Local bias in investor attention: Evidence from China's Internet stock message boards

Journal of Empirical Finance 38 (2016) 338–354

Contents lists available at ScienceDirect

Journal of Empirical Finance journal homepage: www.elsevier.com/locate/jempfin

Local bias in investor attention: Evidence from China's Internet stock message boards☆ Yuqin Huang a,⁎, Huiyan Qiu b, Zhiguo Wu c a b c

School of Finance, Central University of Finance and Economics, 39 South College Road, Haidian District, Beijing, PR China Faculty of Business and Economics, the University of Hong Kong, Porkfulam Road, Hong Kong Pacific Securities, Unit 3, Section D, Huayuan Enterprise Building, No. 9 Beizhan North Road, Xicheng District, Beijing, PR China

a r t i c l e

i n f o

Article history: Received 22 February 2015 Received in revised form 7 July 2016 Accepted 9 July 2016 Available online 12 July 2016 Keywords: Local bias Limited attention Internet stock message boards

a b s t r a c t In contrast to studies that focus on investment accounts, this study examines local bias in investor attention by analyzing messages posted by investors on China's Internet stock message boards. We find that individual investors pay more attention to stocks of local companies than to those of nonlocal companies. Local bias is particularly strong in underdeveloped regions, toward large, non-CSI 300, and low-turnover stocks and toward stocks with names that indicate their localities. The marginal effect of local bias is also considerably strong for distances within 500 km. (G10; G11; G14; G15). © 2016 Elsevier B.V. All rights reserved.

1. Introduction Previous studies have provided ample evidence of the “local-bias” puzzles based on the analyses of investment accounts. Both institutional and individual investors tend to invest in equities of local companies despite the well-documented benefits of diversification [Coval and Moskowitz, 1999; Huberman, 2001; Ivković and Weisbenner, 2005]. In contrast to studies that explored the geography of investors' investment portfolios, this study investigates the geography of individual investor attention. A key obstacle encountered by empiricists is the difficulty of measuring investor attention directly. The majority of researchers rely on news and events considered likely to attract investor attention.1 Da et al. (2011) use Google search frequency to measure individual investor attention directly. Similar to their innovative approach, a direct proxy using message posting activities on China's Internet stock message boards is designed in this study. Posting behavior is a revealed attention measure: if an investor posts a message relating to a specific stock on a message board, then he or she is genuinely paying attention to that stock. In addition, those who post on message boards are more likely to be real-world individual investors than

☆ We gratefully acknowledge helpful comments from anonymous referees, editor Daniel Ferreira, Pengjie Gao, Harrison Hong, Paul Po-Hsuan Hsu, Tong Li, Qiao Liu, Frank M. Song, Zhigang Tao, Ying Wang, Xueping Wu, Xianming Zhou, as well as the comments of conference participants at the 2012 China International Conference in Finance (CICF) and seminar participants at the University of Hong Kong, Central University of Finance and Economics. Yuqin Huang has been supported by National Natural Science Foundation of China (Project No. 71203246, 71673318), Program for Innovation Research in Central University of Finance and Economics and National Social Science Fund (Project No. 14ZDA044). ⁎ Corresponding author. E-mail addresses: [email protected] (Y. Huang), [email protected] (H. Qiu), [email protected] (Z. Wu). 1 Barber and Odean (2008) use news, unusual trading volume, and extreme returns. Grullon et al. (2004) and Chemmanur and Yan (2009) use advertising. Seasholes and Wu (2007) use price limit events. Yuan (2015) uses some market-wide attention-grabbing events, such as record-breaking moments and front-page articles about stock markets.

http://dx.doi.org/10.1016/j.jempfin.2016.07.007 0927-5398/© 2016 Elsevier B.V. All rights reserved.

Y. Huang et al. / Journal of Empirical Finance 38 (2016) 338–354

339

institutional ones, because the latter normally uses sophisticated and powerful information service providers, such as Reuters or Bloomberg. Chinese Internet stock message boards have a unique feature for this empirical study. Unlike their counterpart regulations in the United States and Europe, China's privacy law is not sufficiently comprehensive to protect Internet protocol (IP) addresses. Many Internet stock message boards explicitly display the IP information of posters (see Fig. 1). We can identify the location of posters based on their IP addresses using geolocation technology, and further categorize individuals according to whether they pay more attention to local companies. The first issue we investigate is the existence of local investor-attention bias. We construct local-bias measurements based on data of more than 24 million message postings related to publicly traded companies between July 2008 and June 2010. We determine that individual investors tend to spend most of their time studying the stocks of companies whose headquarters are

Fig. 1. China's Internet stock message board Guba Eastmoney. Notes: The figure is a screenshot of Guba Eastmoney, the most popular Internet stock message board in China. The board provides the IP information of posters who do not log in with a registered account.

340

Y. Huang et al. / Journal of Empirical Finance 38 (2016) 338–354

close to their own geographical locations. During our sample period, the average percentage of posts on a firm from posters in the same province as the firm is 9.75%, which is significantly larger than the average percentage of Internet users in the province (5.01%) and the average percentage of A-share investor accounts in the province (6.11%). For distance-related measures using the distance between the headquarters of a firm and the province of an investor, these measures show that the firms are located 9.85% and 9.04% closer to its actual posters than to its investors as indicated by benchmarks. Furthermore, distance plays a significant role: the marginal effect of local bias is considerably strong for distances within 500 km. Two mainstream explanations for local bias can be found in existing literature. The first is information-based, attributing local bias puzzle to the information asymmetry between local and non-local investors (Coval and Moskowitz, 1999, 2001; Ivković and Weisbenner, 2005; Bernile et al., 2015). The second is behavioral-based, considering that investors prefer local assets because of non-information reasons, such as familiarity or relative optimism (Huberman, 2001; Grinblatt and Keloharju, 2001; Ackert et al., 2005; Seasholes and Zhu, 2010). By the same taken, the sources of local bias of investor attention can also be twofold. Our analyses attempt to explore the characteristics and sources of local bias in investor attention. First, we examine the relationship between the measures of local bias in investor attention and various firm characteristics, including financial and trading variables. We determine that the local bias of investor attention is particularly strong when pertaining to large firms, lowattention stocks (e.g., non-CSI 300) and low-turnover stocks. Second, we construct a measure of local attention bias in each province to explore whether geographic characteristics and economic development are relevant to local bias in investor attention. This pattern of local bias in attention is strong in underdeveloped regions. The messages posted by registered users and local posters can attract a significant level of attention, as measured by the number of replies and clicks to each message. We also uncover an interesting phenomenon in which stocks whose names indicate their locality have a high level of local bias in investor attention. In summary, this study yields several key insights. First, although researchers have traditionally expended most effort on processing the composition of actively managed mutual funds or individual portfolios as a subset of brokerage data, this study proposes a new and direct measure of investor attention using messages posted by investors on China's Internet stock message boards. We confirm the existence of local investor-attention bias, which opens interesting possibilities to test the local bias from a new angle. Second, this research contributes to existing literature on Internet stock message boards by exploring the geographic information to study local bias. Researchers have devoted considerable effort in examining the potential effect of stock message boards on financial market. The literature on message boards has focused on several principal features, including message posting and market volatility (Antweiler and Frank, 2004); investor sentiment and stock returns (Das and Chen, 2007; Sabherwal et al., 2011); and the level of agreement among posting investor opinions, future stock return, and earning surprise (Chen et al., 2014). Chang et al. (2015) also use the same message board data to identify an important cultural and linguistic factor and show that investors living in linguistically diverse areas express diverse opinions on stock message boards and trade stocks actively. Huang and Li (2016) use a dataset to construct a measure of abnormal relative posting that reflects unusual changes in posting activities on stock message boards by local relative to non-local investors, to measure local information advantages. An increase in this measure predicts higher returns in the short term. Third, this study contributes to the empirical study of behaviors of individual investors in China. China has the largest Internet population in the world. A report by the China Internet Network Information Center (CNNIC) states that by June 2015, China had 668 million Internet users, with a penetration rate of 48.8%. However, despite extensive use of the Internet in China, few studies of the behavior of individual Chinese investors on Internet stock message boards have been undertaken. This study focuses on individual investors and their behavior, diverging from the traditional emphasis on developed equity markets. In contrast to the prominence of institutional investors in developed equity markets, the vast majority of investors in the Chinese market are individuals. The findings of this study shed new light on the factors influencing local bias, thereby helping both investors and researchers understand the psychology and behavior of individual investors. The remainder of this paper is organized as follows. Section 2 describes the source of data used to measure local investorattention bias. Section 3 describes the data and methods used to measure investor-attention local bias and the tests conducted to determine the existence of local bias in investor attention using different local-bias proxy variables. Section 4 examines the characteristics of investor-attention local bias. Section 5 focuses on investor attention as expressed in local posts. Section 6 concludes the study.

2. Data 2.1. Stock message board data With the largest population in the world, rapidly developing stock markets, and rapid growth in Internet users, China has particularly active Internet stock message boards. The message board data we use in this study come from Guba Eastmoney, the most popular stock message board in China.2 Guba Eastmoney officially went online in January 2006; however, it reached maturity and became increasingly popular from 2008 onward. We use a Web-scraper program to collect stock message-posting information from July 2008 to June 2010. 2 Guba Eastmoney (http://guba.eastmoney.com) has consistently ranked first in recent years according to statistics for unique visitors and page views. The other two popular sites are Guba Hexun (http://guba.hexun.com) and Istock JRJ (http://istock.jrj.com). All three sites offer an individual message board for each stock.

Y. Huang et al. / Journal of Empirical Finance 38 (2016) 338–354

341

Fig. 1 shows that for each message posted, the discussion board displays (1) the number of clicks, (2) the number of replies, (3) title of the original post, (4) author of the post, (5) time of the most recent reply (or that of the original post in case of no reply), and (6) date of the original post. The poster is represented by a registered account ID or a computer IP address. The majority of messages are posted by non-registered posters. For example, 27,325 and 49,580 messages are posted on the boards for the diesel engine firm Weichai Power (stock code: 000338.CH) and for the Bank of China (stock code: 601988.CH), respectively, with 80.3% and 85.2% of the messages posted by visitors without IDs. We use the location of the headquarters of each firm rather than their sites of domicile or incorporation as proxy for actual location because the majority of the operations of firms occur at their headquarters. Information on the headquarters of firms in our sample is obtained from the Wind database. We use IP address geolocation technology, particularly from the QQ database, which is the most popular proprietary IP address lookup database, to locate the users. Nie et al. (2008) demonstrate that this lookup database has 100% and more than 98% accuracies at the province and city levels, respectively. Two limitations of the design should be noted. First, the location of each poster is identified by their IP address using a proprietary IP address lookup database. An IP address may be shared by a group of users. For example, employees belonging to the same division of a company often share the same IP address. If multiple users access the Internet through the same subscriber account, then the IP address can possibly identify all their Internet traffic; therefore, the IP address cannot be linked to any specific user. This scenario does not alter our basic results, because our main concern is with the province- or city-level locations of posters. Second, all messages posted by an investor who moves from one city to another (and updates his or her location) are considered to originate from the second location. We only care about the source of postings and not the permanent residence of posters. We exclude all messages with IP addresses that indicate locations outside the mainland China. IP addresses are allocated extensively by various regional organizations; thus, each message poster corresponds to 1 of the 31 regions in China (provinces, autonomous regions, and municipalities). Finally, we collect data comprising 24,261,606 posts from 1598 listed companies, which cover the majority of listed companies in China at the time of data collection. Descriptive statistics for the sample are presented in Table 1. From July 2008 to June 2010, Table 1 Descriptive statistics of sample from Guba Eastmoney. Province

Number of IP posts

Percentage Number of listed companies

Total number of posts

Average number of posts to each stock

Average number of IP posts to each stock

Average percentage of IP posts

Min. posts for each stock

Max. posts for each stock

Anhui Beijing Chongqing Fujian Gansu Guangdong Guangxi Guizhou Hainan Hebei Heilongjiang Henan Hubei Hunan Jiangsu Jiangxi Jilin Liaoning Inner Mongolia Ningxia Qinghai Shandong Shanghai Shanxi Shaanxi Sichuan Tianjin Xinjiang Tibet Yunnan Zhejiang Total

648,355 1,282,335 373,621 894,687 130,019 3,314,626 458,771 91,083 111,461 663,652 313,713 722,715 956,407 591,246 1,615,933 486,025 303,595 682,774 87,763

3.13 6.18 1.80 4.31 0.63 15.98 2.21 0.44 0.54 3.20 1.51 3.48 4.61 2.85 7.79 2.34 1.46 3.29 0.42

55 110 29 54 21 200 25 17 20 35 26 38 61 48 116 26 33 50 18

608,740 2,742,708 357,846 778,976 267,737 3,094,508 350,198 236,263 489,470 582,824 383,372 521,806 849,960 696,017 1,336,672 294,027 511,294 712,172 363,127

11,068 24,934 12,340 14,425 12,749 15,473 14,008 13,898 24,474 16,652 14,745 13,732 13,934 14,500 11,523 11,309 15,494 14,243 20,174

9566 21,395 10,693 12,400 10,955 13,116 12,084 12,121 20,858 14,252 12,518 11,628 11,958 12,360 9851 9800 13,028 12,213 17,253

86.19 85.92 86.21 86.30 85.60 85.99 86.56 87.34 85.76 85.21 85.14 85.24 85.75 85.61 85.55 86.07 83.97 85.85 85.59

2973 2107 2928 3084 3707 1329 5059 5013 6607 3966 3397 3675 3653 2916 652 3847 2475 4739 7655

51,496 168,195 39,321 110,399 58,226 209,200 36,053 25,994 75,108 91,056 59,754 68,298 109,264 65,508 62,951 38,626 39,072 47,027 65,414

64,980 18,061 1,375,708 1,652,761 258,562 508,064 928,346 299,676 143,994 3951 183,068 1,579,676 20,745,629

0.31 0.09 6.63 7.97 1.25 2.45 4.47 1.44 0.69 0.02 0.88 7.61 100

11 10 94 154 28 29 66 29 32 8 26 129 1598

133,318 170,666 1,414,089 2,649,082 428,904 465,808 926,817 573,194 490,911 105,623 363,849 1,361,628 24,261,606

12,120 17,067 15,044 17,202 15,318 16,062 14,043 19,765 15,341 13,203 13,994 10,555 15,182

10,363 14,889 12,966 14,676 13,285 13,615 11,898 16,680 13,150 11,435 12,010 9056 12,982

84.87 85.99 86.49 85.11 86.69 85.24 85.48 85.85 85.53 86.48 86.15 85.89 85.77

4832 4308 3401 2542 4874 3669 1920 3275 4299 5793 4473 2468

27,350 62,064 59,780 135,400 70,200 69,927 86,442 103,700 45,738 23,979 60,195 42,598

Notes: This table provides a statistical summary of a sample comprising posts relating to 31 provinces and 1598 firms on Guba Eastmoney message boards from July 2008 to June 2010. This table reports the total number of IP posts from each province, percentage of IP posts, total number of firms for each province, total number of posts for firms in each province, average number of posts for firms in each province, average number of IP posts, average percentage of IP posts, and minimum and maximum of total posts by province.

342

Y. Huang et al. / Journal of Empirical Finance 38 (2016) 338–354

Fig. 2. Hourly distribution of message volume. Notes: This figure presents the distribution of stock messages throughout the day. The results are based on a sample of 20,992,143 stock messages. X-axis designates the hour of the day. Hour 23 represents the time from 11:00 PM to 11:59 PM. Y-axis represents the share of message volume. A substantial spike can be observed in the message volume during trading hours, indicating that individual investors use the stock message board to exchange trading ideas in real time.

15,182 messages are posted on average to each listed company, with 652 being the lowest and 209,200 being the highest. Investors in Tibet post the smallest number of messages (3951), whereas investors in Guangdong post the largest number of messages at 3,314,626 posts. Approximately 86% of the postings are from participants using only their IP addresses, which is stable across provinces. Stock message posting and stock-trading activities in China are closely related. This finding is consistent with those of Antweiler and Frank (2004) and Wysocki (1998). Message traffic varies by hour. Fig. 2 shows the distribution of total messages throughout the day. The x-axis designates the hour of the day. Hour 0 represents the time from 12 midnight to 12:59 AM, whereas Hour 23 represents the time from 11:00 PM to 11:59 PM. China's A-share market opens at 9:30 AM (Hour 9), breaks from 11:30 AM to 12:59 PM, and closes at 3:00 PM (Hour 15). The message traffic is the greatest during trading hours, and a temporary drop can be observed during the break at noon. The message traffic remains stable after the market closes until nearly midnight, and nearly disappears after that time. The message traffic shows another significant spike at around 8:00 AM, before the market opens. This observation provides further evidence that stock messages are used by individual investors to exchange relevant trading ideas in real time. For the two-year sample period, we obtain monthly trading volume data for individual firms from the Wind database. Fig. 3 presents a scatter plot of monthly trading volume against monthly message volume. A positive correlation between firm-level trading volume and message volume can be observed, with a correlation coefficient as high as 0.7382. Aggregated trading volume and message-posting volume also exhibit strong correlation. Fig. 4 depicts the monthly trading volume of the stock market and the monthly message volume of our sample. At the aggregate level, stock message posting and stock-trading activities move together in the majority of the months comprising the sample period. 2.2. Other data Internet user data come from the Twenty Fifth Statistical Report on Internet Development in China issued by the CNNIC.3 Data on the regional distribution of A-share investor accounts are collected from the China Securities Registration and Settlement Statistical Yearbook 2009.4 We translate geographic information for the headquarters of each company and the centroid of each province or city into latitudes and longitudes using the Google map APIs available at http://www.mygeoposition.com/. The relevant financial data, stock trading data, and provincial GDP per capita are collected from the Wind database. 3. Measuring investor-attention local bias In this section, we construct three measures of investor-attention local bias. Their summary statistics are also provided. 3.1. Local investor-attention bias Home bias or local bias is usually measured in terms of the deviation from a benchmark of no bias.5 Following previous work, we measure investor-attention local bias through the deviation of local investors' actual participation in message posting from a benchmark for that participation. Therefore, we use (1) the percentage of Internet users in a province (netizens) and (2) the 3 4 5

http://www.cnnic.cn/hlwfzyj/hlwxzbg/201007/P020120709345290787849.pdf http://www.chinaclear.cn/main/03/0305/1277889134195.pdf See, for example, Grinblatt and Keloharju (2001); Kho et al. (2009); Solnik and Zuo (2012)

Y. Huang et al. / Journal of Empirical Finance 38 (2016) 338–354

343

Fig. 3. Monthly message volume and monthly trading volume. Notes: This figure shows a scatter plot of firm-level monthly trading volume against firm-level monthly message volume. Logarithmic scales are used for each axis. The cloud of points shows a strong positive correlation between monthly message volume and monthly trading volume. The actual correlation coefficient is as high as 0.7382.

percentage of A-share investor accounts (investor accounts) in a province as benchmarks. benchmarkNj ¼

number o f Internet users in province j number of total Internet users in mainland China:

ð1Þ

benchmarkAj ¼

number of A−share investor accounts in province j number of total A−share investor accounts in mainland China

ð2Þ

We calculate the percentage of messages from each province as follows: pct ij ¼

the number of posts from province j for firm i : the number of total posts for firm i

ð3Þ

pctij gives the actual participation in the message posting of investors from province j on firm i. If the hypothesis of local bias in attention is true, then we expect to see more participation from investors in the province where the firm is headquartered (local province of the firm) than indicated by the benchmark. Thus, we compare the percentage of posts from the local province with the benchmark for this province to ascertain the local-bias effect. In particular, pct

LBi

¼

pct ijðiÞ benchmark jðiÞ

−1;

ð4Þ

where j(i) is the province in which firm i is headquartered and pctij(i) represents the actual participation of investors from the local province of firm i. In the absence of local bias, the variable is zero. The variable has a high value with great local bias. For Wuhan Steel (stock code: 600005.CH), which is headquartered in Wuhan City, Hubei Province, 11.67% of the posts originated

Fig. 4. Aggregate message volume vs. aggregate trading volume. Notes: This figure presents the time series of total message volume and market trading volume on a monthly basis.

344

Y. Huang et al. / Journal of Empirical Finance 38 (2016) 338–354

in Hubei Province, although only 3.82% of the total number of Internet users and 3.78% of the total number of A-share investors come from this province. Thus, LBpct for Wuhan Steel is 2.055 and 2.087 when using Internet users and A-share investor accounts i as a benchmark, respectively. 3.2. All investor-attention bias The participation of investors in message posting regarding nonlocal companies is expected to be lower than the benchmark. We construct a general measure to capture this effect: extension

LBij

¼

pctij −1: benchmark j

ð5Þ

Province j may be any province, not solely the province in which firm i is headquartered. If j is equal to j(i), namely, the local province of firm i, then the measure is the same as LBpct i . If not, we expect the local-bias effect to take a negative value. Each data point represents a firm-province combination. Some firms have no investors from underdeveloped regions, such as Tibet; thus, we finally obtain 46,436 firm-province combinations (i ≠j(i)). 3.3. Local bias expressed in terms of distance Coval and Moskowitz (1999) argue that geography plays a key role in economics and that home bias or local bias is a specialization of the effect of geographical proximity. They treat actual distance as “economic distance” and highlight the effect of distance on domestic portfolio choice. We create a distance-related measure to capture local bias in investor attention by following their work. First, we need to calculate the distance between two points representing the company and the location of the investor using the great-circle distance formula. For each firm, we choose either the province or city centroid of its headquarters as the point. For each investor, we use province or city centroid of post source. Thus, we can calculate the distance between provinces, between headquarters city of firms and provinces of posters, and between cities: h
i dist ij ¼ arccos sinLat i  sinLat j þ cosLat i  cosLat j  cos Long i −Long j πr=180;

ð6Þ

where Lati and Latj (Longi and Longj) represent the latitudes (longitudes) of the two points respectively, and r is the radius of the earth (r ≈ 6378 km). The distance between a firm and its investors is measured by a weighted average of the distance, for which the weight is either the percentage of actual participation in message posting (actual distance), the percentage of Internet users, or A-share investor accounts (benchmark distance):

actual

dist i

31

¼ ∑ pct ij dist ij ;

benchmark

dist i

ð7Þ

j¼1

31

¼ ∑ benchmark j dist ij : j¼1

ð8Þ

If investors focus significant attention on companies geographically close to them, then the actual weight of investors from provinces close to a company is higher than the benchmark weight. Therefore, compared with the benchmark distance, the actual distance between a firm and its investors is low. Thus, the ratio of the actual distance to benchmark distance captures local bias, if local bias exists. We construct the distance-related local-bias measure as follows:

distance

LBi

¼ 1−

dist actual i dist benchmark i

:

ð9Þ

is zero in the absence of local bias. The higher the value of the ratio is, the stronger the local bias of the The ratio LBdistance i investor attention is. The actual average distance between Wuhan Steel and its investors, distactual, is 700.26 km. The benchmark average distance is 852.98 km with the percentage of Internet users as weight. The benchmark average distance is 848.27 km with the percentage of A-share investor accounts as weight. Therefore, the distance-related local-bias measures for Wuhan Steel are 21.81% (benchmark 1) and 21.24% (benchmark 2).

Y. Huang et al. / Journal of Empirical Finance 38 (2016) 338–354

345

Table 2 Tests of investor-attention local bias. Panel A: LBpct, measuring local investor-attention bias N 1598 1598

Netizens Investor accounts

Average percent of actual local posts (A)

Average percent of benchmark (B)

A–B

LBpct (t-stat)

Ln(A/B) (t-stat)

9.75 9.75

5.01 6.11

4.73 3.64

1.20*** (33.66) 1.09*** (23.16)

0.62*** (42.86) 0.55*** (39.02)

(i ≠ j(i)), measuring all investor-attention bias Panel B: LBextension ij N

Benchmark

Mean

Median

Standard deviation

Minimum

Maximum

t-Stat

46,436

Netizens Investor accounts

−0.15*** −0.07***

−0.29 −0.18

0.76 0.76

−1.00 −0.99

47.75 43.18

−43.71 −21.01

Panel C: LBdistance, measuring local bias in distance (between provinces) N

Benchmark

Actual

Benchmark

Difference

1598 1598

Netizens Investor accounts

1032.84 1032.84

1134.33 1112.82

101.49 79.97

LBdistance 9.85*** 9.04***

t-Stat 45.28 40.95

Panel D: LBdistance, measuring local bias in distance (between provinces and firm headquarters cities) N

Benchmark

Actual

Benchmark

Difference

1598 1598

Netizens Investor accounts

1041.85 1041.83

1142.47 1121.27

100.61 79.44

LBdistance 9.76*** 8.96***

t-Stat 44.75 40.72

Notes: This table presents the results of local-bias testing using three different measures. Panel A summarizes the results of local investor-attention bias LBpct for 1598 firms in the sample. A local post is defined as a post whose IP address is from the same region (province, autonomous region, or municipality) as the headquarters of the firm. For each of the total 1598 firms in our sample, pctij(i) (A) is calculated and compared with benchmarkj (B) for three variables, namely, A − B, , calculated as the ratio of province A / B − 1 (LBpct), and Ln(A/B). Panel B provides summary statistics for the measure of all local investor-attention bias, LBextension ij j's actual fraction of firm i's total investors, pctij(i), to province j's benchmarks, minus one. Each data point represents a firm–province combination. Panels C and D present the test for local bias in terms of distance. The actual distance of Panel C represents the weighted average of the distance among provinces, weighted by the actual number of investors in each province. The benchmark's weights are the percentage of Internet users or A-share investor accounts. Difference is actual distance minus benchmark distance. LBdistance is constructed as one minus the ratio of the actual distance to the benchmark distance. The actual distance of Panel D is the weighted average distance between firm headquarters city and centroid of the province of its investors, weighted by the actual number of investors in each province. The weights of the benchmark are the percentage of Internet users or A-share investor accounts. *, **, and *** denote statistical significance at the 10%, 5%, and 1% levels, respectively.

3.4. Evidence of local bias in investor attention Table 2 provides the summary statistics for the local-bias measures. Panel A summarizes the results for local investor-attention bias. For each of the 1598 firms in our sample, we calculated pctij(i) (A) and compared it with benchmarkj(i) (B) for three variables, namely, Difference (A–B), LBpct, and Ln(A/B). We can observe the local-bias effects. The average percentage of actual local posters (9.75%) is always larger than the benchmark average percentage (5.01% for netizens and 6.11% for investor accounts). LBpct ranges from 1.09 to 1.20. Ln(A/B) ranges from 0.55 to 0.62. All measures consistently generate results indicating that individual investors tend to pay more attention to the stocks of local companies. The p-value for LBpct and Ln(A/B) is small, and the existence of local bias is statistically significant. Panel B presents the summary statistics for variables measuring the attention bias of all investors, LBextension . When investors ij live outside the province where the firm is located (j ≠j(i)), the average investor-attention bias LBextension is between −0.15 and ij −0.07 using the two benchmarks. These values are less than zero with statistical significance. Panel C summarizes the results for distance-related measure LBdistance, in which we calculate the distance between firm headquarters province and province of the poster. On average, the headquarters of each stock is located 1033 km from its actual investors, and 1113 km to 1134 km, as indicated by the benchmark composition. In percentage terms, the firms are located between 9.85% and 9.04% closer to its actual investors than to its investors, as indicated by benchmark investor composition. This finding shows that stocks are attractive to local investors. In panel D, we present LBdistance using the distance between firm headquarters city and the province of the investors. The firms are located between 9.76% and 8.96% closer to its actual investors than that indicated by benchmark investor composition. We construct LBdistance to test robustness using the distance between the firm headquarters city and poster city. In this approach, the weight is the percentage of the actual participation in message posting for each city. The mean values of LBdistance for benchmark netizens and benchmark investor accounts are 17.33% and 12.31%, respectively, with significant t-statistics. In summary, the tests for local bias in the attention of individual investors, as reflected in China's Internet stock message boards, provide strong and robust evidence using different local-bias measures. 4. Factors contributing to investor-attention local bias In this section, we analyze the factors that affect investor-attention local bias. First, we examine the effect of firm characteristics, such as financial and trading variables, on investor attention. We then consider whether geographic characteristics contribute

346

Y. Huang et al. / Journal of Empirical Finance 38 (2016) 338–354

to the phenomenon of investor-attention local bias. We cover two types of geographic characteristics: regional development and distance. Finally, we test whether stock names containing location information influence local bias in investor attention.

4.1. Local bias in investor attention and firm characteristics Early works (Coval and Moskowitz, 2001; Kang and Stulz, 1997) demonstrate that certain firm characteristics are significant determinants of local bias. For example, fund managers tend to hold securities in large firms in good financial condition because of the availability of transparent information on these companies. These findings confirm that advantageous information can partly explain the local bias behavior of institutional investors. In this section, we seek to determine whether advantageous information plays a similar role in motivating individual investors. We also test the relationship between firm-specific characteristics and investor-attention local bias. In the following analysis, the dependent variables are local investor-attention bias, LBpct i , and local bias in terms of distance, LBdistance , using the two different benchmarks (i.e., netizens and investor accounts). The financial characteristics of the i firms are incorporated in regressions, which are the same as those in Coval and Moskowitz (1999): firm size (market capitalization), leverage, current ratio, return on assets, and market-to-book ratio. Size is defined as the natural logarithm of the market capitalization LN(MV). Leverage and current ratio are a pair of accounting figures commonly used to measure firm distress. Leverage is defined as the ratio of total liabilities to total assets, and current ratio is the ratio of current assets to current liabilities. Leverage and current ratio capture the long-term and short-term financial health of a firm, respectively. Return on assets (ROA) of a firm is calculated by dividing earnings before interest and taxes by total assets. ROA is a good indicator of the profitability of a firm. The market-to-book ratio of a firm is defined as the ratio of the market value of its equity to the book value of its equity. It indicates the expectations of investors on the growth opportunities and potential profitability of a firm. Including this measurement allows us to determine whether local investors pay significant attention on local firms expected to have great opportunities to grow in the future. We also add the number of firm employees, a dummy variable for the CSI 300 index, the number of investor accounts, and turnover ratio as independent variables. The number of employees helps us to examine the local bias from the labor side of production. CSI 300 is defined as a dummy variable that is equal to one, if the firm is in the Shanghai-Shenzhen 300 index, or zero, if not. Firms in the CSI 300 index are well known nationally, and could possibly attract significant attention from investors across China. Thus, we expect firms in the CSI 300 index to show a small local-bias effect.

Table 3 Regression of firm characteristics on investor-attention local bias. LBpct BN

ln (MV) Leverage Current ratio Return on assets Market-book ratio ln (Employees) CSI 300 ln (Investor Accounts) Turnover Constant Adjusted R2 N

LBpct BA

LBdistance BN

LBdistance BA

(t-Statistic)

(t-Statistic)

(t-Statistic)

(t-Statistic)

0.181*** (3.60) 0.016 (1.12) −0.014 (−1.16) 0.053 (1.49) 0.002 (1.30) −0.026 (−1.06) −0.274** (−2.45) −0.027 (−0.68) −0.061*** (−4.60) −2.590*** (−2.63) 0.423 1539

0.269*** (3.83) 0.006 (0.29) −0.003 (−0.15) 0.029 (0.58) 0.004* (1.93) −0.069** (−2.02) −0.423*** (−2.71) −0.032 (−0.58) −0.060*** (−3.22) −3.974*** (−2.88) 0.355 1539

0.006** (2.16) 0.001 (0.64) −0.001* (−1.69) 0.002 (0.91) 0.000 (0.78) 0.001 (0.72) −0.018*** (−2.78) −0.002 (−1.07) −0.003*** (−3.65) 0.054 (0.98) 0.497 1539

0.007** (2.41) 0.001 (0.65) −0.001* (−1.94) 0.002 (0.89) 0.000 (0.76) 0.001 (0.68) −0.017*** (−2.64) −0.003 (−1.35) −0.003*** (−3.75) −0.033 (−0.58) 0.176 1539

Notes: This table reports the results of regressions for alternative measures of investor-attention local bias from the perspective of listed firms. We also control for pct cross-sectional persistence in investor involvement using dummy variables for provinces, whose coefficients are not shown in the table for clarity. LBpct BN and LBBA distance denote local investor-attention bias using netizens and investor accounts as the benchmarks. LBdistance and LB denote local bias expressed in terms of disBN BA tance using netizens and investor accounts as the benchmarks. *, **, and *** denote statistical significance at the 10%, 5%, and 1% levels, respectively.

Y. Huang et al. / Journal of Empirical Finance 38 (2016) 338–354

347

We also add the number of investor accounts to study the local bias from the investor side of participation. Turnover, which is defined as the ratio of the number of shares traded to the number of shares outstanding in the stock, is a useful measure of stock liquidity. The regression results are reported in Table 3. The coefficients of firm size, CSI 300 index, and turnover ratio are significant. The positive coefficient on size suggests local bias is stronger in large firms. Both the coefficients of CSI 300 dummy and turnover are significantly negative at 1% level. The possible reason is that CSI 300 stocks and high turnover tend to attract significant attention from all investors and to decrease local bias of attention. Unlike Coval and Moskowitz (1999), we find no evidence of relationships between the financial variables of firms and local bias in investor attention. 4.2. Local bias in investor attention and regional development In addition to firm characteristics, we consider whether geographic characteristics are relevant to local bias in investor attention. In this section, we provide another measure of local bias in each province, autonomous region, or municipality as follows:

province LB j


 0 Mean percentage of province j s posts in local firms
 −1: ¼ 0 Mean percentage of province j s posts in nonlocal firms

ð10Þ

We compare posters in local and nonlocal regions, which allows us to explore geographic trends in the distribution of local bias. The advantage of this calculation is that it does not require an external benchmark. We sort provinces into seven groups according to the degree of local bias LBprovince and plot a map of China in Fig. 5. In this map, the provinces, autonomous regions, and municipalities are colored with different intensities. Dark colors represent high levels of local bias. Fig. 5 shows a clear pattern, with mid-west regions of China exhibiting higher levels of investor-attention local bias than in the south-east regions. Table 4 presents the relationship between the investor-attention local bias and the variables related to regional development. province The measures of investor-attention local bias, namely, LBpct , are negatively correlated with variables related to reBA , and LB gional development, such as GDP per capita, Internet popularization rate, and the number of listed companies. This observation verifies that the investor-attention local bias is strong in many underdeveloped regions. One possible explanation is that in addition to their local firms, investors in developed regions have easier access to high-quality information on firms in other parts of

Fig. 5. Distribution of local bias in investor attention across regions of China. Notes: The figure presents the distribution of local bias across 31 regions (provinces, autonomous regions, and municipalities) in China. The provinces, autonomous regions, and municipalities are colored with different intensities according to the level of local bias in each region LBprovince. The darker the color is, the stronger the local bias in investor attention.

348

Y. Huang et al. / Journal of Empirical Finance 38 (2016) 338–354

Table 4 Pearson correlations for measures of investor-attention local bias and regional development.

Investor accounts Netizens Investor accounts LBpct BN pct LBBA LBprovince GDP per capita (yuan) Internet popularization rate (percent)

0.842***

LBpct BN

LBpct BA

LBprovince

−0.300 −0.373** −0.562*** 0.009 −0.387** −0.496*** 0.447** 0.347* 0.844***

GDP per capita (yuan) 0.312* 0.621*** 0.258 −0.337* −0.347*

Internet popularization rate (percent) 0.384** 0.663*** 0.231 −0.339* −0.349* 0.921***

ln (number of listed companies) 0.799*** 0.856*** −0.060 −0.512*** −0.722*** 0.634*** 0.656***

Notes: This table reports the Pearson correlation coefficients between various measures of investor-attention local bias and variables concerning regional development, such as GDP per capita, Internet popularization rate, and number of listed companies. The correlations among all of the variables are based on a sample of 31 pct regions (provinces, autonomous regions, and municipalities). LBpct BN and LBBA denote local investor-attention bias using netizens and investor accounts as the benchmark. *, **, and *** denote statistical significance at the 10%, 5%, and 1% levels, respectively.

the country. This scenario is one of the consequences of what is known as “information explosion era,” in which “a wealth of information creates a poverty of attention” (Simon, 1971). As a result, investor-attention local bias is slightly severe in underdeveloped regions. 4.3. Local bias in investor attention and distance effects A considerable number of studies on the threshold values of locality have been conducted in different countries. Although these threshold values vary among countries, they all measure “economic distance,” which denotes the driving distance of a daily round trip. Events that occur within this area are usually reported by various kinds of local media. For the United States, which has an area of 9,629,091 km2, Ivković and Weisbenner (2005) and Seasholes and Zhu (2010) consider approximately 402 km (250 miles) as an appropriate threshold value to distinguish local investors from nonlocal investors. For Finland, which has an area of 338,424 km2, Grinblatt and Keloharju (2001) regard 100 km as the threshold value of locality. We follow the method of Grinblatt and Keloharju (2001) in studying the threshold value of locality in investor attention of China. Fig. 6 shows that the coefficients are plotted using a group of distance interval dummy variables ranging from 0 km to 4000 km. These coefficients indicate the marginal distance effect on investor-attention local bias, LBextension . The present study ij tests three cases, with intervals representing 20, 30, and 50 km. Fig. 6 shows a sudden change in the slope of the distance coefficient at the 500 km point. However, the pattern of marginal distance effect on the local bias remains unchanged. We conduct regressions with distance regressors in the form of piecewise linear functions based on these figures. Table 5 is based on a regression in which the dependent variable LBextension , that is, the measure of local bias in all investor ij attention, was projected onto (1) dummy variables for each province (except one), (2) the maximum of the log of 500 km and the log of the distance (in kilometers) of province j from the province in which the headquarters of firm i are located, and (3) the minimum of the log of 500 km and the log of the distance (in kilometers) of province j from the province in which the headquarters of firm i are located. The use of a log function form for distance follows the method used by Grinblatt and Keloharju (2001). Table 5 reports the coefficients and t-statistics for the two distance regressors and a constant, with LBextension as the dependent ij variable. We also control for cross-sectional persistence in investor involvement as a result of developed-developing differences in familiarity with the stock market using dummy variables for provinces (coefficients are not reported). The regression exhibits a diminishing marginal distance effect within 500 km. The underlying t-statistic is as high as −54.98 with netizens as the benchmark and reaches −43.98 with investor accounts as the benchmark. The marginal distance effect of local bias is negligible for distances that exceed 500 km, but it remains statistically significant with t-statistics of −3.36 and −5.68 for two benchmarks. Our proposal of 500 km as China's threshold value of locality is compatible with the suggestions made in previous literature. The land area of China is 9,640,011 km2, which is roughly the same as that of the United States. A minimal difference exists between their locality cutoffs. The land area of China is 28.5 times larger than that of Finland, and the locality cutoff for China is five times that for Finland. The ratio of an area is the square of the ratio of its linear dimensions, such as its side, radius, or apothem. 4.4. Local bias in investor attention and stock name According to traditional financial theory, which assumes frictionless and efficient markets, stock names are irrelevant. However, anomalies are pervasive. For instance, Rashes (2001) presents considerable evidence for the co-movement of the MassMutual Corporate Investors (MCI) and MCI Communication funds, namely, two stocks with similar ticker symbols. Cooper et al. (2001) document a positive correlation between stock price movement and the announcement of changes in corporate names to popular Internet-related dotcom names during 1998 and 1999. Therefore, investors tend to pick their stocks in accordance with simple cognitive criteria, which simplifies a complex task.

Y. Huang et al. / Journal of Empirical Finance 38 (2016) 338–354

349

Fig. 6. Marginal distance effect of investor-attention local bias. Notes: The graphs plot coefficients depict the marginal distance effect of investor-attention local bias LBextension using a group of distance interval dummy variables ranging from 0 to 4000 km. The dependent variable in the regression LBextension is the ratio of provij ij ince j's actual fraction of firm i's total investors, pctij, to province j's benchmark fraction of firm i's total investors, benchmarkj, minus one. In this study, “benchmark fraction” signifies the percentage of investor accounts from the province. Each data point represents a firm–province combination. This study tests three cases, with intervals representing 20, 30, and 50 km. Similar results are determined when the percentage of netizens from the province is used as the benchmark. The results are available upon request.

In particular, whether stock names containing information on its geographic location affect investor attention and increase the message-posting activities of investors still needs to be addressed. We show that a stock name containing this information can induce a high level of local bias in investor attention; this finding is consistent with intuitive assumptions. Table 6 reports the summary statistics for the effect of stock name on local bias. Nameloca is a dummy variable that distinguishes the name locality of the stock. This dummy variable takes a value of one when the stock name contains information about its geographic location, such as the name of Shanghai Pharmaceuticals (stock code: 601607.CH); the dummy variable takes a value of zero when the stock name does not contain information regarding its geographic location, such as the name of the real-estate developer Vanke (stock code: 000002.CH). The results in Table 6 show that when the investor lives in the same province as the firm, the means of LBpct are 1.58 and 1.64 if the stock name contains information on the geographic location of the firm, whereas the means of LBpct are 1.07 and 0.90 if the stock name does not contain information regarding geographic location. We conclude that stocks whose names indicate their geographical location exhibit a stronger local-bias effect than those without geographical indication. Both the t-test and the Wilcoxon

350

Y. Huang et al. / Journal of Empirical Finance 38 (2016) 338–354

Table 5 Regression of marginal distance effect of investor-attention local bias. Benchmark netizens LBextension ij

Independent variables

LBextension Benchmark investor accounts ij

(t-Stat) −0.269*** (−54.98) −0.024*** (−3.36) 3.276*** (66.14) 0.395 48,014

Min [ln500, ln(distance)] Max [ln500, ln(distance)] Constant Adjusted R2 N

(t-Stat) −0.256*** (−43.98) −0.049*** (−5.68) 1.473*** (24.95) 0.178 48,014

Notes: This table reports the regression results for investor-attention local bias on the two distance regressors. The dependent variable is the local bias of all investor attention LBextension , each data point of which represents a firm–province combination. We also control for cross-sectional persistence in investor involveij ment as a result of developed–developing differences in familiarity with the stock market using dummy variables for provinces, whose coefficients are not shown in the table for clarity. *, **, and *** denote statistical significance at the 10%, 5%, and 1% levels, respectively.

test confirm that investor-attention local bias is considerably strong for firms whose stock names contain information on their localities. The other two measures of investor-attention local bias, namely, LBdistance and LBextension , provide similar results. ij Table 7 presents the results of multivariate regressions of stock name on investor-attention local bias. Panel A reports the results of a similar regression in Table 3. The dependent variables are LBpct and LBdistance with netizens and investor accounts, respectively, as benchmarks. All regressions show that the coefficient of nameloca is significantly positive at the 1% significant level. Firm size, CSI 300 dummy, and turnover ratio also remain statistically significant for all of the regressions. The contribution of these factors to investor-attention local bias strongly supports the hypothesis that local bias is affected by factors capable of attracting the attention investors. Panel B reports the results of a regression similar to that in Table 5. The dependent variable LBextension is projected onto ij (1) dummy variables for each of 30 provinces, (2) the maximum of the log of 500 km and the log of the distance (in kilometers) of province j from the province in which the headquarters of firm i are located, (3) the minimum of the log of 500 km and the log of the distance (in kilometers) of province j from the province in which the headquarters of firm i are located, and (4) the interaction terms between the dummy of stock name with locality and items (2) and (3). The regression results on dummies for each province, autonomous region, or municipality are not reported. The coefficient of the first locality slope dummy, “Min[ln500, ln(distance)] × dummy for firms with stock names indicating locality,” is −0.115 and −0.138 with netizens and investor accounts, respectively, as benchmarks. This finding indicates that the marginal distance effect of investor-attention local bias is greater for firms whose stock names indicate their localities below 500 km. Fig. 7 further supports the conclusion that a great distance effect is associated with firms whose stock names indicate their localities, because doing so makes them well known in the local area and capable of attracting the attention of local investors. In this study, we propose some explanations for the particular effect of stock name on local bias. First, the stock names of firms tend to evoke evaluative reactions, which affect the investment decisions of investors. A stock name that specifies locality may provide investors with a tangible connection to the firm, thereby evoking affective associations and predisposing investors to invest large amounts in local securities. Second, local investors can recognize a stock as a local investment opportunity when the Table 6 t-Test and Wilcoxon test of stock-name effect on investor-attention local bias.

Local bias

Benchmark

LBpct

Netizens

LBdistance

Investor Accounts Netizens

LBextension ij

Investor Accounts i ≠ j(i) Netizens Investor Accounts

Mean Median Mean Median Mean Median Mean Median Mean Median Mean Median

Stock name with locality

Stock name without locality

t-Test

Wilcoxon rank-sum test

Nameloca = 1 (N = 418)

Nameloca = 0 (N = 1180)

t-Statistic (p-value)

p-Value

1.58 1.16 1.64 1.03 12.22 12.09 10.39 10.13 (N = 12,186) −0.17 −0.31 −0.09 −0.20

1.07 0.69 0.90 0.50 9.01 8.85 6.79 6.28 (N = 34,250) −0.15 −0.29 −0.07 −0.17

5.92 (0.000) 7.03 (0.000) 6.22 (0.000) 8.59 (0.000)

0.000

−1.76 (−0.078) −2.09 (−0.037)

0.016

0.000 0.000 0.000

0.001

Notes: This table reports summary statistics for stock name effect on investor-attention local bias. Nameloca is a dummy variable that distinguishes the name locality of a stock. This table reports the sample mean and median values of LBpct, LBdistance, and LBextension , which are grouped by nameloca, the t-statistics arising from ij sample mean-comparison tests, and the z-statistics resulting from Wilcoxon tests conducted, respectively, to assess whether two groups come from the same distribution.

Y. Huang et al. / Journal of Empirical Finance 38 (2016) 338–354

351

Table 7 Regression of stock name effect on investor-attention local bias. Panel A: LBpct and LBdistance (percent) Independent variables

LBpct BN (t-Statistic)

LBpct BA (t-Statistic)

LBdistance BN (t-Statistic)

LBdistance BA (t-Statistic)

ln (MV)

0.165*** (3.38) 0.018 (1.30) −0.010 (−0.87) 0.053 (1.52) 0.002 (1.52) −0.034 (−1.42) −0.220** (−2.03) −0.060 (−1.57) −0.055*** (−4.30) 0.592*** (9.37) −1.993** (−2.08) 0.454 1539

0.255*** (3.67) 0.008 (0.39) 0.001 (0.04) 0.029 (0.58) 0.005** (2.07) −0.076** (−2.25) −0.376** (−2.43) −0.062 (−1.14) −0.055*** (−2.99) 0.530*** (5.89) −3.439** (−2.51) 0.369 1539

0.005* (1.90) 0.001 (0.81) −0.001 (−1.41) 0.002 (0.94) 0.000 (0.99) 0.001 (0.39) −0.014** (−2.36) −0.004** (−1.98) −0.002*** (−3.32) 0.034*** (9.54) 0.089 (1.64) 0.526 1539

0.006** (2.15) 0.001 (0.82) −0.001* (−1.68) 0.002 (0.91) 0.000 (0.98) 0.000 (0.36) −0.014** (−2.21) −0.005** (−2.28) −0.003*** (−3.42) 0.035*** (9.65) 0.002 (0.04) 0.224 1539

Leverage Current ratio Return on assets Market-book ratio ln (Employees) CSI 300 ln (Investor accounts) Turnover Nameloca Constant Adjusted R2 N Panel B: LBextension ij Independent variables

Min [ln500, ln(distance)] Max [ln500, ln(distance)] Min [ln500, ln(distance)] × dummy for firms with stock names indicating locality Max [ln500, ln(distance)] × dummy for firms with stock names indicating locality Constant Adjusted R2 N

LBextension ij

LBextension ij

Benchmark Netizens (t-statistic)

Benchmark Investor accounts (t-statistic)

−0.238*** (−42.47) −0.052*** (−6.78) −0.115*** (−11.41) 0.100*** (11.42) 3.280*** (66.29) 0.397 48,014

−0.219*** (−32.83) −0.082*** (−9.01) −0.138*** (−11.47) 0.120*** (11.51) 1.478*** (25.05) 0.180 48,014

pct pct distance Notes: Panel A reports the results of regressions for alternative measures of investor-attention local bias from the perspective of listed firms (LBBN , LBBA , LBBN , pct distance and LBdistance ). LBpct and LBdistance denote local BA BN and LBBA denote local investor-attention bias using netizens and investor accounts as the benchmarks. LBBN BA bias expressed in terms of distance using netizens and investor accounts as the benchmarks. The regressions were run across all firms on various financial characteristics, trading characteristics, and nameloca. Nameloca is a dummy variable that distinguishes the name locality of the stock. We also control provinces, whose coefficients are not shown in the table for clarity. Panel B reports the regression results for investor-attention local bias on two distance regressors. The dependent variable is the local bias of all investor attention LBextension , each data point of which represents a firm–province combination. We also control provinces, whose coefficients are not shown in the table for clarity. ij *, **, and *** denote statistical significance at the 10%, 5%, and 1% levels, respectively.

name of a stock contains information about its geographic location. Finally, stock name, which is the actual identity of an investment opportunity, can also serve as an advertisement. 5. Investor attention to local posts In 1971, when the world had no communication tools, such as e-mail, Google, Facebook, and Twitter, Simon wrote the following in Computers, Communications and the Public Interest: In an information-rich world, the wealth of information means a dearth of something else: a scarcity of whatever it is that information consumes. What information consumes is rather obvious: it consumes the attention of its recipients. Hence a wealth of information creates a poverty of attention and a need to allocate that attention efficiently among the overabundance of information sources that might consume it.

352

Y. Huang et al. / Journal of Empirical Finance 38 (2016) 338–354

Fig. 7. Real effect of stock name on investor-attention local bias. Notes: The graph plots coefficients depict the marginal distance effect of investor-attention local bias LBextension using a group of distance interval dummy variables ranging from 0 km to 4000 km. Separate sets of dummies for stock names that do and do not ij indicate locality are plotted. The dependent variable in the regression LBextension is the ratio of province j's actual fraction of firm i's total investors, pctij, to province ij j's benchmark fraction of firm i's total investors, benchmark j, minus one. In this study, “benchmark fraction” signifies the percentage of investor accounts from the province. Each data point represents a firm–province combination. Similar results are obtained when the percentage of netizens from the province is used as the benchmark. The results are available upon request.

The Internet stock message board is a platform for noisy talk. Initially, the posted content may give the impression that many of the messages posted are simply noise. However, we can still extract useful information by examining the characteristics and behavior of participants who receive relatively significant attention on the stock message board, and the resulting data are used for research purposes. First, we hypothesize that the messages posted by registered users, known as non-IP posts, capture significant attention from individual investors. We consider the participants who are willing to spend time registering an ID for message posting instead of posting messages directly. These individuals are more likely to be loyal participants in the online community than online passersby, and their messages are costly signals to some extent, particularly when these messages are compared with those posted by participants with only their IP address displayed. Thus, we can roughly distinguish the posts according to two separate categories: credible signals and cheap talk. Therefore, the “not-so-cheap talk” of registered users is expected to be informative and to attract significant attention from individual investors. We use the number of replies to each message and the number of clicks on each message as proxy measures of the attention of individual investors to explore the hypothesis. First, we use nonparametric techniques to rank all posts based on the number of replies and clicks separately. Then, we construct Rationon−IP/IP as the mean (median) rank of replies to, clicks on non-IP posts divided by the mean (median) rank of replies to, or clicks on IP posts. Table 8 reports the sample mean and median values of Rationon−IP/IP for replies and clicks, the t-statistics resulting from the sample mean comparison tests, and the z-statistics arising from the Wilcoxon tests to determine whether two post ranks come from the same distribution. The t-test indicates that the average Rationon−IP/IP is significantly larger than the ratio for both replies and clicks (p-values are both 0.000). The Wilcoxon test yields the same results, thereby confirming these findings. In brief, the ranking of non-IP messages is significantly higher than that of IP messages. Second, we expect that the messages posted by local investors can capture significant attention from individual investors. The information-based explanation of the local-bias puzzle is that local investors have easier access to information on firms than nonlocal investors by communicating with the insiders and suppliers of firms or through the local media. Given this local information

Table 8 t-Test and Wilcoxon rank-sum test of investor attention to local posts. t-Test of H0: the rank means of non-IP posts and IP posts are equal

Replies Mean Median Clicks Mean Median

Wilcoxon rank-sum test of H0: the medians of non-IP posts and IP posts are equal

t-Test of H0: the rank means of local posts and non-local posts are equal.

Wilcoxon rank-sum test of H0: the medians of local posts and non-local posts are equal

Rationon−IP/IP t-Statistic (p-value)

p-Value

Ratiolocal/non−local t-Statistic (p-value)

p-Value

1.128 1.127 1.097 1.106

0.000

1.024 1.024 1.029 1.033

0.000

69.35 (0.000) 41.52 (0.000)

0.000

13.69 (0.000) 12.06 (0.000)

0.000

Notes: All posts are based the separate categories of number of replies to and clicks on each message using nonparametric techniques. Rationon−IP/IP denotes the mean rank of local posts divided by non-local posts. Ratiolocal/non−local denotes the mean rank of non-IP posts divided by IP posts. This table reports the sample mean and median values of Rationon−IP/IP and Ratiolocal/non−local for replies and clicks, the t-statistics arising from sample mean-comparison tests and the zstatistics resulting from the Wilcoxon rank-sum tests conducted to determine whether two post ranks come from the same distribution.

Y. Huang et al. / Journal of Empirical Finance 38 (2016) 338–354

353

advantage, investors tend to hold large shares in locally headquartered firms. If this explanation valid, then we can expect to observe that local investors posting on message boards also possess a local information advantage. The messages posted by local posters should attract significant attention from readers on the board. We can indirectly test the information-based explanation for the local-bias puzzle. Similar to the previous test, we construct Ratiolocal/non−local as the mean (median) rank of replies to, clicks on local posts divided by the mean (median) rank of replies to, or clicks on non-local posts. Table 8 reports the sample mean and median values of Ratiolocal/non−local for replies and clicks. The t-test indicates that the average Ratiolocal/non−local is significantly larger than the one for both replies and clicks (p-values are 0.000). The Wilcoxon test yields the same results. In general, the rank of local IP messages is significantly higher than that of non-local IP messages. Rationon−IP/IP is also observed to be consistently larger than Ratiolocal/non−local for both measures (replies and clicks). This result indicates that the cheap-talk effect completely dominates the local-bias effect in our experimental setting. Therefore, we conclude that the messages posted by registered users are the most attractive posts, which are followed successively by local posts and nonlocal posts. Whether messages posted on the Internet stock message boards receive considerable attention is determined by the quality of information provided, among other factors. However, we cannot judge the quality solely from the numbers of replies to and clicks on a message because follow-up messages and replies may degenerate into meaningless conversation and even personal attacks that have little relation to the subject. Substantial analysis of the content of these messages should still be conducted. These analyses may even involve analysis of message sentiment: the use of computational methods to identify positive and negative opinions. However, these methods are beyond the scope of this study. 6. Conclusion Many studies indicate that investors tend to invest considerably in locally headquartered stocks. Using data from Chinese Internet stock message boards, we test whether investors pay significant attention to locally headquartered stocks. Our methodology involves forming multiple local-bias attention variables. We classify local and nonlocal posting by comparing the location revealed by the IP address of the poster with the headquarters of the firm. The use of IP address information of stock message boards to study local bias in attention is an aspect that sets our study apart from prior works. We find that local bias in attention exists. We examine the factors that affect investor-attention local bias and document that this bias is particularly strong in underdeveloped regions, large companies, non-CSI 300, low-turnover stocks, and stocks whose names indicate their localities. Distance plays a significant role: the marginal effect of local bias is considerably stronger for distances within 500 km. In addition, we show that posts from registered users and local users can attract significant attention. Our research is mainly descriptive. Our empirical evidence cannot provide specific causal claims for the source of local attention bias. We use the quantity of postings to analyze local bias in investor attention. Hence, some related questions may warrant future research. For example, the relationship between local bias of investor attention and stock return should be explored. The posted content should be also analyzed directly to disentangle sentiment and information. Future researchers may ask whether any link exists between local bias in investor attention and related well-known financial market phenomena. Our findings illustrate the usefulness of Internet stock message board data in financial applications. Message posting is an objective means of revealing and quantifying investor attention. We fully anticipate that exploring this dataset will continue to yield innovative insights into the behaviors of market participants and market rationality in the future. Appendix A. Supplementary data Supplementary data to this article can be found online at http://dx.doi.org/10.1016/j.jempfin.2016.07.007. References Ackert, L.F., Church, B.K., Tompkins, J., Ping, Z., 2005. What's in a name ? An experimental examination of investment behavior. Eur. Finan. Rev. 9, 281–304. Antweiler, W., Frank, M.Z., 2004. Is all that talk just noise? The information content of internet stock message boards. J. Financ. 59, 1259–1294. Barber, B.M., Odean, T., 2008. All that glitters: the effect of attention and news on the buying behavior of individual and institutional investors. Rev. Financ. Stud. 21, 785–818. Bernile, G., Kumar, A., Sulaeman, J., 2015. Home away from home: geography of information and local investors. Rev. Financ. Stud. 28, 2009–2049. Chang, Y.-C., Hong, H.G., Tiedens, L., Wang, N., Zhao, B., 2015. Does diversity lead to diverse opinions? Evidence from languages and stock markets. Rock Center for Corporate Governance at Stanford University Working Paper No. 168; Stanford University Graduate School of Business Research Paper No. 13–16. Chemmanur, T., Yan, A., 2009. Product market advertising and new equity issues. J. Financ. Econ. 92, 40–65. Chen, H., De, P., Hu, Y.J., Hwang, B.H., 2014. Wisdom of crowds: the value of stock opinions transmitted through social media. Rev. Financ. Stud. 27, 1367–1403. Cooper, M.J., Dimitrov, O., Rau, P.R., 2001. A Rose.com by any other name. J. Finance 56, 2371–2388. Coval, J.D., Moskowitz, T.J., 1999. Home bias at home: local equity preference in domestic portfolios. J. Financ. 54, 2045–2073. Coval, J.D., Moskowitz, T.J., 2001. The geography of investment : informed trading and asset prices. J. Polit. Econ. 109, 811–841. Da, Z., Engelberg, J., Gao, P., 2011. In search of attention. J. Financ. 66, 1461–1499. Das, S.R., Chen, M.Y., 2007. Yahoo! For Amazon: sentiment extraction from small talk on the web. Manag. Sci. 53, 1375–1388. Grinblatt, M., Keloharju, M., 2001. How distance, language, and culture influence stockholdings and trades. J. Financ. 56, 1053–1073. Grullon, G., Kanatas, G., Weston, J.P., 2004. Advertising, breadth of ownership, and liquidity. Rev. Financ. Stud. 17, 439–461. Huang, Y., Li, T., 2016. Local Information Advantage, Investor Attention, and Stock Returns. AFA San Fransico Meeting Paper. Huberman, G., 2001. Familiarity breeds investment. Rev. Financ. Stud. 14, 659–680.

354

Y. Huang et al. / Journal of Empirical Finance 38 (2016) 338–354

Ivković, Z., Weisbenner, S., 2005. Local does as local is: information content of the geography of individual Investors' common stock investments. J. Financ. 60, 267–306. Kang, J.K., Stulz, R., 1997. Why is there a home bias? An analysis of foreign portfolio equity ownership in Japan. J. Financ. Econ. 46, 3–28. Kho, B.C., Stulz, R.M., Warnock, F.E., 2009. Financial globalization, governance, and the evolution of the home bias. J. Account. Res. 47, 597–635. Nie, R., Yu, J., Zhang, H., LV, Y., 2008. IP address geomapping technique. Comput. Eng. 15, 102–104. Rashes, M.S., 2001. Massively confused investors making conspicuously ignorant choices (MCI–MCIC). J. Financ. 56, 1911–1927. Sabherwal, S., Sarkar, S.K., Zhang, Y., 2011. Do internet stock message boards influence trading? Evidence from heavily discussed stocks with no fundamental news. J. Bus. Financ. Account. 38, 1209–1237. Seasholes, M.S., Wu, G., 2007. Predictable behavior, profits, and attention. J. Empir. Financ. 14, 590–610. Seasholes, M.S., Zhu, N., 2010. Individual investors and local bias. J. Financ. 65, 1987–2010. Simon, H., 1971. Computers, Communications and the Public Interest. Johns Hopkins Press, Baltimore, pp. 40–41. Solnik, B., Zuo, L., 2012. A global equilibrium asset pricing model with home preference. Manag. Sci. 58, 273–292. Wysocki, P., 1998. Cheap Talk on the Web: The Determinants of Postings on Stock Message Boards, Working Paper. University of Michigan Business School. Yuan, Y., 2015. Market-wide attention, trading, and stock returns. J. Financ. Econ. 116, 548–564.