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

Securities investment funds (SIFs) play an essential role in Chinese capital markets and exhibit pronounced herding, driven by factors such as rankings and personal incentives. Evidence from holdings shows many funds synchronously adjusting positions in the same stocks within a quarter. As fund managers are embedded in dense social networks (alumni, colleagues, peers), these networks facilitate information exchange that may lead to convergence in asset allocation decisions. This paper argues that similarities in valid information acquired through managers' social networks drive convergence in behaviors, manifesting as herding. Key research questions include: how to measure fund managers' social networks; and what characteristics and mechanisms link managers' social networks to SIFs' herding. The study builds on information asymmetry theory, behavioral finance, and social relations theory; proposes hypotheses regarding centrality, size, constraint, and heterogeneity; measures herding with the CSAD model; and estimates regression models linking social network metrics to herding, with further heterogeneity analyses by herding degree, gender, diploma, and region. The contributions include clarifying theoretical mechanisms, constructing social network measures (centrality, size, constraint) for SIF managers, applying CSAD-based herding measures, and documenting heterogeneous effects across manager characteristics.

The literature indicates social networks shape investors’ trading behavior: crosstown managers share information and trade similarly; alumni ties influence stock selection; close network ties correlate with more similar trades; and proximity fosters information sharing that yields trading similarity. Regarding fund performance, managers from prestigious universities possess richer social capital; network centrality confers information advantages that can improve performance; and fund network structure influences flows and performance, with direct/close alumni ties most impactful. Herding formation has been explained by reputation concerns (managers align with peers to mitigate reputational risk), information cascades (public information embedded in prices and observation of informed peers), and incentive structures (principal–agent driven compensation). Chinese SIFs exhibit pronounced herding, often asymmetric between buys and sells, and vary with stock characteristics. Two main herding measurement approaches exist: return-dispersion based (CSSD, CSAD) and investor-level (e.g., LSV). CSAD better captures nonlinearities and weaker herding. Prior studies document significant herding in U.S. and Chinese markets and across fund types. Building on these, the paper develops hypotheses: H1 centrality significantly affects herding; H2 larger network size inversely relates to herding; H3 network constraint positively contributes to herding; H4 heterogeneity across herding degree, gender, diploma, and region affects these relationships.

Data and sample: Daily return data for all Chinese open-end SIFs from 2012-07-06 to 2022-07-07 were collected (Wind and Eastmoney). Money, bond, and some mixed funds with low volatility were excluded; equity and equity-biased hybrid funds were retained (6,078 fund entries covering 30+ industries). Market return is proxied by the CSI 300 (Shanghai and Shenzhen 300). Fund managers’ CVs were used to construct social networks through two relationship layers: alumni networks (schools attended) and colleague networks (workplaces). Due to incomplete education disclosures for some managers, colleague networks were combined with alumni networks to better capture overall network structure. The social network adjacency matrices were built in Excel and analyzed in UCINET to compute network metrics. Herding measurement (explained variable): Herding is measured via the CSAD framework. CSAD_t = (1/N) Σ_i |R_{i,t} – R_{m,t}| measures cross-sectional absolute deviation of fund returns from market return. Under CAPM, CSAD increases linearly with |R_m|; a significant nonlinear (concave) relation indicates herding. The empirical specification: CSAD_{n,t} = α + γ1 R_{m,t} + γ2 R_{m,t}^2 + ε. A significantly negative γ2 indicates herding, and its magnitude reflects herding strength. For each fund manager n, daily CSAD across funds managed by that manager was computed over the sample period; the quadratic coefficient γ2 (denoted HB) serves as the manager-level herding measure (more negative = stronger herding). Newey–West standard errors were employed. Explanatory variables (social network): Centrality and structure measures computed in UCINET include: FreeClo (Freeman closeness centrality), ValClo (Valente–Foreman closeness centrality), Const (network constraint per Burt capturing structural holes/constraint), and Size (network size, i.e., count/extent of connected alters across alumni/colleague layers). Centrality captures position and access; Size captures breadth of connections; Const captures degree of redundancy versus structural holes. Control variables: Sex (gender; 1=male, 0=female), Dip (diploma; coded to reflect undergraduate/master/Ph.D.), Jnum (number of schools attended), Snum (number of funds managed). In extended heterogeneity regressions, additional controls include Cnum (number of inaugural firms), Jyear (years of experience), Soum (fund size under management), and Garr (geometric mean annualized return). Baseline regression: HB_i = β0 + β1 FreeClo_i + β2 ValClo_i + β3 Const_i + β4 Size_i + β5 Sex_i + β6 Dip_i + β7 Jnum_i + β8 Snum_i + ε_i. Sampling for estimation: Due to computational intensity, 250 managers were randomly selected; herding measured via CSAD for each. Of these, 100 managers exhibited significantly negative γ2 (herding) and were retained for the main regression analyses. Descriptive statistics (N=100): HB mean -72.047 (sd 65.537), FreeClo mean 1.357, ValClo mean 15.253, Const mean 0.646, Size mean 1.355; Sex mean 0.85; Dip mean 19.21; Jnum mean 0.38; Snum mean 5.02. Heterogeneity analyses: Separate regressions were run by (a) herding degree (split by HB into high vs. low herding groups; each N=50) with additional controls Soum and Garr; (b) gender (male N=85; female N=15), replacing Sex with Cnum; (c) diploma (master’s N=87; Ph.D. N=12), replacing Dip with experience Jyear; and (d) region (Shanghai N=53; Beijing N=19; Shenzhen N=22).

- Baseline herding evidence: Among 100 focal managers, CSAD quadratic terms γ2 were significantly negative, indicating prevalent herding; extreme cases included γ2 = -483.277 (Feng Jiang) and -312.029 (Yixiang Fu). Most managers exhibited medium-level herding (HB in approximately [-100, -20]). - Baseline regression (N=100): - Network size (Size): coefficient = 16.970, significant at 5% (SE 7.059). Given HB is negative when herding is stronger, a positive coefficient implies larger networks are associated with less negative HB (i.e., weaker herding). This supports H2 (larger network size inversely relates to herding). - Centrality (FreeClo, ValClo): coefficients 108.798 and -3.916, respectively; both not significant. H1 not supported. - Network constraint (Const): coefficient -16.890; not significant. H3 not supported. - Controls generally not significant in the baseline. - Heterogeneity by degree of herding (N=50 per group): - High herding group: Size = 27.713**, significant (SE 12.742), indicating network size has stronger association when herding is high; FreeClo, ValClo, Const not significant. - Low herding group: none of the social network variables significant. - Gender heterogeneity (Male N=85; Female N=15): - Male managers: Size = 18.620**, significant (SE 7.692); FreeClo, ValClo, Const not significant. - Female managers: None of the four social network variables significant; Soum significant at 5% with negative sign. - Diploma heterogeneity (Master’s N=87; Ph.D. N=12): - Master’s group: Size = 17.261**, significant (SE 7.207); centrality and constraint not significant. Several controls significant (e.g., Jyear negative at 1%; Snum positive at 1%; Soum negative at 1%). - Ph.D. group: none of the social network variables significant at 10%. - Regional heterogeneity (Shanghai N=53; Beijing N=19; Shenzhen N=22): - Beijing: ValClo = -21.960*, significant at 10%, suggesting centrality (ValClo) increases herding (more negative HB) — a “status advantage effect.” Size positive but not significant. - Shanghai and Shenzhen: social network variables not significant. Overall: Larger network size is consistently associated with weaker herding in the full sample and in several subgroups (especially where herding is high and among male or master’s-level managers). Centrality and constraint generally are not significant, with the exception of Beijing (ValClo).

The findings indicate that while herding among Chinese SIF managers is prevalent (negative γ2 in CSAD regressions), the breadth of a manager’s social network (Size) is associated with attenuated herding. This aligns with information asymmetry and social relations theories: broader networks expand access to diverse, potentially superior information, improving independent decision-making and reducing reliance on peer imitation. The lack of robust effects for centrality (FreeClo, ValClo) and network constraint (Const) in the aggregate suggests that position and structural hole utilization, as measured here, do not systematically drive herding across all managers, possibly because high-centrality managers’ superior information buffers them from conformity pressures. Heterogeneity analyses nuance these results. When herding pressure is high, network size’s mitigating association is stronger, consistent with broader networks enabling better information screening under stress. Gender differences show that network size relates to herding primarily for male managers, consistent with documented higher risk-taking and reputational pressures among men that might otherwise promote herding; broader networks may counteract such tendencies. Educational heterogeneity suggests that master’s-educated managers’ herding is more sensitive to network size, whereas Ph.D.-educated managers may rely on alternative information-processing advantages not captured by the network variables. Regionally, Beijing’s significant ValClo effect suggests that certain network configurations can increase herding through status and visibility advantages in specific ecosystems. Together, the results support the view that social networks matter most through breadth (Size), with their effects conditioned by manager and regional characteristics.

This study develops and applies a social-network-based framework to examine SIF fund managers’ herding in China. Using CSAD-derived herding measures and UCINET-based network metrics, the paper shows that: (i) larger social networks are linked to lower degrees of herding; (ii) the mitigating association of network size is strongest when herding is high; (iii) effects are more pronounced for male managers, with no significant social-network effects for female managers; (iv) master’s-level managers exhibit significant network-size effects, whereas Ph.D.-level managers do not; and (v) regional heterogeneity exists, with Beijing showing a centrality-driven “status advantage effect” that increases herding. These contributions extend the literature by integrating network metrics with fund-level herding measures and documenting heterogeneous impacts across manager attributes and regions. Future research could incorporate additional relationship channels (e.g., mobile social platforms, inter-company business ties, family/relatives), distinguish intentional versus spurious herding using behavioral and agency-theoretic constructs, and directly measure information transmission efficiency within networks.

The study’s social network measurement is limited to alumni and colleague ties summarized by centrality, size, and constraint, excluding other channels (e.g., mobile social networking platforms, inter-firm business contacts, relatives/friends). Educational information in CVs may be incomplete, potentially biasing network construction. The CSAD method can understate herding in certain conditions and does not differentiate intentional versus spurious herding. Information transmission efficiency within networks is not modeled. The empirical sample for regression focuses on 100 managers with significant herding, and subgroup sample sizes (especially female and Ph.D. groups, and some regional groups) are small, limiting statistical power and generalizability.