Introduction

SIFs play an essential role in Chinese capital markets. As of March 31, 2020, the market value of Chinese SIFs equity investments reached 13.56% of the total market value of stocks outstanding in Shanghai and Shenzhen. Due to the overall low level of internal governance of Chinese fund companies, the herding phenomenon of the asset allocation of Chinese SIFs, driven by fund ranking and personal interests, is exceedingly severe. Data from the Wind database shows that, in the first quarter of 2020, the number of single-shareholding funds in the top 20 stocks heavily held by Chinese SIFs both reached 138, with the number of funds holding the top3 Kweichow Moutai (SH600519), Wuliangye (SZ000858), and CATL (SZ300750) being higher than 95. Typically, high returns and low risk are always the goals pursued by fund managers, and thus they may make similar trading decisions based on reputation, compensation, and other considerations, which in turn trigger herding behavior (Lim et al., 2016; Yin, 2016; Lu et al., 2021).

Herding behavior can be established through direct communication in social networks (Apesteguia et al., 2020). Investors’ position in the network can impact their trading behavior (Ozsoylev et al., 2014), which will imitate and disseminate trading strategies through social networks, to some extent affecting the structure of traders and the stability of the market (Gong & Diao, 2023). As the foremost decision maker and soul of SIFs asset allocation (Hong et al., 2005), fund managers’ investment decision-making activities are embedded in various social networks. Such networks bring the potential for personal interaction and information exchange, which will influence fund managers’ behavior (Pool et al., 2015). In addition, the quarterly reports of SIFs show the synchronization of many funds increasing /decreasing their holdings of a particular stock simultaneously during the same investment cycle. It also suggests that many fund managers have obtained similar decision information in advance through different channels before asset allocation. Furthermore, the frequent emergence of fund managers’ “crosstown effect,” insider information, and alumni connections in the market further illustrates the deep-rooted “relationship” culture in Chinese society, which promotes the formation of their social networks.

This paper argues that the similarity of valid information obtained by various fund managers through their social networks leads to the convergence of their asset allocation behaviors, manifested as the herding behavior of concentrated shareholding. To deeply explore the formation mechanism of fund managers’ social networks, then further probe its influence mechanism on Chinese SIFs herding behavior, it is necessary to clarify and discuss several scientific issues. How do we measure the social network of SIF managers? What are the specific characteristics of the influence mechanism of fund managers’ social networks on the herding behavior of Chinese SIFs?

With this in mind, we first explore the influence of fund managers’ social networks on the herding behavior of SIFs by information asymmetry theory, behavioral finance theory, and social relations theory. Then we propose hypotheses from the perspectives of network centrality, size, constraint, and heterogeneity. Secondly, we adopt the CSAD model to measure the herding behavior of SIFs. Next, we construct a regression model to conduct empirical research. Finally, with the heterogeneity of herding behavior degree, gender, diploma, and other perspectives, we further compare the impact of fund managers’ social networks on the herding behavior of SIFs.

The main contributions of this paper are as follows. (i) From the perspective of the fund managers’ social network, we explore its influence on the herding behavior of SIFs and expand the research field of SIFs. (ii) Combined with information asymmetry theory, behavioral finance theory, and social relation theory, the theoretical mechanism of the influence of fund managers’ social networks on the herding behavior of SIFs is clarified. (iii) From the three dimensions of network centrality, network size, and network constraint, it constructs the measuring method of the social network of SIF managers. Using the quadratic coefficient of the CSAD model to measure the herding behavior of the fund and the regression model of the influence of fund managers’ social network on the herding behavior of SIFs is further constructed. (iv) From the perspective of herding behavior degree, gender, and other heterogeneity, we clarify the difference in the influence of the social network of fund managers on fund herding behavior.

The main research components of this paper are as follows: Section “Literature review and research hypothesis” presents the research question, then analyses the relationship between fund managers’ social networks and the herding behavior of SIFs at the theoretical level and proposes research hypotheses. Section “The effect of fund managers’ social networks on the herding behavior of SIFs in China” examines the impact of fund managers’ social networks on the herding behavior of SIFs. Section “Further study” is a further investigation of heterogeneity and concludes with the conclusions of this paper.

Literature review and research hypothesis

Literature review

Related research on social networks

(i) The impact of social networks on investors’ trading behavior

Hong, Kubik & Stein (2005) pointed out that there are more information transfer behaviors among crosstown fund managers, and the investment behaviors between them are more correlated. Cohen et al. (2008) empirically found that fund managers prefer stocks of listed companies whose executives are alumni. Colla & Mele (2010) found a significant positive correlation between the investment behaviors of investors with close network ties, while the opposite is true for more dispersed ones. Pool et al. (2015) believed that fund managers with social relationships such as colleagues and alumni have certain similarities in transaction behaviors, the closer they live, the more influenced each other. Baltakys et al. (2019) discovered that adjacent investors are likely to talk to each other and share information about each other’s trades, leading to similar trading behavior.

(ii) The influence of fund manager’s social networks on the SIFs performance

Xiao & Peng (2012) pointed out that fund managers who graduated from prestigious universities have better interpersonal resources, and alumni connections can bring them more valuable investment information. Berk & Binsbergen (2015) discovered that fund managers at the center of the network have the information advantages to improve investment performance. Liu & Su (2016) found that fund network structure will influence fund flow, which improves fund performance. Shen et al. (2016) argued that only direct and close alumni relationships could impact fund performance. Calluzzo & Kedia (2019) proposed that fund managers with social ties to executives push up the share prices of public companies through the funds they manage, thus allowing executives to profit.

Research on herding behavior of SIFs

(i) The formation mechanism of herding behavior of SIFs

Available literature explained the formation mechanism of the herding behavior of SIFs mainly from the perspective of reputation, compensation structure, and information flow. Bolton & Scharfstein (1990) based on a reputation perspective, stated that given the greater reputational risk involved in making independent decisions, fund managers are likely to align their trading strategies with most fund managers or more professional investors. From the point of view of information flow, Bikhchandani, Hirshleifer & Welch (1992) believed that the market price faced by investors is identical to public information, which makes fund managers fair information decision-making conditions (Deng, 2013). To obtain excess returns, some fund managers can observe the investment decision-making process of fund managers who may hold more private information. Maug & Naik (2011) argued that compensation systems through principal-agent theory harmed the trading behavior of fund managers, thus leading to herding behavior. In addition, Cheng, Xing & Yao (2022) suggested that the herding behavior of smaller funds is mainly caused by non-fundamental factors, while fundamental drivers primarily determine the herding behavior of Large-scale funds.

(ii) Herding behavior characteristics of Chinese SIFs

Chu & Qin (2008) measure the herding behavior of Chinese mutual funds and show that the herding behavior is more pronounced for bulk-holding stock, and the buying effect of such shares is weaker than the selling effect. Chen (2004) adopted the heterogeneity index to measure the degree of herding behavior of Chinese SIFs, it found that the fund return rate is positively proportional to the degree of herding behavior, but the size of outstanding shares is inversely proportional to it. By the LSV model, Zhang & Li (2005) discovered an inverse effect of the herding behavior of our closed-end funds on stock prices during stock trading and showed a certain degree of asymmetry (Gu et al., 2015). Furthermore, Wu et al. (2004) found a significant convergence in some “buy but not sell” transactions of Chinese SIFs, but there is no such phenomenon in “buy and sell” and “sell but not buy” transactions. Wang et al. (2012) explored the behavioral patterns of fund managers and showed that fund managers’ behavior of often using technical analysis rather than fundamental analysis leads to herding effects.

Related studies on herding behavior measurement models

(i) Development of herding behavior measurement models

It has been argued in the literature that there are two main types of approaches to measure herding behavior in financial markets. The first type is to explore the herding effect in market trading from a yield perspective, which mainly includes models such as CSSD and CSAD. When the market cross-sectional returns in the CAPM model are negatively linearly related to the average market returns (Christie & Huang, 1995), the CSSD model can be used to measure the herding behavior of investors in the stock market. Conversely, the CSAD model can measure herding behavior when a non-linear relationship between the two rates of return (Chang et al., 2000). Many studies have confirmed that the CSAD model is a better measure than the CSSD model, with its ability to fully account for asymmetries in the distribution of returns and its ability to capture weaker herding behavior in the market (Economou et al., 2011; David et al., 2019). The second type of herding behavior measure is mainly used to measure the herding effect of specific institutional investors, the most representative of which is the LSV model proposed by Lakonishok et al. (1992). The model measures herding behavior by comparing the actual shares of investors buying and selling decisions over the same period with the expected values under the independent trading assumption. For example, Caglayan et al. (2021) explored the herding of mutual funds in China using the LSV model and found that the herding effect of mutual funds significantly increases the convergence of return volatility across stocks. Guo et al. (2020) used the model to conduct an empirical study and found an intentional herding effect among institutional investors in the US.

(ii) CSAD model

The existing literature mainly uses the CSAD model to measure herding behavior in stock markets and funds. In the research on stock market herding behavior, Ahmed (2017) modified the CSAD model to include trading volume and investor sentiment as triggers of herding behavior, showing that there is significant herding behavior in the U.S. stock market during turbulent times. Lee et al. (2017) used the CSAD model to reveal the dynamic herding behavior of stock fund managers in the stock market in the U.S. Zhou et al. (2019) used the CSAD approach to test the herding effect in the Chinese stock market and found significant herding behavior in the Chinese stock market, indicating that contagion effects of investment do exist in the Chinese stock market. In the research on fund herding behavior, Cui et al. (2019) extended the CSAD model to closed-end funds and found that the herding effect of investors in U.S. closed-end funds is significant and mainly driven by non-fundamental factors. Ukpong et al. (2021) used the CSAD model to empirically show that the herding effect is driven by noise and is inversely related to funding size. Wang et al. (2021) used the CSAD model to explore and confirm that the herding effect exists in all types of open-end funds in China except income funds. Among mutual funds, growth funds tend to make investment decisions independently, while balanced and value funds exhibit herding behavior under certain circumstances (Cheng et al., 2022).

Theoretical analysis and research hypothesis

Theoretical analysis

(i) Fund manager social network and herding behavior of SIFs

Analysis from information asymmetry theory

According to the theory of information asymmetry, the gap between fund managers’ private information is one of the main reasons for herding behavior (Banerjee, 1992). When there is a more acute information asymmetry in the market, considering that other fund managers may have more private information advantages, some fund managers, to obtain excess returns or reduce the risk of independent decision-making, then choose to imitate the trading strategies of others. Simultaneously, the development of information technology and economic globalization make the scale of fund managers’ social networks expand rapidly. However, this may lead to more severe information asymmetry, which further aggravates the convergence of fund managers’ investment strategies and manifests in more serious herding behavior of SIFs (Jackson, 2008).

Analysis from behavioral finance theory

The herding effect is one of the main supporting theories of behavioral finance theory. The herding behavior of fund managers is affected to some extent by external factors, regular information, market noise, etc., as the main external factors, will be transmitted through the social network of investors. Simultaneously, driven by the maximum utility principle and avoiding losses caused by “selective bias” and “conservative bias”, fund managers may try their best to adopt social networks for information gathering. Furthermore, they will choose herding behavior to adjust to the “group pressure” decision-making psychology, which eventually leads to the emergence of blindly following others and irrational behaviors.

Analysis from social relations theory

Social relations theory states that the actions of each “social person” are influenced by diverse social relations, and individuals may form a complex social network organism when participating in economic activities. As a node of the social network, the SIF manager is embedded in many alumni, collegial and peer relationships, etc. Simultaneously, the transmission of information and of exogenous factors such as reputational pressure forces fund managers to discard and select messages from this complex social network when making decisions. Moreover, they may refer more to the trading behaviors of the other managers in neighboring nodes, which may lead to the convergence of the investment behaviors and eventually manifest as herding behaviors.

(ii) Measuring methods of social network centrality, network size, and network constraint

Network centrality

Centrality is an essential reference indicator in social network analysis, which judges the degree of centrality of an actor in a network and reflects the power that the actor has in the network. Since there are differences in the scale and number of members of various social networks, “relative centrality” is introduced as one of the principal indicators of social network measurement, avoiding the defect that “absolute centrality” cannot be compared across networks. The most academically accepted and applicable metrics are degree centrality, betweenness centrality, and closeness centrality. Table 1 shows the comparison of several different centrality indicators.

Table 1 Comparison of three centrality indexes.
Full size table

Network size

Coleman & James (1988) pointed out that social capital, the capital wealth belonging to social structural resources, often exists in individual social relationship networks and social groups. The larger the size of an individual’s social network relative to others in the network, the higher the individual’s social capital will be. Network size reflects the ability of an individual to obtain corresponding resources in the network. The larger the value of this index is, the more network members the individual is connected to, and the stronger its ability to master relevant information resources is (Ritter et al., 2002). When the scale of the network is large enough, the embedded relationship between individuals in the network increases the access channels of resources in the network. Consequently, it is conducive to the development of individuals in the network.

Network constraint

According to Burt (1995), if some network members are not connected, there are a certain number of structural holes among them. Individuals occupying positions of structural holes can minimize redundant connections between adjacent node individuals and obtain more information and resource advantages. The main indexes to measure the characteristics of structural holes include network size, network constraint, efficiency, and hierarchy. Among them, the index of “network constraint” can better reflect the control of each node, and in practice, this index is more accurate than the other three indexes. Hence, we can apply network constraints to measure the ability of individuals to use structural holes in their networks (Burt, 2004).

Research hypotheses

(i) The effect of fund manager social network centrality on herding behavior of SIFs

Liu et al. (2020) argued that fund managers’ degree of social network centrality significantly affects their information sharing and trading behavior. Lin et al. (2021) showed that the information advantage from social network centrality could influence fund managers’ investment style. Fund managers with a higher degree of centrality in the same social network have some influence on the dissemination of information in that social network and are more authoritative than other members, and therefore may also influence the decisions of other fund managers. Social networks with higher network centrality facilitate the dissemination of homogeneous information to a certain extent (Jackson & Rogers, 2007), and herding effects tend to occur in networks with higher centrality (Wang et al., 2021). In addition, fund managers with a higher degree of centrality can obtain valuable information at a lower cost, which reduces the occurrence of herding behavior to some extent. Therefore, we propose the following hypothesis.

H1 The social network centrality of SIF managers has a significant effect on their degree of herding behavior.

(ii) The effect of fund manager social network size on herding behavior of SIFs

The size of an individual’s network affects how well the individual can access resources in the network. The larger the network size, the more network members the individual is connected to, and therefore the more information resources the individual can access (Ritter et al., 2002). From the fund manager’s working background, fund managers have more job-hopping experience, and working in larger fund companies will have more colleagues. From the perspective of the educational background of fund managers, some fund managers do not disclose all the educational information in their resumes. Therefore, fund managers with relatively comprehensive information disclosure and more alumni in the same industry have more closely connected individuals, and it has a significant advantage in the alumni network. Combined with the various relationship networks of fund managers, the denser the network relationships presented in the fund manager’s resume data, the larger the size of in network. It will increase the fund manager’s source of information channels, which lead to effective processing of various information, manifesting as a reduction in herding behavior. Therefore, we propose the following hypothesis.

H2 The network size of SIF managers has an inverse contribution to their degree of herding behavior

(iii) The effect of fund manager social network constraint on herding behavior of SIFs

The social network constraint reflects an individual’s network information acquisition and external communication expansion ability. This indicator can better portray the tightness of direct or indirect relationships between individuals (Giuliani, 2013) and be adopted in empirical studies to explore the information mobility among individuals in social networks and the actual control of individuals over information dissemination. It found out after measuring the related indexes of fund manager social networks in China, the network constraint indicator that quantifies the degree of individual restriction captures the actual ability of the fund manager to access information and regulate the flow of resources in the network from a reverse perspective. Therefore, the network constraint indicator also simultaneously links the herding behavior of SIFs to the social network characteristics of fund managers from the view of information access. In general, the lower the constraint indicator, the more structural holes it occupies, which suggests that the fund manager has more information resources and thus can screen for more appropriate investment strategies. Therefore, fund managers with lower network constraints but more control over network information have information resources that are difficult to obtain by other fund managers, thus inhibiting the occurrence of herding behavior to a certain extent. Therefore, we propose the following hypothesis.

H3 The network constraint of SIF managers positively contributes to their degree of herding behavior.

(iv) The effect of fund manager’s social networks on the herding behavior of SIFs from a heterogeneity perspective

First, the number of male fund managers in China is twice as high as that of female fund managers. Since information exchange is relatively easy between the same gender, male fund managers differ to a certain extent from women in terms of investment strategies and risk preferences. It may affect the fund managers’ social networks on their herding behavior in terms of gender.

Secondly, nearly 80% of fund managers in China have master’s degrees, and the correlation between diplomas and fund managers’ business ability is close, which will affect their trading decisions. Meanwhile, the degree also impacts the characteristics of fund managers’ social networks through alumni relationships, so there may be differences in the influence of fund managers’ social networks on the herding behavior of SIFs with diverse levels of diploma. Finally, considering fund managers in the same region are more likely to exchange information and the efficiency of oral communication is high, it provides favorable conditions for fund managers in the same area to understand each other’s trading decisions. Meanwhile, the transmission of information in the same region will also reduce information asymmetry to a certain extent (Hong et al., 2005). Therefore, the impact of fund managers’ social networks on their herding behavior may also vary by region. When the degree of herding behavior is high or low, the specific characteristics of the influence of the fund manager’s social network on herding behavior may be different. On that account, findings on the impact of fund managers’ social networks on the herding behavior of SIFs may differ when the above factors change. Hence, we can obtain the following assumptions.

H4 When there are differences in the degree of herding behavior, gender, diploma, and region of SIF managers, the centrality of their social networks, network size, and network constraint have heterogeneous effects on the degree of herding behavior.

The effect of fund managers’ social networks on the herding behavior of SIFs in China

Data and model

We choose the daily return data of all open-end SIFs in China from July 6, 2012, to July 7, 2022, as the sample data source, based on which further screening is needed based on empirical research. As the daily returns of the money fund, bond fund, and part of the commingled fund are less variable, and the trend of net value fluctuation is smooth, the investment strategy is relatively stable. Therefore, we exclude the data of this part of the open-end fund and reserve the data of the equity fund and mixed partial stock fund, a total of 6078 entries. The data covers over thirty industries in China’s A-share market, including the semiconductor industry, restaurant and tourism industry, electric power grid industry, real estate industry, precious metals industry, electronic components industry, household appliances industry, retail industry, and other industries. Moreover, considering that most SIFs in China have long-held blue-chip stocks in Shanghai and Shenzhen, the sample data of market returns in this paper apply the daily returns of the Shanghai and Shenzhen 300 indices, data from the Wind database and Eastmoney.

Through preliminary screening of fund manager resume data, we deconstruct the fund manager social network into two parts—alumni relationship network and colleague relationship network. Alumni relationships are an essential part of people’s social relationships, and interpersonal relationships formed based on the same school significantly impact the behavior and decisions of financial practitioners (Guan et al., 2016). However, since some fund managers do not disclose their educational backgrounds or attend relatively cold overseas institutions, it is difficult to portray the overall structure of fund managers’ social networks through alumni relationships alone. It has shown that fund managers with a common brokerage background hold and trade more of the same stocks (Spilker, 2022). Given this, in addition to alumni networks, it is necessary to combine with colleague networks to construct the overall social network structure.

Regression model

First, considering the matching principle between variables, the explained variables need to be divided into serial data according to fund managers. Therefore, take the quadratic coefficient γ2 obtained by the CSAD method as the explained variable in this paper. The significant negative value of this index indicates the existence of herding behavior, and its size indicates the degree of herding behavior. Secondly, FreeClo (“Freeman closeness centrality”), ValClo (“Valente-Foreman closeness centrality”), Const (Network constraint), and Size (Network size), which can quantify the specific characteristics of fund managers’ social networks, are selected as explanatory variables. Finally, combined with the impact of individual features of fund managers on their investment decisions, the control variables of the model are determined as Sex (Gender), Dip (Diploma), Jnum (Number of schools attended), Snum (Number of funds under management).

$$\begin{array}{l}HB_{i,t} = \beta _0 + \beta _1FreeClo_{i,t} + \beta _2ValClo_{i,t} + \beta _3Const_{i,t} + \beta _4Size_{i,t}\\ \qquad\qquad+\, \beta _5Sex_{i,t} + \beta _6Dip_{i,t} + \beta _7Jnum_{i,t} + \beta _8Snum_{i,t} + \varepsilon _{i,t}\end{array}$$
(1)

where, HBi,t represents the degree of herding behavior of the SIFs managed by fund manager i at time t, namely, the quadratic coefficient γ2 of cross-sectional return deviation obtained through the CSAD model and market return rate. β0 is a constant term, β1, β2, β3, and β4 are the regression coefficients of the explanatory variables FreeCloi,t(FreeClo), ValCloi,t(ValClo), Consti,t(Const), and Sizei,t(Size), respectively. β5, β6, β7 and β8 are the regression coefficients of the control variables Sexi,t(Gender), Dipi,t(Diploma), Jnumi,t(Number of schools attended) and Snumi,t(Number of funds under management), respectively.

Explained variable

We take degree HBi,t of the herding behavior of each fund manager’s fund in the sample period as the explained variable. The herding behavior of fund managers refers to whether the trading strategies of the funds they manage are consistent with the tactics of most other fund managers, which can reflect by the trend of fund net value and return.

Based on Chang et al.’s theory (2000), the CSAD for testing herding behavior is defined as follows.

$$CSAD_{i,t} = \frac{1}{N}\mathop {\sum }\limits_{i = 1}^N \left| {R_{i,t} - R_{m,t}} \right|$$
(2)

where Ri,t is the return of the ith fund at moment t and Rm,t is the return of the average market portfolio return at moment t. CSADi,t is the cross-sectional absolute deviation of the single fund return from the market portfolio return at moment t.

Depending on the Capital Asset Pricing model, when the market price changes, various fund returns have different sensitivity to the average market return, which leads to increased diversification among fund returns. Simultaneously, the expected return rate of a single fund maintains a linear relationship with the expected return rate of the market. However, if there is significant herding behavior in the market, the deviation of an individual fund’s expected return from the market’s expected return is significantly reduced. Thus, this leads to less dispersion between fund returns, which destroys the linear relationship. Therefore, it establishes a regression model to test the linear relationship between the expected return rate of a single fund and the return rate of the market to judge whether there is herding behavior in the market. According to the asset pricing model, the expected returns of assets are as follows.

$$E\left( {R_{i,t}} \right) = R_f + \beta E\left( {R_{m,t}} \right) - R_f + \varepsilon _t$$
(3)

where Ri,t is the return of fund i in period t, Rf is the risk-free rate of return, Rm,t is the market return in period t, E(Rm,t) is the expected return of the market portfolio, and β denotes the systematic risk of the asset. According to Sun & Shi (2002), Eq. (3) can be deformed as

$$E\left( {R_{i,t}} \right) - E\left( {R_{m,t}} \right) = \left( {\beta - 1} \right)\left[ {E\left( {R_{m,t}} \right) - R_f} \right] + \varepsilon _t$$
(4)

Expected returns for a market portfolio consisting of N funds equally weighted

$$\frac{1}{N}\mathop {\sum }\limits_{i = 1}^N E\left( {R_{i,t}} \right) = R_f + \frac{1}{N}\mathop {\sum }\limits_{i = 1}^N \beta _i \times E\left( {R_{m,t} - R_f} \right) + \varepsilon _t$$
(5)

The deformation of Eq. (5) yields

$$\frac{1}{N}\mathop {\sum }\limits_{i = 1}^N \left| {E\left( {R_{i,t}} \right) - E\left( {R_{m,t}} \right)} \right| = \frac{1}{N}\mathop {\sum }\limits_{i = 1}^N \left| {\beta _i - 1} \right|\left[ {E\left( {R_{m,t}} \right) - R_f} \right] + \varepsilon _t$$
(6)

The left side of Eq. (6) equals the expected value of the absolute deviation of the cross-section of the return rate, then

$$E\left( {CSAD_{i,t}} \right) = \frac{1}{N}\mathop {\sum }\limits_{i = 1}^N \left| {\beta _i - 1} \right|\left[ {E\left( {R_{m,t}} \right) - R_f} \right] + \varepsilon _t$$
(7)

The first and second derivatives of both sides of Eq. (7) concerning E(Rm,t) can be obtained

$$\frac{{\partial E\left( {CSAD_{i,t}} \right)}}{{\partial E\left( {R_{m,t}} \right)}} = \frac{1}{N}\mathop {\sum }\limits_{i = 1}^N \left| {\beta _i - 1} \right| > 0$$
(8)
$$\frac{{\partial ^2E\left( {CSAD_{i,t}} \right)}}{{\partial E\left( {R_{m,t}} \right)^2}} = 0$$
(9)

Depending on the basic theory of the CAPM, the absolute deviation CSADi,t of the cross-section of asset returns is linearly increasing against the market returns Rm,t. If investors ignore their private information and imitate the trading behavior of other investors, CSADi,t and CSADi,t will present a non-linear relationship. By the above principles, the absolute deviation AVDi,t between the expected return of fund i and the market return in period t is

$$AVD_{i,t} = \left| {\beta _i - 1} \right|E\left( {R_{m,t} - R_f} \right) + \varepsilon _t$$
(10)

Further, CSADi,t expectations could be obtained

$$E\left( {CSAD_{i,t}} \right) = \frac{1}{N}\mathop {\sum }\limits_{i = 1}^N \left| {\beta _i - 1} \right|E\left( {R_{m,t} - R_f} \right) + \varepsilon _t$$
(11)

The above equation represents the equal-weighted average of the absolute deviation of the selected n sample funds’ returns from the market returns in period t, which can be applied as a reference indicator of the dispersion between the individual funds’ returns and the market returns of both. Depending on Eq. (11), E(CSADi,t) and E(Rm,t) are linearly positively correlated in an efficient market. However, investors will no longer maintain the original relationship if they make a consistent decision. Given that E(CSADi,t) and E(Rm,t) are difficult to measure, we adopt CSADi,t and Rm,t as the proxy variables of the first two, the total herding behavior of the n funds managed by the fund manager can be obtained as

$$CSAD_{n,t} = \frac{1}{N}\mathop {\sum }\limits_{i = 1}^N \left| {\beta _i - 1} \right|\left( {R_{m,t} - R_f} \right) + \varepsilon _t$$
(12)

Since it is difficult to meet the assumptions of the CAPM in practice, therefore, we adopt Feng (2011) for exponential function correction.

$$CSAD_{n,t} = \alpha + \gamma _1\left| {R_{m,t}} \right| + \gamma _2R_{m,t}^2 + \varepsilon _t$$
(13)

In this model, if γ2 is significantly negative, there is evidence of severe herding behavior, and the magnitude of γ2 can indicate the degree of herding behavior.

Since this paper intends to discuss the relationship between the social network of SIF managers and herding behavior, in the measurement process of herding behavior, we adopt the herding behavior of funds managed by fund managers to conduct a sample division. Therefore, in Eq. (12), CSADn,t denotes the cross-sectional return deviation of the fund managed by the nth fund manager at time t. Specifically, after the completion of the data selection, the CSAD values for each day of the individual funds managed by the sample fund managers were calculated separately. The cross-sectional absolute deviations of each fund manager’s managed funds are then averaged for each day of the sample period to obtain the CSAD index of the sample fund managers.

Explanatory variables

To explore the extent of the influence of fund managers’ social networks on the herding behavior of SIFs, we need to establish a system of measurement indicators around fund managers’ social networks, and appropriate explanatory variables need to be set accordingly. We finally select FreeClo, ValClo, Const, and Size as the four explanatory variables of the regression model.

The most widely used social network measure is centrality. A higher centrality metric means the more information resources the fund manager has at its disposal in the network. Since closeness centrality can portray the value and position of actors in a social network, this indicator is more applicable to the research of this paper. Moreover, Freeman (1978), Valente & Foreman (1998), and Sabidussi (1966) put forward three closeness centrality calculation methods, which are FreeClo, ValClo, and RecipClo, respectively. Since FreeClo can better characterize the value of individuals in the network, and ValClo measures the location of individuals from the perspective of network edges, these two indicators are more suitable for describing the value and location characteristics of individuals in fund managers’ social networks. Therefore, we select FreeClo and ValClo as explanatory variables.

Under the different perspectives of the analysis on social network heterogeneity, most indicators and centrality indicators have similar change trends. In terms of its final function of quantifying results, these indicators and centrality analysis indicators have a certain degree of overlap. To comprehensively describe the social network characteristics of fund managers, Const will also be selected as an explanatory variable in this paper. Compared with the centrality index, this index can quantify the extent to which fund managers are restricted in the social network and value the actual utility individuals can play. Furthermore, the three indicators, FreeClo, ValClo, and Const, are all relative indicators, so it should also consider the promotion or inhibition effect of network size on information transfer and includes the Size as an explanatory variable.

The indicators of the above explanatory variables were processed and calculated using UCINET software. Specifically, we extracted the alumni and colleague networks of the screened fund managers from the schools and workplaces disclosed in the fund managers’ CV data. Then we transformed the data and organized them into a matrix form recognizable by UCINET software through Excel, and finally substituted them into the software for calculation.

Control variables

From the perspective of personal characteristics of fund managers, this paper intends to adopt the Sex (Gender), Dip (Diploma), Jnum (Number of schools attended), Snum (Number of funds under management), Soum (Unit: billion) (Fund size under management), Garr (Geometric mean annualized return) as the control variables of the regression model.

The most basic natural attributes of a person are gender, age, height, and weight. Since the latter two physical characteristics do not affect the professional ability of fund managers, it excludes this part of the natural attribute index. Considering that most fund managers are in the prime of life between 30 and 45 years old, it is not easy to judge the role of age on the professional competence of fund managers, so this paper also does not consider the natural attribute of their age. Furthermore, studies have shown that gender factors impact individuals’ social behaviors. Male focus on the selection process and acquire information more purposefully, while females value the comprehensiveness of information collection and pay more attention to the opinions of others. Because of this, we select Sex as the control variable of the model. Secondly, excellent academic qualifications are one of the fundamental requirements of the fund industry, and therefore we take Dip as the primary control variable. Thirdly, some fund managers do not disclose or do not fully disclose information about the schools they attended, which may reduce the accuracy of the social network analysis results. Hence, it takes the Jnum in the fund managers’ CV data as a control variable. Moreover, since the explained variables are calculated by the CSAD method, and the fund management mode in China is mostly one person managing many funds, there must be herding behavior among these funds. Therefore, the number of funds managed by fund managers will also influence the results of herding behavior measured by the CSAD method, so we should also consider this factor (Snum) in control variables. Finally, since there are differences in the size of funds managed by different fund managers, considering the impact on performance, fund managers will be more prudent in deciding their investment strategies, which may consequently affect herding behavior. In addition, the geometric mean annualized rate of return provides an accurate measure of the actual return on fund performance and, to some extent, reflects the personal risk appetite characteristics of the fund manager. Therefore, we include the fund size under management (Soum) and the geometric mean annualized return (Garr) as control variables in the model in further studies.

Herding behavior of SIF managers

Given the number of fund managers in China and a lengthy sample span, measuring the herding behavior of each fund manager’s managed funds is high volume and difficult. Therefore, we conduct random screening on the sample data after processing and finally choose the daily rate of return data of 250 fund managers, using the CSAD method to measure the herding behavior of SIFs. Among them, 100 fund managers managed securities with significant herding behavior. Due to space limitations, only the measurement results of 100 fund managers whose CSAD quadratic coefficient γ2 is significantly negative in Eq. (13), as shown in Table 2. Since Newey-West regression helps to refine the errors, both t-statistics and standard errors in this paper based on this method were obtained (Tan et al., 2008).

Table 2 Herding behavior measurement of SIFs from the perspective of fund managers.
Full size table

As the results of the herding behavior measured in Table 2, the quadratic coefficients γ2 of the funds managed by these 100 fund managers are most in the range of (−100, −20), which indicates that the herding behavior of most of the funds managed by fund managers is at the same level. However, among these 100 fund managers, the absolute values of the quadratic coefficients of “Feng Jiang” and “Yixiang Fu” are too large, 483.277 and 312.029, respectively. It shows that the daily returns of the funds managed by these two managers have a negative non-linear relationship with the market returns, which means there is a significant herding behavior. Overall, the sample of most fund managers of herding behavior is at medium levels, but individual fund managers who manage funds still have a higher degree of herding behavior.

Meanwhile, the test results of the remaining 150 show that their quadratic term coefficients are insignificant or significantly positive. It indicates that most funds managed by fund managers in the random sample do not have significant herding behavior when using the CSAD method to test the herding behavior of Chinese SIFs. Since the CSAD applies to markets with more volatile prices (Sun & Shi, 2002) and may underestimate the overall level of herding behavior in the actual measurement process, the test results in this section may deviate from the actual herding behavior of Chinese SIF managers. Considering that the CSAD method can quantify the herding behavior of SIFs from the perspective of fund managers, the subsequent research does not take these 150 fund managers as the research objects, thus overcoming the defects of the CSAD model.

An empirical study on the influence of fund manager social networks on the herding behavior of SIFs

Descriptive statistics of variables

The results of descriptive statistics for each variable are in Table 3. The explained variable HBi,t is the quadratic coefficient obtained by the CSAD method in the section “Data and model.” Sex (Gender), Dip (Diploma), Jnum (Number of schools attended), and Snum (Number of funds under management) are derived from fund manager biographical data, and the first two are transformed into dummy variables via text data. The “0” and “1” of Sex mean “female” and “male,” respectively. The Dip variable values “16”, “19”, and “21” represent the average years of attainment for each degree, which means “undergraduate,” “master,” and “Ph.D.” degrees, respectively.

Table 3 Descriptive statistics of variables.
Full size table

From Table 3, the extreme values of the explanatory variable HBi,t have a significant difference and high standard deviation. The explanatory variable “Valente-Foreman closeness centrality” has a higher value, and “Freeman closeness centrality” (FreeClo) has a mean value of 1.357, and network size (Size) has a maximum of 6.289 and a minimum value of 0. In addition, the network constraint (Const) has relatively small, with the minimum values of both Dip and Snum are not less than 1.

Regression analysis

According to the previously constructed model of the influence of fund managers’ social networks on the herding behavior of SIFs, it brought the sample data with 100 groups of variables in for empirical regression. It uses StataMP 16 software to implement the regression process, and the regression results are in Table 4.

Table 4 Regression results of fund manager social network on herding behavior of SIFs.
Full size table

Studies have shown that a higher level of social networks promotes the dissemination of homogenous information to a certain extent (Hana et al., 2013). As can be seen from the regression results in Table 4, the regression coefficients of FreeClo and ValClo are 108.798 and −3.916, respectively. The absolute values of the two coefficients differ by a bit large margin and do not pass the significance test, which suggests that the centrality of SIFs has no significant effect on the herding behavior of SIF managers, and hypothesis H1 is not confirmed. It may be because fund managers with higher network centrality are more dominant in the network, and their access to information is more valuable. Therefore, their investment decisions are less likely to be disturbed by other information in the market, thus showing no significant effect of FreeClo and ValClo on herding behavior HB. In addition, Table 4 shows that the coefficient of the variable Size is 16.970 and significant at the 5% confidence level. Hence, network size is one of the main reasons for the impact of fund managers’ social networks on herding behavior, and hypothesis H2 was tested. Further, the regression coefficient of the variable Const in Table 4 is −16.890 but insignificant. It suggests that there is no significant effect of the fund manager’s network constraint on the degree of herding behavior of SIFs, and hypothesis H3 is not tested. It may be because the network constraint mainly reflects individuals’ network information acquisition ability and external communication expansion ability, then can better portray the tightness of direct or indirect relationships between individuals. Given this, fund managers with higher network constraints have lower status and influence in the network, and their speed of information acquisition is relatively low, which makes it more difficult for fund managers to be informed of others’ trading strategies and thus cannot provide the necessary information base for herding behavior.

As with ValClo, which has a negative regression coefficient, the regression result for Const is also negative. Since the explained variables were negative, both had an enhancing effect on herding behavior. When the variable Const takes a high value, the fund manager has a single source of information and is prone to make decisions blindly based on the information currently available, so the variable Const shows a positive relationship with the explanatory variables. Since the regression results for network size (Size) in Table 4 are significant and positive, indicating that fund managers with larger networks have a stronger information access advantage, which leads to herding behavior, they can also use their information advantage for effective information screening in this condition, which makes the relationship between Size and HB negative. In addition, Table 4 also shows that fund managers with higher network centrality (FreeClo and ValClo) have a better network position and information access advantage, and they are less susceptible to interference from external information in the investment, therefore, less likely to form herding behavior, which shows that the regression results of variables FreeClo and ValClo are not significant with variable HB, respectively.

Further study

To further explore the impact of fund managers’ social networks on the herding behavior of SIFs, the heterogeneity investigations will conduct from the perspectives of the degree of herding behavior, gender, diploma, and region.

An investigation on the perspective of the degree of herding behavior heterogeneity

Regression model

Depending on the descriptive statistical results of variables mentioned above, the standard deviation of explained variable HBi,t, which measures the degree of herding behavior of SIFs, is high, and there are too small extreme values in the sample data. Considering the various degrees of herding behavior, the social networks of fund managers may have varying effects on the herding behavior of SIFs. Therefore, sort the 100 samples data according to the size of HBi,t, then divide it into two groups of high and low of the difference of herding behavior degree. The regression models of high and low herding behavior degrees will be respectively established with the following equations.

$$\begin{array}{l}HB_{j,i,t} = \beta _{j,0} + \beta _{j,1}FreeClo_{j,i,t} \\\qquad\qquad+\, \beta _{j,2}ValClo_{j,i,t} + \beta _{j,3}Const_{j,i,t} + \beta _{j,4}Size_{j,i,t}\\ \qquad\qquad+\, \beta _{j,5}Sex_{j,i,t} + \beta _{j,6}Dip_{j,i,t} + \beta _{j,7}Jnum_{j,i,t} + \beta _{j,8}Snum_{j,i,t}\\ \qquad\qquad+\, \beta _{j,9}Soum_{j,i,t} + \beta _{j,10}Garr_{j,i,t} + \varepsilon _{j,i,t}\end{array}$$
(14)

where j is taken as h and l in turn, denoting the grouping of high and low herding behavior, respectively, HBj,i,t represents the explained variable, the degree of herding behavior. The other variables are the same as defined in Eq. (1).

Empirical analysis

The regression results of high and low herding behavior degree groups are in Table 5.

Table 5 Regression results from the perspective of heterogeneity of herding behavior degree.
Full size table

According to Table 5, we can see that the regression results of the higher group of herding behavior showed that the regression coefficients of the variables FreeClo, ValClo, and Size were 247.284, −7.449, and 27.713, respectively. The regression results of the lower group showed that the regression coefficients of FreeClo, ValClo, and Size were 20.234, −0.676, and 1.795. The regressions of network size (Size) are significant in the higher group of herding behavior, while none of the variables in the lower group passed the significance test, indicating that from the perspective of network size, social networks had a greater influence on herding behavior in the higher group.

As with the regression results under the whole sample in Table 4, the regression results for the network constraint (Const), “Valente-Foreman closeness centrality” (ValClo), and “Freeman closeness centrality” (FreeClo) in both the higher and lower groups in Table 5 are not significant. Meanwhile, the coefficient of the variable Size is significantly positive at the 5% confidence level in the regression results for the group with higher levels of herding behavior. It indicates that the individual network size of fund managers has a negative contribution to herding behavior at higher levels of herding behavior, but such an effect is not significant at lower levels of herding behavior. Since individual network size reflects the number of individuals in the fund manager’s social network with whom it has strong ties when the fund manager has connections with more individuals, its influence on the overall network is also greater, and while obtaining information from many other fund managers, it also gives it a certain degree of control over the network, so the individual network size of the fund manager can provide an information base for herd behavior, which shows a significant influence relationship between network size (Size) and the degree of herding behavior (HB).

In summary, the positive and negative and significance levels of the coefficients of the variables changed to different degrees under the perspective of differences in the degree of herding behavior, which indicated that the centrality of social networks, network size, and network constraint had heterogeneous effects on the degree of herding behavior, which partially verified hypothesis H4. In addition, “Valente-Foreman closeness centrality” and individual network size have different degrees of facilitating or inhibiting effects when the degree of herding behavior of SIFs is high.

An investigation on the perspective of gender heterogeneity

Regression model

Research by Atkinson et al. (2003) confirmed that gender differences lead to differences in trading habits and that men are more risk-taking (Niessen & Ruenzi, 2019), which leads to heterogeneity in the social network structure and characteristics of male and female fund managers. Meanwhile, some scholars have found that female fund managers are more risk-averse (Bliss & Potter, 2002). Though there is no inherent difference in skills between female and male fund managers, only better performing female managers will be able to continue managing the fund (Rajesh & Nicole, 2016). Given this, we divided the sample data into two groups of male and female fund managers from the perspective of gender differences in fund managers in this section, and separate regression models are developed and empirically analyzed. The sample consisted of 85 male fund managers and 15 female fund managers.

$$\begin{array}{l}HB_{j,i,t} = \beta _{j,0} + \beta _{j,1}FreeClo_{j,i,t} \\\qquad\qquad+\, \beta _{j,2}ValClo_{j,i,t} + \beta _{j,3}Const_{j,i,t} + \beta _{j,4}Size_{j,i,t}\\ \qquad\qquad+\, \beta _{j,5}Cnum_{j,i,t} + \beta _{j,6}Dip_{j,i,t} + \beta _{j,7}Jnum_{j,i,t}\\ \qquad\qquad+\, \beta _{j,8}Snum_{j,i,t} + \beta _{j,9}Soum_{j,i,t} + \beta _{j,10}Garr_{j,i,t} + \varepsilon _{j,i,t}\end{array}$$
(15)

where j is taken as n and f in turn, denoting the female and male fund manager groups, respectively, HBj,i,t represents the explained variable, the degree of herding behavior. The other variables are the same as defined in Eq. (1). Furthermore, since the grouping is through gender in this section, the control variable Sex in Eq. (1) is replaced with the number of fund managers’ inaugural firms Cnum in Eq. (15).

Empirical analysis

The regression results of the female and male fund manager groups of samples are shown in Table 6.

Table 6 Regression results from a gender heterogeneity perspective.
Full size table

As can be seen from Table 6, in the regression results of the male fund management group, the regression coefficients of the variables FreeClo, ValClo, and Const are 80.584, −3.181, and −14.459, respectively, and only the network size (Size) is significant at the 5% confidence level. While in the female fund manager group, the regression coefficients of the variables FreeClo, ValClo, and Const are 808.568, −20.573, and −6.194, respectively, and the regression coefficient of the variable Size is 29.469, all four explanatory variables are not significant. In both samples, the positive and negative coefficients of the regression results for male and female fund managers are the same as in Table 4.

According to social psychology theory, investors, as individuals in society, are often influenced by social psychology in their trading strategies. Although fund managers have a high level of expertize compared to individual investors, some fund managers may abandon independent decision-making and imitate others’ trading strategies due to personal reputation and performance compensation considerations. Research has shown that women prefer conservative trading strategies (Charness & Gneezy, 2012), while men place more emphasis on “success” and “reputation” in their careers due to social values and other factors (Blazina, Pisecco & O’Neil, 2005; Jiang & Lv, 2012). Given this, male fund managers have a higher risk appetite than female fund managers and bear more pressure from social values. When faced with the problem of information asymmetry and peer competition, male fund managers may engage in herding behavior due to the overlay of stress from others having access to more information. Moreover, female fund managers are more likely to be risk-averse and make independent decisions with more information about their peers. Therefore, the regression results in Table 6 show that only the variable Size in the social network of male fund managers is significant with the degree of herding behavior HB, while all four indicators of the social network of female fund managers, FreeClo, ValClo, Const, and Size, are not significant with HB.

In summary, the significance levels of the regression results are different for males and females, which argues hypothesis H4 to some extent. In addition, since male fund managers account for a higher proportion of the whole sample, the regression results of the male fund manager group after grouping by gender are consistent with those before grouping. Female fund managers will make more independent decisions because of risk aversion and other reasons, which shows that their social network doesn’t have a significant facilitating or inhibiting effect on their herding behavior.

An investigation on the perspective of diploma heterogeneity

Regression model

The percentage of fund managers with master’s degrees in China is about 84%, while those with doctoral and bachelor’s degrees are about 16%. Considering that there are fewer fund managers with bachelor’s degrees and fewer bachelor’s degree fund managers in charge of equity and equity-biased hybrid funds, this section focuses on the heterogeneity of master’s and doctoral degree fund managers. From the 100 samples in Table 2, fund managers with master’s and doctoral degrees are screened separately based on their diploma levels, with 87 of the former and 12 of the latter. Subsequently, we model the influence of social networks of fund managers with master’s and doctoral degrees on the herding behavior of SIFs as follows.

$$\begin{array}{l}HB_{j,i,t} = \beta _{j,0} + \beta _{j,1}FreeClo_{j,i,t}\\ \qquad\qquad+\, \beta _{j,2}ValClo_{j,i,t} + \beta _{j,3}Const_{j,i,t} + \beta _{j,4}Size_{j,i,t}\\ \qquad\qquad+\, \beta _{j,5}Sex_{j,i,t} + \beta _{j,6}Jyear_{j,i,t} + \beta _{j,7}Jnum_{j,i,t} + \beta _{j,8}Snum_{j,i,t}\\ \qquad\qquad+\, \beta _{j,9}Soum_{j,i,t} + \beta _{j,10}Garr_{j,i,t} + \varepsilon _{j,i,t}\end{array}$$
(16)

where j takes m and d in turn, representing the master’s degree and Ph.D. degree groups, respectively. HBj,i,t is the explained variable, the degree of herding behavior. Other variables and constant terms have the same basic meaning as Eq. (1). In addition, since this section is grouped based on diploma, the control variable Dip in Eq. (1) is replaced with the fund manager’s years of experience Jyear in Eq. (16).

Empirical analysis

The regression results of the master’s degree and Ph.D. degree groups are as Table 7.

Table 7 Regression results from a diploma heterogeneity perspective.
Full size table

According to Table 7, in the regression results of the master’s degree fund management group, the positive and negative coefficients of the four explanatory variables are the same as those in Table 4, and only the network size (Size) is significant at the 5% confidence level. However, all the variables of the doctoral degree fund manager group are not significant at the 10% confidence level, which is completely different from the previous findings. Given this, in the heterogeneity investigation from the perspective of diploma differences, the regression results for the master’s degree fund manager group are generally consistent with those of the whole sample. However, the doctoral degree fund manager group does not show a significant effect of social networks on herding behavior. Some scholarly studies pointed out that fund managers’ performance is correlated with the ranking of the school they attended but is less correlated with whether they have obtained a Ph.D. degree. (Gottesman & Morey, 2006). While, some studies have argued that there is a difference in investment style and performance level between Ph.D.—and master’s-educated fund managers (Berk & Binsbergen, 2015), this difference may have different performance in bull and bear markets. Therefore, there is still disagreement on whether the level of diploma impact fund managers’ investment behavior. However, it is undeniable that there are some differences in trading behavior or strategies between fund managers with Ph.D. and masterșs degrees.

In summary, there is heterogeneity in the influence of social networks of fund managers with master’s and doctoral degrees on herding behavior, with the former largely consistent with the regression results under the whole sample, while the latter does not exhibit this characteristic, which to some extent supports hypothesis H4 of this investigate. Given that the proportion of fund managers with masterșs degrees is about 87%, this may lead to the former largely consistent with the regression results under the whole sample. Moreover, fund managers with a Ph.D. degree tend to have more complex alumni relationships, so they are more likely to obtain valid information through social networks and are less likely to herding behavior due to information asymmetry. Hence, social network centrality, constraint, and size may influence the herding behavior of fund managers with a Ph.D. degree to some extent but are not the main reasons for their herding behavior. Given this, social network centrality, network constraint, and network size may have some influence on the herding behavior of Ph.D.-educated fund managers, but they are not the main reason for their herding behavior, thus manifesting as the coefficients of the explanatory variables FreeClo, ValClo, Const, and Size are not significant in the regression results.

An investigation on the perspective of region heterogeneity

Regression model

The higher transmission efficiency of information in the same region also helps fund managers in the same region obtain important market information from others, which may reduce the information asymmetry of fund managers in the same area to a certain extent. Therefore, this section intends to discuss the influence of fund manager social networks on the herding behavior of SIFs from the perspective of regional differences.

Considering that the number of fund managers and the number of fund companies is higher in Shanghai, Beijing, and Shenzhen, and the social network structure and characteristics of fund managers may be more evident, this section conducts a regional heterogeneity investigation of these three cities. Among the 100 samples in Table 2, There are more fund managers in Shanghai with 53 samples, Beijing fund managers have a sample of 19, and Shenzhen fund managers have a sample of 22. Subsequently, we model the influence of social networks of fund managers in Shanghai, Beijing, and Shenzhen on the herding behavior of SIFs as follows.

$$\begin{array}{l}HB_{j,i,t} = \beta _{j,0} + \beta _{j,1}FreeClo_{j,i,t}\\ \qquad\qquad+\, \beta _{j,2}ValClo_{j,i,t} + \beta _{j,3}Const_{j,i,t} + \beta _{j,4}Size_{j,i,t}\\ \qquad\qquad+\, \beta _{j,5}Sex_{j,i,t} + \beta _{j,6}Edu_{j,i,t} + \beta _{j,7}Jnum_{j,i,t} + \beta _{j,8}Snum_{j,i,t} \\\qquad\qquad+\, \beta _{j,9}Soum_{j,i,t} + \beta _{j,10}Garr_{j,i,t} + \varepsilon _{j,i,t}\end{array}$$
(17)

where j takes s, b, and z in turn, representing the influence model of the social network of fund managers in Shanghai, Beijing, and Shenzhen herding behavior of SIFs, respectively. HBj,i,t is the explained variable, the degree of herding behavior. Other variables and constant terms have the same meaning as Eq. (1).

Empirical analysis

The regression results of three groups of samples in Shanghai, Beijing, and Shenzhen can be obtained, as shown in Table 8.

Table 8 Regression results from a region heterogeneity perspective.
Full size table

As can be seen from Table 8, the positive and negative signs of the regression coefficients of “Freeman closeness centrality” (FreeClo), “Valente-Foreman closeness centrality” (ValClo), network constraint (Const), and network size (Size) in the regression results of Shanghai and Beijing fund managers are consistent with the results in Table 4, and the variable ValClo in the Beijing fund manager group is significant at the 10% confidence level. In addition, all four explanatory variables in the Shenzhen fund manager group are insignificant, and the coefficients of the variables are completely different from the findings obtained in the previous paper. Therefore, under the perspective of regional differences, the “Valente-Foreman closeness centrality” indicator of Beijing fund managers shows a promotion effect on herding behavior with an absolute value of 21.960, indicating that fund managers have increased access to information due to their network position advantage, which increases the possibility of fund managers imitating others, which we defined as the “status advantage effect”.

However, the regression results for Shanghai and Shenzhen fund managers do not show a significant effect of social networks on herding behavior, which partially supports hypothesis H4 to some extent. From the behavioral finance perspective, there is variability in the external influences on fund managers in distinctive regions, coupled with the ease of access to information in the form of verbal communication among fund managers in the same area. Therefore, the trading strategies of fund managers in the same region may exhibit convergence, but there may be heterogeneity in the trading strategies of fund managers in various areas. Therefore, for fund managers in Shanghai, Beijing, and Shenzhen, social network centrality, network constraint, and network size have diverse effects on the herding behavior of SIFs.

Conclusion

This paper studies the influence of fund managers’ social networks on the herding behavior of SIFs in China. First, through information asymmetry, behavioral finance, and social relationship theories, it investigates the influence mechanism of the social network of SIF managers on herding behavior, proposing research hypotheses. Second, using the data of Chinese SIFs from 2012–2022, the CSAD model is applied to measure SIF’s herding behavior. Accordingly, a regression model of the influence of the managers’ social network on herding behavior is constructed and empirically studied by combining indicators such as network centrality, constraint, and size. Third, it explores the heterogeneity of this influence in four dimensions, degree of herding behavior, gender, diploma, and region, and it can be found that:

(i) The social network of fund managers has a certain degree of influence on the herding behavior of SIFs, which mainly shows that the larger the social network of fund managers, the lower the degree of herding behavior of fund managers, but the influence of the network constraint on the degree of herding behavior is not significant. (ii) When the degree of herding behavior is high, the network size of fund managers’ social networks has a significant effect on the heterogeneity of herding behavior, but not significant when the degree of herding behavior is low. (iii) The relationship between the social network of male fund managers and the degree of herding behavior is more significant than that of female fund managers, as shown by the significant network size of male fund managers, while the regression results of the four indicators of the social network of female fund managers (“Freeman closeness centrality”, “Valente-Foreman closeness centrality”, network constraint, and network size) and the degree of herding behavior are not significant, thus indicating that male fund managers may be more prone to herding behavior. (iv) The influence of the social network of fund managers with master’s degrees on the herding behavior of SIFs is more significant, but it is not significant for fund managers with doctoral degrees due to their more complex alumni relationships, etc. (v) Regional differences have a heterogeneous effect on fund managers’ trading strategies. Beijing fund managers’ herding behavior is influenced by the “status advantage effect” of social network centrality, which shows a higher facilitation effect on herding behavior.

Given the limitation of space and the difficulty of data collection, this paper only constructs a measurement model of the social network of Chinese SIF managers in terms of the dimensions of network centrality, network constraint, and network size, without considering the influence of mobile social networking platforms, company business contacts, and relatives or friends’ relationships, etc. In addition, this paper does not combine behavioral finance and principal-agent theory to construct measures of intentional or spurious herding behavior of SIFs from the perspective of trading motives and does not consider the impact of information transmission efficiency in the social network of fund managers on the herding behavior of SIFs. Hence, this will be the future research focus.