Trust and Local Bias of Individual Investors

Transcription

1 Trust and Local Bias of Individual Investors Ran Shao and Na Wang August 2016 Abstract It has been widely documented that investment decisions of individual investors exhibit local bias. Yet little is known of how societal forces affect investors portfolio allocations of local versus nonlocal assets. We propose that social capital, and trust in particular, decreases the local bias of individual investors. Using the trading records of Chinese retail investors, we find that investors with a higher level of trust exhibit less local bias in their stock investments, and this pattern survives under various robustness checks. Furthermore, we document that trust has less effect on local bias of small stocks and potentially informed investors. These findings suggest that trust plays an important role in individual investors financial decisions, especially for their noninformation-driven trading behaviors. Key words: trust, local bias, portfolio holdings, individual investors, small stocks, information advantage. JEL Classification: G02, G11, G15. We are grateful to David Hirshleifer, Andrew Karolyi, and Jianfeng Yu for their helpful comments and suggestions. Ran Shao is at the Department of Economics of Yeshiva University, New York, NY 10016, rshao@yu.edu; Na Wang is at the Department of Finance of Frank G. Zarb School of Business, Hofstra University, Hempstead, NY 11549; (516) , Na.Wang@hofstra.edu. 1

2 1 Introduction One of the puzzling facts of individual investors investment behavior is that their portfolios are locally biased. The strong bias in favor of domestic securities and securities close to home has been evident in many financial markets and among different groups of investors. For example, Ivkovic and Weisbenner (2005) examine the investments of a large number of individual investors made through a U.S. discount broker, and find that households exhibit a strong preference for local investments. Grinblatt and Keloharju (2001) document the behavior of Finnish investors, who are more likely to trade stocks that are geographically close to them and that have chief executives with similar cultural backgrounds. Coval and Moskowitz (1999) document that U.S. investment managers are in favor of investing in locally headquartered firms, particularly small and highly leveraged firms. It has been extensively studied in the literature whether local bias is driven by investors information advantage or by behavioral biases (such as the familiarity bias). The local biases of professional investors and analysts are mostly supported by their information advantages (e.g., Coval and Moskowitz 1999 and 2001; Bae, Stulz, and Tan 2008; Baik, Kang, and Kim 2010). In contrast, there is a lack of consensus regarding the causes of individual investors local bias. Ivkovic and Weisbenner (2005) find that households on average generate larger returns from their local holdings relative to their nonlocal holdings, suggesting that individual investors can exploit local advantages. In contrast, Seaholes and Zhu (2010) use calendar-time portfolios showing that the local holdings of individual investors do not generate abnormal performance. Zhu (2002), Huberman (2001), and Baltzer, Stolper, and Walter (2015) further support the non-fundamentally based familiarity as a more plausible explanation for the local bias of individual investors. Recently, Korniotis and Kumar (2013) suggest that psychological and information-based explanations for individual portfolio distortions apply to distinct subsets of investors based on the investor s level of smartness. Aside from debating between the information-driven or familiarity-driven local biases, the literature remains silent on how societal forces, and trust in particular, contribute the local 2

3 investment decisions of individual investors. 1 Guiso, Sapienza, and Zingales (2008) argue that a general lack of trust can affect stock market participation. They show that investors factor in the risk of being cheated when deciding whether to buy stocks. Guiso, Sapienza, and Zingales (2009) document that cultural biases, such as a lower level of bilateral trust, can lead to less trade between two countries, and less portfolio or direct investment. In addition, social capital or trust has also been shown playing a sizeable role in economic growth and financial development (e.g., Algan and Cahuc 2010; Guiso, Sapienza, and Zingales 2004 and 2006). This emerging literature indicates that cultural and societal forces are likely to affect investors decisions on local versus nonlocal investments. In this paper, we hypothesize that trust, forming the foundation of cooperation and trade, decreases the local bias of individual investors, especially for their non-information-based investments. On the one hand, individual investors may invest disproportionally in local stocks due to familiarity and, therefore, exhibit local bias. While trust can increase stock market participation and stock investment in general (see, for example, Guiso, Sapienza, and Zingales 2008), it is likely that investors trusting attitudes play a larger role on less familiar investments, so that the increase in nonlocal investments is larger than that in local investments, causing a reduction in local bias. On the other hand, if local bias is indeed driven by the relative information advantage toward local assets, trust may have a weaker or no effect on these information-driven trades, hence exhibiting less or no effect in terms of reducing local bias. In our study, we do not aim to distinguish between the causes of local bias. Instead, we argue that a weaker trust effect on information-driven trades further supports our hypothesis. We test our hypothesis using various trust measures and stock investments of individual investors across Chinese provinces. China is ideal for testing the hypothesis for the following reasons. First, due to the household registration system, China has low population mobility. Residential stability facilitates the application of province-level trust measures to individuals, 1 One exception is the concurrent paper of Wei and Zhang (2016) focusing on institutional investors. We will further distinguish our findings from theirs later in the introduction. 3

4 making it relatively more accurate compared to other studies using location-based trust measures for a population with high mobility. Second, again due to the low population mobility, trust levels vary a lot across different provinces. According to Ang, Cheng, and Wu (2015), the differences in social and cultural characteristics among provinces in China are often greater than the differences across 13 European countries combined. 2 Third, Chinese stock markets (established in early 1990s) have a short history, which allows trust to play an important role for investors facing relatively new and unfamiliar circumstances. Our main independent variables of interest relating to trust are measured at the provinceyear level. In particular, we utilize the Chinese General Social Survey conducted annually by academic institutions in China since We construct the average trust level in each province based on the respondents answers to a survey question: "Generally speaking, do you agree that most people in the society can be trusted?" In addition, we explore two other trust measures from different sources: a bilateral trust measure of the perceived trustworthiness of enterprises in each province, developed by Ke and Zhang (2002); and a city-level trust measure in the Spiritual Life Study of Chinese Residents, provided by the Association of Religious Data Archives. Our dependent variable the local bias of individual investors is measured based on their stock investment portfolios. 3 We obtained the trading records of about 300,000 individual investors from one of the largest brokerage houses in China. We calculate the local bias of each investor using the percentage of local stock holdings in the investor s portfolio minus the weight of local stocks in the market portfolio. This local bias measure at the investor level is similar to that at the stock level used in Coval and Moskowitz (1999) by assuming that every investor holds the market portfolio as a benchmark. Our baseline analysis regresses the investor local bias on the average trust level of the investor s home province, controlling for investor characteristics as well as other province 2 Our bilateral trust measure is based on the same survey of Ke and Zhang (2002) used in Ang, Cheng, and Wu (2015). 3 An alternative transaction-based measure of local bias is discussed in the Appendix. 4

5 characteristics. We find that investors from provinces with higher levels of trust exhibit less local bias, and this finding is both statistically and economically significant. One standard deviation increase in trust measures can decrease the local bias by around 20% of its standard deviation. Our results hold under a variety of robustness checks, such as examining different sample periods and investors with different characteristics, using alternative specifications with only cross-sectional variations, and measuring the average of investor local bias at the province-year level. We further examine the link between the trust effect and the potential information advantage of investors holding local stocks. If investors do have information advantages over local stocks and tilt their investment to achieve higher returns, we expect trust to have less or even no effect on their information-driven local bias. We first identify this link with local bias measured on small stock holdings. Small stocks have less analyst or media coverage, are diffi cult to value, and, consequently, have higher information asymmetry between local and nonlocal investors. Indeed, we find that trust shows a lower (less than one quarter) economic effect on the local bias of small stocks than on that of total stocks. We then examine the performance of investors local investments in small stocks and identify a group of potentially informed local investors (around 4.5%). As we conjectured, we find little evidence that trust affects the local bias of small stocks among this group of informed investors. We document the trust effects on the local bias of individual investors based on the average trusting attitudes of the investors home provinces. Some possible confounds are that the trusting attitudes of investors in one area might be correlated with the trustworthiness of the environment and stocks located in more trustworthy areas could be more attractive investments (see, for example, Ang, Cheng, and Wu 2015). To exclude the possibility that the variations in local bias of investors in different provinces are caused by the differences in local investment opportunities, we adopt the bilateral trust measure based on the survey of Ke and Zhang (2002) to control for the local investment properties and further pin down the effects of the investors trusting attitudes. This approach allows us to include both 5

6 investment fixed effects at the province level and individual fixed effects at the investor level. 4 We find that the provincial investments of individual investors have a significant positive relation with their trusting attitudes toward that province. Lastly, we repeat our main analysis of the trust effects on local bias of individual investors with a city-level trust measure provided by the Association of Religious Data Archives. We include investor province fixed effects to control for potential omitted province characteristics. Again, we find strong evidence that investors in the cities with a higher trust level exhibit a lower level of local bias in their portfolios. Our study contributes to the literature on understanding culture effects on local bias and portfolio diversification. In particular, Anderson, Fedenia, Hirschey, and Skiba (2011) show that a society s culture and the culture distance between two markets play an important role in asset allocation. Similarly, cultural differences also show a significant influence on corporate diversification decisions, such as oversea listing decisions and international mergers (e.g., Sarkissian and Schill 2004; Ahern, Daminelli, and Fracassi 2015). We explore a different source of societal force trust and provide the first evidence that trust reduces the local bias of individual investors, especially for their non-information-based trading behaviors. Our work is related to a concurrent study by Wei and Zhang (2016), which focuses on institutional investors. They argue that institutional investors who lack trust tend to form relationship with local companies to guide their investment decisions. Hence, their investments exhibit greater biases toward local stocks. In other words, lower trust increases institutional investors information advantage, which, in turn, induces more informationdriven local biases. In contrast, we focus on individual investors who are less likely to obtain information through connections with local companies. Hence, their trust levels are not directly related to their information advantages. For the individual investors who are 4 In the previous analysis, we controlled for the characteristics of investors and investors home provinces, such as investor age, gender, and wealth, and province GPD per capital, population, and trading technology. Including investment province fixed effects and investor fixed effects allows us to further control for the variations in investment opportunities in different provinces, as well as for other unobserved investor characteristics, such as eduction, trading experience, and risk aversion. 6

7 potentially informed on local small stocks, we find that trust has no significant effect on their local biases of small stocks. Instead, trust plays a more important role in reducing the non-information-based local bias at the investor level. Our study can further pin down the trust effects using a bilateral trust measure and controlling for the fixed effects at both the stock investment level and the investor level. The rest of the paper proceeds as follows. We discuss our data and variables in Section 2. We present our main regression results, robustness, and further classifications using small stocks and other trust measures in Section 3. Section 4 concludes. Appendix discusses an alternative measure of local bias based on transactions instead of portfolio holdings. 2 Data and Variables 2.1 Investor and Stock Data The main analysis for the local bias of individual investors is conducted at the province level. A sample of about 300,000 individual investors from a nationwide brokerage house in China was randomly selected from all the existing accounts at the end of The sample covers investors from 21 provinces (70 cities) in China. For each investor, we observe the customer code, age, gender, total account value at the end of the sample period, and a complete transaction record from January 2006 to December Each record has a customer code, a security code, the trade date, a buy-sell indicator, and the price and quantity of the trade. We also observe the portfolio holdings of each investor at the end of sample period. Combining the holdings with transaction records, we can obtain the portfolio holdings of each investor at any specific point of time. In particular, we use investor portfolio holdings at the end of each month during Most importantly, the data include the location information of individual investors, that is, where they opened their brokerage accounts. As previously mentioned, due to the household registration system, most residents in China 5 It is the same data set used in Hong, Jiang, Wang, and Zhao (2014). 7

8 have low mobility, which makes it reasonable to connect an investor s trusting attitude with the average trusting attitudes of people in the same city or province. The stocks in our data set are A-shares listed on the Shanghai and Shenzhen Stock Exchanges. 6 Stock prices, trading volumes, and market values are from the China Stock Market and Accounting Research (CSMAR) database. CSMAR also provides the location of firms headquarters. For each investor, we define local stocks as those headquartered in the investor s home province. The investor sample is from 2006 to We obtain several province-level characteristics as our control variables. The gross domestic product (GDP) per capita and population data are from National Bureau of Statistics of China for each province-year observation. We also obtain from the Shanghai Stock Exchange the number of brokerage house branches located in the different provinces that send trades to the stock exchanges. We use the ratio of number of brokerage house branches to population to measure the access of trading technology in each province. 2.2 Trust Measures Our main measures of trust are collected from the Chinese General Social Survey (CGSS), which launched in 2003 and is the earliest national representative annual survey project run by academic institutions in China. 7 Since 2010, the CGSS has included a question on social trust: "Generally speaking, do you agree that most people in the society can be trusted?" The answers range from "totally agree (scaling 5)", "agree (4)", "not agree nor disagree (3)", "disagree (2)" to "totally disagree (1)". Our sample of trust measures covers 2010, 2011, and 2012, when the investor sample ends. In each year, the CGSS randomly sampled around 12,000 households in each province in China. Our first trust measure, TRUSTN, is calculated as the average of the trust scales of households in each province, normalized 6 Only A-shares are included in our analysis. B-shares are denominated in dollars and traded mostly by foreign investors. 7 CGSS is aimed to systematically monitor the changing relationship between social structure and quality of life in China. The survey data and detailed description of the CGSS are available at: 8

9 across provinces in each year. Alternatively, we construct another trust measure, TRUSTP, as the percentage of households in each province answering "totally agree" or "agree" to the survey question on social trust. We use both TRUSTN and TRUSTP in most tests to see how robust our results are. In addition, we consider a bilateral trust measure of the perceived trustworthiness of the enterprises in each province, BILTRUST. Ke and Zhang (2002) conduct a survey of over 5,000 managers from all provinces in China and ask them to rank the trustworthiness of each province based on their business experiences there. We adopt this bilateral trust measure to pin down the trust effects on the regional investment of investors. Lastly, we utilize a city-level trust measure (CITYTRUST) from the Spiritual Life Study of Chinese Residents provided by the Association of Religious Data Archives. The study surveyed a series of questions related to trust in Though it is a snap-shot measure of trust, it allows us to perform another identification strategy with province fixed effects and show the robustness of our results across different sources of trust measures. We conduct our main analyses based on TRUSTN and TRUSTP. We will provide detailed discussions and results on the BILTRUST and CITYTRUST measures in the last two subsections of Section Local Bias Measures For each investor (i) in province (j) at year (t), we calculate his/her local bias (LB ijt ) as the average of the monthly local bias during that year. We define the local stocks for each investor as those headquartered in the investor s home province. Based on the portfolio holdings (H) of each investor at the end of each month (t ), we calculate the yearly local bias as the average of monthly local bias (LB ijt ) defined as follows: LB ijt = H ijt H mjt, 9

10 where H ijt measures the value of local stocks divided by the value of total stocks in investor i s portfolios at month t ; H mjt measures the value of all stocks headquartered in province j divided by the value of total stocks in the markets at month t. In addition to using the value of stocks, we also calculate local bias using the number of shares of the corresponding stocks as an alternative measure after excluding the effects of stock price fluctuations. We define these two local bias measures as LBVALUE and LBSHARE, respectively. Essentially, H ijt measures the percentage of local stocks in investors portfolios, while H mjt measures that percentage in market portfolios. The difference, namely local bias, measures how an investor s portfolio deviates from a benchmark (towards local stocks). This measure is similar to that used in Coval and Moskowitz (1999), but we calculate it at the investor level instead of the stock level. 8 In addition, if we calculate investor s yearly local bias directly using the portfolio holdings at the end of each year, we find very similar results as those using the average of the monthly local bias during each year. Panel A of Table I summarizes the average local bias of investors at each province and province characteristics over the sample period , and our two main measures of trust over the sample period Provinces are sorted on the descending order of the log GDP per capita, LNGDPPC. On average, individual investors exhibit local bias of around 4%, measured either by LBSHARE or LBVALUE. With the exception of Beijing, investors in all other provinces show positive local biases, ranging from 2% to 10.5%. Our results are robust with or without including investors from Beijing. Looking at trust measures, there is no obvious monotone relation between TRUSTN and LNGDPPC, and TRUSTP shows similar patterns as TRUSTN. LNPOP is the log population in millions. BRANCH counts the ratio of number of brokerage houses to population in millions. BRANCH decreases dramatically as LNGDPPC decreases, showing that underdeveloped areas have less access to trading technologies. 8 For a given stock, Coval and Moskowitz (1999) calculate the local bias as the difference between the actual local investment of local investors in excess of a benchmark investment, assuming every investor holds the market portfoilo. 10

11 Panel B reports the pairwise correlation matrix of our main variables at the province-year level during the sample period Several pairs of correlations are worth mentioning. Investor local biases, LBSHARE and LBVALUE, are negatively correlated with both trust measures. TRUSTN and TRUSTP are highly correlated, with a correlation coeffi cient of 92%, and they are both negatively correlated with LNGDPPC and BRANCH and positively correlated with LNPOP. To make sure that our trust measures do not capture purely the economic or financial development, in the unreported analysis, we obtain the orthogonalized trust measures. We use the residuals from the regressions of trusts on LNGDPPC, LNPOP, and BRANCH, and our results of the trust effects on local bias are virtually unchanged. 3 Methodology and Results 3.1 Trust Effects on Local Bias We examine whether investors living in areas with a higher level of trust are more likely to exhibit a lower level of local bias. Our main variable of interest is Trust (TRUSTN or TRUSTP), calculated based on the survey question in the CGSS. We run the panel regression of investor local bias as follows: LB ijt = α + βt rust jt + γ Controls + Y eardummy t + ɛ ijt, where the Controls include the province characteristics LNGDPPC jt, LNPOP jt, and BRAN CH jt, as well as the investor characteristics LNAGE it, GENDER i, LNSIZE it, and VOL it. For the investor characteristics, LNAGE it is the log age of investor i at the end of year t; GENDER i takes the value of one for female investors and zero for male investors; LNSIZE it is the average value of the monthly stock investment of each investor in each year; VOL it is the standard deviation of the monthly value-weighted portfolio returns of investor i during year 11

12 t, and we use VOL it to measure investors risk-taking behaviors. 9 Province characteristics are defined similarly as those in Table I. We also include YearDummy in the above regressions. Note that our sample of investor local bias spans from 2006 to 2012, while the surveybased trust measures have been available only since Thus, in our main analysis of TRUSTN and TRUSTP, we focus on the sample period of In addition, we also provide evidence using the entire investor sample period of by using the trust measures in 2010 for those during The summary statistics of the investor-year local bias sample are shown in Table II. Panel A summarizes the sample periods of The mean and standard deviation of LBSHARE (LBVALUE) are and (0.040 and 0.206), respectively. The standard deviation of TRUSTN (TRUSTP) is (0.058). The mean investor age is around 42 (LNAGE = 3.739), and there are slightly more male investors than female investors in our sample (GENDER = 0.432). The summary statistics for the sample periods of reported in the Panel B of Table II are quite similar to those reported in the Panel A. Table III reports the ordinary least square (OLS) estimation results, with standard errors clustered at two dimensions (province and year). Columns (1)-(4) report the estimation results using the sample of In Column (1), we regress LBSHARE on TRUSTN with all the controls included. The coeffi cient of TRUSTN is 0.095, with a t-statistic of The negative coeffi cient of TRUSTN shows that individual investors with higher level of trust exhibit significantly lower level of local bias. The economic significance of the trust effect is also large. Considering the standard deviations of LBSHARE and TRUSTN are and 0.541, respectively (see Table 2), a one-standard-deviation increase in TRUSTN leads to a decrease in the left-hand-side variable by 23% (= /0.220) of its standard deviation. Column (2) documents similar results when using LBVALUE as the 9 To fully control for investors characteristics, such as risk aversion or education, we include investor fixed effects in the bilateral trust analysis in Section The average time-series correlation of trust measures in 2010, 2011, and 2012 are around 70%. The results are almost identical if we substitute the trust measures during with the average of trust measures during

13 dependent variable. In Columns (3) and (4), using TRUSTP as the independent variable shows again both statistically and economically significant trust effects. The corresponding economic magnitudes are around 19% to 25% of the dependent variable s standard deviation. In Columns (5) through (8), we report the estimation results using the sample of The results are statistically stronger than those in the first four columns, given that we have almost twice observations. For example, in Columns (7) and (8), the effects of TRUSTP on local bias measures are significant at the 1% level, and the absolute values of the t-statistics are larger than 3. The economic significance is slightly weaker, with a one-standard-deviation increase in the trust measures, the dependent variables decreases by about 10% 15% of its standard deviation. 11 The effects of other control variables are also worth mentioning. Investor age, gender and portfolio volatility (VOL) are not significantly related to local versus nonlocal investment decisions. LNSIZE is significantly positively related to the local bias of individual investors in the last four columns. This result, though weak, can be attributed to the information advantage of wealthier investors, which leads them to invest proportionally more on local stocks. As for the province characteristics, LNGDPPC is mostly negatively related to the local bias of individual investors. LNPOP shows some positive effects, but only in the first four columns. BRANCH does not significantly relate to our local bias measures. The fact that trust effects persist after controlling for the investor and province characteristics shows that trust is important in accounting for the variations of local bias across different provinces and in different sample periods. In the rest of this subsection, we examine the robustness of trust effects on the local bias of different groups of individual investors and whether trust effects vary among different investor characteristics. During each year of , we split our sample into two groups based on investor age, gender, portfolio size (LNSIZE) and volatility (VOL), respectively. The regression results are presented in Table IV, with TRUSTP as our main variable of 11 The weaker economic significance could be due to the fact that we have only the cross-sectional variations of trust measures during 2006 to

14 interest and LBVALUE as the dependent variable. 12 The first distinct pattern is that trust effects are significant among all of the eight groups of individual investors. In addition, we find no significant differences in the trust effects on investors with different ages, genders, and portfolio sizes. When we look at investors with different portfolio volatilities, in Columns (7) and (8), TRUSTP shows a larger effect on the local bias of investors with higher portfolio volatilities, and this difference is statistically significant (with the untabulated t-statistic of 5.54). Overall, our main conclusions survive with the different sample selections. 3.2 Alternative Specifications The local bias in our baseline analysis is measured at the investor-year level. Though we have clustered standard errors by two dimensions (province and year), we consider two alternative specifications to fully address the potential correlations of investor local bias and provide further robustness for our main results. We first examine the trust effects using only the cross-sectional variations of investors local biases. For each investor, we calculate the local bias as the average of the yearly local bias over the sample period of We match the trust measures, investor characteristics, and province characteristics using the sample average as well. The crosssectional regression results are presented in Table V, with the standard errors clustered at the province level. We can see that both TRUSTN and TRUSTP are significantly negatively correlated with our average local bias measures. For example, the coeffi cient of TRUSTP in Column (3) is with a t-statistic of The statistical significance is similar to that in our baseline analysis in Table III, and the point estimates are slightly larger in the cross-sectional setup. Next, we construct a local bias measure at the province-year level to mitigate the trading correlation of investors from the same province. For each of the 21 sampled provinces, we calculate the province-level local bias using the average of the local bias of all investors in 12 The results using TRUSTN as the main variable of interest or LBSHARE as the dependent variable, or using the sample period of , are almost identical. 14

15 that province. 13 We then regress the province average local bias on our trust measures and control for province characteristics and year fixed effects. Table VI reports the regression results, with the standard errors clustered by two dimensions (province and year). Given the limited number of observations we have, it is comforting to see that our trust measures are correlated with province-level local bias with t-statistics close to 2. In addition, trust is the only significant factor in explaining the variations in the province average local bias, and the rest of the province characteristics show no significance in this specification. Overall, the robust results using the alternative specifications provide further support for our hypothesis that trust reduces the local bias of individual investors. 3.3 Local Bias, Small Stocks, and Investors with Information Advantage Local Bias of Small Stock Holdings We have documented that trust decreases the local biases of individual investors. However, if investors local biases are driven by a relative information advantage on local stocks, we would expect trust to have less or no effect on their local versus nonlocal stock trading. Even investors with higher levels of trust may tilt their portfolio holdings toward local stocks due to their information advantage. There is little consensus in the local bias literature about whether or not investors possess private information about their local stocks. But if they do, it would be more evident in small local stocks than in large local stocks (see, for example, Ivkovic and Weisbenner 2005). In this subsection, we examine whether trust shows less effect on the local bias calculated using small local stocks. In each year, we rank all stocks based on their market capitalization, and we group the bottom 50% of stocks as the small stocks. 14 We then calculate the local bias 13 Using the investment-size weighted average of the local bias of all investors in the province yields similar results. 14 The stock s market capitalization is measured in each December of the previous year, and our results are identical when using a 40% or 60% criterion to classify small stocks. 15

16 of small stocks using the holding percentage of small local stocks in investors portfolios of small stocks minus the corresponding percentage calculated with the market portfolios. This method is similar to that used in Subsection 3.1; instead of focusing on all stocks, we calculate the percentage difference based on small stocks only. 15 We define the small stock local biases as SLBSHARE and SLBVALUE in terms of stock share and value, respectively. The summary statistics using only small stocks are in Table VII. SLBSHARE and SLBVALUE have the mean of and 0.042, respectively, with standard deviations of These numbers are comparable to those in Table II. The investor and province characteristics are also similar to those used in the analysis of all stocks. Table VIII presents the trust effect on investor local bias of small stocks. The point estimates of the trust effect on investments in small stocks are generally smaller than those on all stocks (in Table III). For example, the coeffi cient of TRUSTN in Column (1) is 0.021; though it is statistically significant, the magnitude is smaller than the corresponding number of in the analysis of all stocks. In Column (3) and (4), TRUSTP has the coeffi cients of and The corresponding numbers in Table III are and Putting together, we find statistically significant trust effect on local bias of small stocks, but the economic significance is much smaller. A one-standard-deviation increase in trust leads to a decrease in SLBSHARE or SLBVALUE by only 3% 5% of its standard deviation, while the trust effect on the local bias calculated using all stocks in Table III is around 19% 25%. This finding confirms our conjecture that trust has less effect on the potentially information-based trading, such as small local stock trading. Wei and Zhang (2016) document that for institutional investor trading, the trust effect on small stocks is nearly three times larger than that on large stocks, supporting their argument that institutional investors who lack trust are more likely to form business ties with small local companies to gain information advantages, which, in turn, strengthen the informationdriven local bias. In contrast, the weaker economic effect of trust on the small stock local 15 If an investor did not hold any small stocks (local or nonlocal) in a given year, that investor-year observation is excluded from our analysis in this subsection. 16

17 bias of individual investors implies that trust affects the local trading of individual and institutional investors differently. Instead of forming business relations, which is unlikely for individual investors, a lack of trust does not necessarily change their relative information advantage in local versus nonlocal investments. Hence, trust has a smaller effect on the potentially information-based (small stock) local bias of individual investors Local Bias of Investors with Information Advantage in Small Stocks We further look at whether the effects of trust differ among investors with heterogeneous information on small local stocks. The potential heterogeneity in the amount of information investors possess provides us with another testable hypothesis. We expect that the trust effect is stronger among investors without local private information, and weaker, or even insignificant, for investors who have local information advantages and deliberately tilt their portfolios toward small local stocks. To identify a group of (potentially) informed local investors during our sample period, we employ the criterion similar to that used in Odean (1999) and Seasholes and Zhu (2010). For each investor at each year, we calculate differences of the future one-month returns of investor s monthly buying portfolios and those of the monthly selling portfolios of the small local stocks. We classify investors whose buy-sell return differences are statistically positive as the informed local investors. 16 We define a dummy variable INF O it, which takes the value of one for the potentially informed investors during year t, and zero, otherwise. The summary statistics of INF O it are listed in the last row of Table VII. The mean value of INFO is 0.045, which means on average around 4.5% of investors are identified as relatively informed about their small local stocks each year. We include INFO and its interaction with trust in the analysis of the local bias of small 16 We set the statistical significance at 10% level (one-tailed t-test). 17

18 stocks. The regression model is as follows: SLB ijt = α + β 1 T rust jt + β 2 T rust jt INF O it + β 3 INF O it + γ Controls + Y eardummy t + ɛ ijt, where SLB represents SLBSHARE or SLBVALUE, Trust takes the value of TRUSTN or TRUSTP, and Controls and YearDummy are defined similarly as those in Section 3.1. In Panel A of Table IX, we present the panel regression results with standard errors clustering at two dimensions (province and year). In Columns (3) and (4), the interaction terms of TRUSTP and INFO have positive coeffi cients of 0.22, and they are statistically significant at 1% level with t-statistics around 3.6. When compared to the coeffi cients of TRUSTP, the trust effect disappears or even reverses among the group of informed investors. In Columns (1) and (2), the coeffi cients of TRUSTN INFO are also positive and statistically significant at 5% level and 1% level, respectively. To further identify whether trust affects the local bias of informed investors, we restrict our investor sample and focus only on the 4.5% of potentially informed investors. The regression results with only the informed investors are in Panel B of Table IX. We can see clearly that none of the specifications have any significant trust effects. The trust coeffi cients in the regressions of SLBSHARE (SLBVALUE) are negative (positive), and the absolute values of the t-statistics of these coeffi cients are around 1 or less. These results confirm our hypothesis that investors who possess private information tend to tilt their portfolios toward small local stocks, and this tendency is less likely to be affected by the level of trust. 3.4 Bilateral Trust and Province-Level Investment We have established our main results that the local bias of individual investors decreases with investors trust level, where trust level is measured by the average trusting attitudes of 18

19 surveyed households in the same province. However, the trusting attitudes of investors in an area could also be related to the trustworthiness of the environment. Whether stocks located in more trustworthy areas tend to be more attractive investments is a point of discernment. For example, Ang, Cheng, and Wu (2015) show that foreign investors favor regions in which local partners and employees are considered more trustworthy and that they are more likely to establish joint ventures in these regions. In this subsection, we control for the variations of the trustworthiness of companies in different provinces using a province-to-province bilateral trust measure, constructed from the survey in Ke and Zhang (2002). The bilateral trust measure is listed in Table A.2, which shows the trusting attitudes of the investors from provinces in the first row toward the business environment of provinces in the first column. It seems that people tend to rank their home provinces as the most trustworthy provinces, and the rest of the ranks vary from province to province. 17 Though the bilateral trust is a snap-shot measure, it gives us enough cross-sectional variations to explore the trust effects while controlling both investor and stock province fixed effects. We examine the relation of the bilateral trust and province-level investment using the following regressions: P ROSHARE ijj t = α + βbilt RUST jj + γ Controls + InvestorDummy i + StockP rovdummy j + Y eardummy t + ɛ ijj t, where P ROSHARE ijj t represents the average of monthly percentage of stocks from province j held by investor i (in province j) at the end of each month during year t in terms of stock shares. We also consider another dependent variable, P ROV ALUE ijj t, which is defined analogously in terms of stock value. BILT RUST jj is the trusting attitude of investors in province j toward investment opportunities in province j, Controls include investors home province characteristics LNGDPPC jt, LNPOP jt, and BRANCH jt, as well as 17 The fact that investors ranking their home province as the most trustworthy province could also explain the existence of local bias to some extent. 19

20 the time-varying investor characteristics, such as age (LNAGE it ), portfolio size (LNSIZE it ) and volatility (VOL it ). In addition to Y eardummy t, we also include the investor fixed effect, InvestorDummy i, and the investment province fixed effect, StockP rovdummy j. Note that this regression setup enables us to address the concerns of several alternative explanations. Investors characteristics, such as risk aversion, trading experience, education, and financial knowledge, could also affect the local bias of investors. For example, investors with more financial knowledge might know the benefit of portfolio diversification, and hence exhibit less local bias. By including the investor fixed effects, InvestorDummy i, we are able to control for all the fixed investor characteristics regardless whether they are captured by investor gender, age, wealth, and portfolio volatility in our previous analysis. Another possibility is that the difference in local stock investment opportunities could also drive the variations of the local bias of individual investors instead of their trusting attitude. To address this concern, we include the investment province fixed effect, StockP rovdummy j, to control for the local stock differences. Now we are able to fix the same set of stocks, fix the individual investors, and test whether the provincial investment decisions of investors do vary with their trusting attitude toward that province. Panel A of Table X lists the summary statistics of the variables used in the bilateral trust analysis during the sample period of The mean and standard deviation of PROSHARE are and 0.296, respectively, and these values of PROVALUE are similar. BILTRUST has the mean of and the standard deviation of Panel B of Table X reports the regression results. Columns (1) and (2), using observations from all provinces, show that BILTRUST is significantly positively related to PROSHARE and PROVALUE, with the t-statistics as large as 12. The economic significance is around 4%, meaning that a one-standard-deviation increase in BILTRUST leads to an increase in the dependent variable by around 4% of its standard deviation. To make sure that our results here are not totally driven by the high level of trust toward home province, in Columns (3) and (4), we exclude the observations of investors local investments (i.e., keeping only P ROSHARE ijj t,j j ). Again, 20

21 we find that investors provincial investments are significantly positively related to their corresponding trusting attitude toward each province, though the magnitudes are slightly smaller after excluding observations from home provinces. Hence, we find strong evidence that the trust attitude of individual investors affects their regional investment decisions, and these results hold after including various fixed effects at investor level, stock province level, and year level. 3.5 Local Bias and City-Level Trust Lastly, we perform a classic identification strategy with province fixed effects using trust variations at the city level. The city-level trust measure is constructed based on the Spiritual Life Study of Chinese Residents provided by the Association of Religious Data Archives. 18 The study surveyed 7,021 households with one of the questions asking "How do you trust businessman in general, scaling from 1 (not at all) to 9 (always)?". 19 While the survey has a limited coverage of only 18 cities overlapping with our sampled investors, the data gives us enough variations of trust on the city level. The city-level trust not only provides us a finer geographic measure of trust, but serves as an out-of-sample test since it is constructed using a completely different source of survey. Our main independent variable of interest, CITYTRUST, is calculated as the equal weighted average of the trust scales over all the surveyed households in each city. The mean and standard deviation of CITYTRUST in our sample are and 0.529, respectively. In this analysis, we control for investor characteristics and the available city-level characteristics, and we include province fixed effects. We report the regression results of LBSHARE and LBVALUE in Table XI. 20 We can see that CITYTRUST is significantly related to the local bias of individual investors. For example, in Column (1), the coeffi cients of CITYTRUST is with a t-statistic of 18 The study can be accessed at 19 There are other survey questions asking about how you trust your family, friends, relatives, and etc. Our main focus of stock investment is more related to the general trust attitude toward businessman. 20 The dependent variables, LBSHARE and LBVALUE, are constructed similarly as those in previous analyses. We keep using the province-level local bias variables, because many cities do not have enough stocks located in them to construct a solid local bias measures. 21

22 3.55, and a one-standard-deviation increase in CITYTRUST leads to an increase in the LBSHARE by around 13% of its standard deviation. The results are similar in Column (2) with LBVALUE as the dependent variable. Overall, our results are robust with city-level trust measure and we document that investors in cities with higher level of trust exhibit lower level of local bias in their investment portfolios. 4 Conclusions Local bias in investment decisions has been documented in many financial markets and among many groups of investors, yet little is known on how societal forces affect investors portfolio allocations in local versus nonlocal assets. In this paper, we examine whether trust, as the foundation of cooperation and trades, affects the local bias of individual investors. Using the trading records of Chinese individual investors, we document that investors with higher levels of trust exhibit less local bias in their stock portfolios, and the trust effect is evident among different groups of investors. Our finding survives a set of robustness checks. Furthermore, we find that local biases of small stocks and potentially informed investors are less affected by trust. We also find consistent evidence by employing the bilateral trust and the city-level trust measure with province fixed effects. Our findings shed light on how the cultural and societal forces affect investors financial decisions. We contribute to the literature following Guiso, Sapienza, and Zingales (2008), where they find that trust increases stock market participation. We further document that trust affects the local bias of individual investors, especially for their non-information-based trading behaviors. Given the fact that local bias (and the corresponding portfolio under-diversification) deleteriously affects the financial well being of most individual investors (see, for example, Goetzmann and Kumar 2008; Barber and Odean 2013), our paper has the following implication. Any policies or societal development promoting the cumulation of trust could also 22

23 indirectly benefit individual investors by reducing their local bias and hence improving their financial welfare. 23

24 Appendix Transaction-Based Local Bias Considering that many investors may engage in short-term trading and hold only a few stocks in their portfolios at the end of each month, the percentage of local stocks in portfolio holdings may not fully represent investors trading concentration. Hence, we define, in a similar spirit of Seaholes and Zhu (2010), the transaction-based local bias (T LB ijt ) for each investor (i) in province (j) at each month (t ) as the following: T LB ijt = T ijt T mjt, where T ijt measures the value of local stocks traded by investor i divided by the value of total stocks traded by the same investor during each month (t ), and the value is calculated using the buying and selling prices; similarly, T mjt measures the value of all the local stocks in province j traded in the entire market divided by the value of total stocks traded in the markets during month t. Then the yearly transaction-based local bias (T LB ijt ) at year (t) is calculated as the average of the corresponding monthly local bias (T LB ijt ) during that year. Again, we calculate this transaction-based local bias using the number of shares traded in addition to the value of the shares. Panel A of Table A.1 summarizes the transaction-based local bias and other variables in regressions during the sample period of 2010 to The mean values of TLBSHARE and TLBVALUE are around 0.04 and standard deviations are The transaction-based local biases have slightly larger means but smaller standard deviations than the measures based on portfolio holdings. Panel B shows the regression results of transaction-based local bias. It is comforting to see that all regressions yield statistically significant trust effects. For example, in Column (3), the coeffi cient of TRUST is 0.118, with a t-statistic of nearly 3. A one-standard-deviation increase in TRUSTB leads to a decrease in TLBSHARE by 24