decrease 0.00613 in the average returns. This can be explained by the overconfidence and over-optimism hypothesis of De Bondt and Thaler (1995) in that while investors’
confidence about Taiwan Stock Price is positive, the actual TAIEX stocks returns are not so good as they believe.
2. Ticker Symbols vs. Company Names: SVIs for Investors Attention
Table 1 shows that the SVI based on the ticker symbols of all companies is not correlated to the average TAIEX returns. Since we used SVI to represent investors’
attention and companies with bigger market capital normally receive more attention, the SVI of companies with higher ranks in TAIEX might better reflect investors’ attention.
To verify this hypothesis, we used a top-down approach by separating the SVI of all companies’ ticker symbols into various ranks (top-800, top-700, top-600, top-500, top-400, top-300, top-200, top-100). After that, we ran regressions using each of the
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
26
SVI series. The results show that only the SVI by ticker symbols of the top-100 companies have an impact on the average TAIEX returns while all others have not.
Table 3 presents the results of SVI based on ticker symbols of the top-100 companies.
Table 3
Regression of two interview sentiment indexes, trading volume and the SVI based on ticker symbols of the top-100 companies on average TAIEX returns
Parameter Estimates
Parameter Estimate Approx. Std Err T-value Approx. Pr > |t|
Intercept 0.000336 0.000756 0.44 0.6576
𝑥𝑥1 0.016388 0.00736 2.23 0.0276*
𝑥𝑥2 0.030368 0.00915 3.32 0.0012**
𝑥𝑥3 -0.00563 0.00228 -2.47 0.0147*
𝑥𝑥4 0.002768 0.00135 2.05 0.0422*
Note. t = 146. 𝑙𝑙(𝑛𝑛) = 4. *p < 0.05. **p < 0.01. ***p < 0.001.
y ∶ average returns of TAIEX 𝑥𝑥1: Google SVI
𝑥𝑥2: trading volume of TAIEX 𝑥𝑥3: Taiwan Stock Price Index 𝑥𝑥4: Taiwan Economic Situation Index
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
27
The result triggered our curiosity about whether the same pattern also exists in the SVI by company names for different ranks. To answer that question, we also separated the SVI of all company names into various ranks (top-800, top-700, top-600, top-500, top-400, top-300, top-200, top-100). We then ran regressions using each of the SVI series.
To our surprise, we found that all SVIs are significantly and positively related to the TAIEX average returns. Table 4 shows the regression result of the SVI searched by the top-100 company names. Regression of other SVI series has similar results. This result supports the attention hypothesis of Barber and Odean (2007) that individual investors would buy stocks that attract more attention where attention is measured by the Google search volume in our study. We will provide more in depth analysis of the p-value of SVIs under these regressions in the next Section.
Table 4
Regression of two interview sentiment indexes, trading volume and the SVI based on company names of the top-100 companies on average TAIEX returns
Parameter Estimates
Parameter Estimate Approx. Std Err T-value Approx. Pr > |t|
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
28
Intercept 0.000275 0.000760 0.36 0.7183
𝑥𝑥1 0.038288 0.0132 2.89 0.0044**
𝑥𝑥2 0.02717 0.00811 3.35 0.0010**
𝑥𝑥3 -0.00606 0.00230 -2.63 0.0096**
𝑥𝑥4 0.003069 0.00138 2.22 0.0282*
Note. t = 146. 𝑙𝑙(𝑛𝑛) = 4. *p < 0.05. **p < 0.01. ***p < 0.001.
y ∶ average returns of TAIEX 𝑥𝑥1: Google SVI
𝑥𝑥2: trading volume of TAIEX 𝑥𝑥3: Taiwan Stock Price Index 𝑥𝑥4: Taiwan Economic Situation Index
In Fan, Liao, and Chen (2014), the authors reported that the SVI based on the top-50 company names is significantly related to the average returns of TAIEX. To compare our results with theirs, we made two more SVIs based on the top-50 companies.
One used ticker symbols and the other used company names. We then ran two regressions using each of the series. The results are given in Tables 5 & 6.
Table 5
Regression of two interview sentiment indexes, trading volume and the SVI of ticker symbols for the top-50 companies on average TAIEX returns
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
29
Parameter Estimates
Parameter Estimate Approx. Std Err T-value Approx. Pr > |t|
Intercept 0.000311 0.000756 0.41 0.6810
𝑥𝑥1 0.016982 0.00693 2.45 0.0155*
𝑥𝑥2 0.029725 0.00923 3.22 0.0016**
𝑥𝑥3 -0.00566 0.00225 -2.52 0.0130*
𝑥𝑥4 0.002779 0.00133 2.08 0.0391*
Note. t = 146. 𝑙𝑙(𝑛𝑛) = 4. *p < 0.05. **p < 0.01. ***p < 0.001.
y ∶ average returns of TAIEX 𝑥𝑥1: Google SVI
𝑥𝑥2: trading volume of TAIEX 𝑥𝑥3: Taiwan Stock Price Index 𝑥𝑥4: Taiwan Economic Situation Index
Table 6
Regression of two interview-sentiment indexes, trading volume and the SVI based on top-50 company names on average TAIEX returns
Parameter Estimates
Parameter Estimate Approx. Std Err T-value Approx. Pr > |t|
Intercept 0.000232 0.000764 0.30 0.7618
𝑥𝑥1 0.033182 0.0109 3.05 0.0028**
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
30
𝑥𝑥2 0.028716 0.00868 3.31 0.0012**
𝑥𝑥3 -0.00597 0.00231 -2.59 0.0107*
𝑥𝑥4 0.003052 0.00139 2.20 0.0292*
Note. t = 146. 𝑙𝑙(𝑛𝑛) = 4. *p < 0.05. **p < 0.01. ***p < 0.001.
y ∶ average returns of TAIEX 𝑥𝑥1: Google SVI
𝑥𝑥2: trading volume of TAIEX 𝑥𝑥3: Taiwan Stock Price Index 𝑥𝑥4: Taiwan Economic Situation Index
As the above two tables show, both SVIs searched by the top-50 company names and by the top-50 ticker symbols are significantly related to the average TAIEX returns, which agrees with that reported in Fan, Liao, and Chen (2014).
3. Analysis of Investors’ Behaviours
In the previous section, we have presented the regression results using the SVIs of company names and ticker symbols of various ranks. To better understand the relationship between these SVIs and investors’ attention, we plotted the p-value of SVIs for company names and ticker symbols under various ranks on Figure 1. The lower the p-value is, the more significant the SVI is in impacting the TAIEX average returns.
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
31
Figure 1
p-value of the SVI searched by companies with different ranks in TAIEX
Figure 1 shows that all SVIs by ticker symbols have p-value indicating they are not
Figure 1 shows that all SVIs by ticker symbols have p-value indicating they are not