This study investigates the asymmetric relationship between investor sentiment and mutual fund market resilience in China’s accommodative monetary policy environment. Following the September 2024 reserve requirement ratio and interest rate cuts by the Chinese central bank, we employ FinBERT, a financial domain large language model, to construct a textual sentiment index and use normalized Shannon entropy to capture the dynamic stability of the fund market. The empirical results show that investor sentiment has exerted a significant positive effect on fund market resilience. Specifically, a 1-unit increase in investor sentiment is associated with a 0.034-unit increase in fund market resilience. However, the relationship is asymmetric. Excessively high sentiment weakens resilience, whereas low sentiment strengthens it, revealing a nonlinear relationship between sentiment and resilience. Heterogeneity analysis reveals that funds managed by more highly educated managers are less influenced by sentiment, whereas balanced-style funds and those domiciled in financially developed regions are more sensitive to sentiment fluctuations. The endogeneity test employing the proportion of heavily weighted stocks as the instrumental variable further confirms the asymmetric effect of investor sentiment on fund market resilience. These findings highlight the dual role of investor sentiment as both a stabilizing and destabilizing force and offer insights for designing sentiment-aware regulatory and resilience frameworks in knowledge-driven financial systems.
The COVID-19 pandemic has fundamentally reshaped China’s financial landscape, triggering unprecedented volatility in capital markets and profound shifts in investor behavior. Following the initial shock of lockdowns, China’s equity market experienced dramatic fluctuations: after hitting multiyear lows in early 2022, the Shanghai Stock Exchange Composite Index (SSEC) rebounded by more than 14% during the National Day holiday in October 2024, achieving its best performance since 2008. However, prior to that, much of the economic slowdown was characterized by a decline in market sentiment. Households generally exhibited historically low levels of confidence, manifesting in strong reluctance to consume and the urge to hoard savings. Moreover, their investment behavior toward property was equally conservative, with the majority of potential homebuyers leaving capital idle in deposits instead of flowing into markets. Businesses have also contracted: private fixed-asset investment in manufacturing and services has either stagnated or decreased, and less than 30% of small and medium-sized enterprises (SMEs) intended to expand their investment. These patterns of behavior are indicative of a general pessimism that goes beyond conventional economic principles.
In response to deteriorating market confidence, Chinese authorities implemented comprehensive stimulus packages on September 24, 2024, encompassing monetary easing, fiscal expansion, and capital market reforms (including lower reserve requirements, targeted lending support, and long-term funding mechanisms for institutional investors). This move released a favorable signal through executive decision-making, further sparking an instantaneous spike in trading activity, with several brokerage systems experiencing temporary system disruptions due to unprecedented transaction loads. While such episodes illustrate the amplified impact of sentiment shifts on market microstructure, they also expose fundamental vulnerabilities: sentiment-driven capital flows can generate systemic pressures that cascade from individual securities to mutual funds and broader financial ecosystems. This phenomenon is particularly concerning in China’s retail-dominated market, raising critical questions about the resilience of financial markets under sentiment-driven conditions.
Despite extensive research on investor sentiment (Baker et al., 2007; Bashir et al., 2024; Phan et al., 2023), stock returns (Chen & Haga, 2021; Dias et al., 2022; Jing et al., 2021; Seok et al., 2024), and volatility (Gong et al., 2022; Gong et al., 2024; Hsu et al., 2022), critical gaps remain in understanding its asymmetric effects on fund market resilience—defined as the capacity of mutual funds to maintain stability and recover from shocks. The existing literature has predominantly focused on the following areas but has its own limitations:
1) Most studies have focused mainly on equity stock markets. However, mutual fund markets also present a distinct and equally important context for understanding sentiment effects. Fund markets, owing to their interconnected nature, where redemption pressures and portfolio rebalancing can amplify initial shocks, are particularly susceptible to sentiment contagion (Arfaoui et al., 2024; Jeffers et al., 2024).
2) The existing literature mainly explores changes in investor sentiment under external shocks (Anand et al., 2013; Cerqueti et al., 2022; Tang et al., 2022). Few studies have examined the impact of investor sentiment on fund resilience, especially given the potential for asymmetric effects where negative sentiment may have disproportionately stronger effects than positive sentiment does.
3) Several studies have employed market-based proxies such as stock returns (Nguyen et al., 2025), trading volume (Chung et al., 2024), turnover rate (Zhang et al., 2025), and the VIX, or survey-based proxies such as the AAII, and SICI to measure investor sentiment. However, the selection of proxy variables fails to capture all relevant dimensions of investor sentiment, and nuanced information is inevitably discarded through dimensionality reduction.
On the basis of these discussions, three fundamental questions remain unresolved: First, how can the investor sentiment and the fund market resilience index be accurately constructed so that they can more comprehensively reflect the investor sentiment? Second, do positive and negative sentiment shocks affect fund market resilience asymmetrically or symmetrically? Third, what are the transmission channels through which investor sentiment propagates to fund market resilience and shapes its dynamics?
To address these research gaps, we adopt three interrelated methodological approaches. First, we construct the sentiment index by applying FinBERT—a finance-specific language model pretrained on 4.9 billion words of financial text—to investor comments from a Chinese online fund forum. To measure fund market resilience, we construct a complex dynamic network combined with Markov chains to derive steady-state transition matrices and ultimately employ the normalized Shannon entropy as a proxy for resilience. Second, we estimate a time-fixed effects econometric model in which investor sentiment serves as the key explanatory variable and fund resilience as the dependent variable. This framework allows us to capture how mutual funds respond to external sentiment fluctuations and to quantify the magnitude of the sentiment–resilience linkage. Third, to identify potential transmission mechanisms, we conduct heterogeneity analysis across three fund characteristics—manager educational background, investment style, and registration location. This multidimensional approach reveals how fund-specific attributes moderate the sentiment-resilience dynamics. Based on this analytical framework, our study offers three primary contributions:
1) From an optimization perspective, we construct a high-frequency, domain-specific sentiment index by applying the FinBERT model, which leverages the architecture of large language models. Unlike traditional sentiment proxies, FinBERT captures contextual semantics and financial terminology more effectively through iterative parameter optimization, thereby mitigating the information loss commonly induced by dimensionality reduction.
2) From a scope perspective, while prior research has largely focused on sentiment effects in equity markets, we extend the analysis to mutual funds and systematically examine the relationship between investor sentiment and fund market resilience. Our findings reveal an asymmetric relationship: sentiment weakens resilience during high-sentiment regimes but strengthens it during low-sentiment periods.
3) From a mechanism perspective, we comprehensively investigate the transmission path of the sentiment–resilience relationship through multiple fund-level attributes. The heterogeneity tests further indicate that different fund types exhibit varying sensitivities to sentiment shocks. As China continues to deepen its integration into global capital markets, insights from this study contribute to the broader discussion on sentiment contagion and financial stability in emerging economies.
The remainder of the study is structured as follows. Section 2 reviews the related literature on investor sentiment and fund market resilience. Section 3 describes the methodological framework. Section 4 outlines the data sources and variable construction. Section 5 presents the empirical findings and the heterogeneity analysis. Section 6 concludes and discusses directions for future research.
Literature reviewMeasurement of investor sentimentInvestor sentiment represents the overall psychological disposition and expectations of investors toward the financial market. It reflects the degree of optimism or pessimism of market participants in their investment decisions. It can mainly be divided into two types: a comprehensive index and a text mining index.
The composite index approach is based on a mature empirical method, which can effectively integrate multi-dimensional data. Baker and Wurgler (2006) integrated multiple proxy variables reflecting market sentiment, such as adding the closed-end fund discount rate and IPO number into the comprehensive sentiment measurement index through principal component analysis. Glushkov (2006) constructed the sentiment beta index to measure investor sentiment by controlling risk factors related to market, scale, book-to-market ratio, and liquidity. Antoniewicz et al. (2014) concentrated on a questionnaire survey targeted at professional and individual investors to elicit information including investors’ expectations of future market developments and risk preferences, thereby constructing a comprehensive sentiment index. Huang et al. (2015) used partial least squares regression to improve the traditional principal component analysis method, thereby extracting emotional factors most relevant to future returns and eliminating macroeconomic noise to construct a new sentiment index. Ung et al. (2024) improved the ability of empirical models to capture underlying investor sentiment by observing how the various components of emotional indicators change over time.
The text-mining index approach, based on large-scale text data, is constructed by using a text analysis model to distinguish positive, negative, and neutral sentiment tendencies. Antweiler and Frank (2004) used the Naive Bayes algorithm to classify Yahoo Finance posts into three sentiment types: short, long, and hedge. Bollen et al. (2011) crawled Twitter texts and used OpinionFinder to classify texts into positive and negative sentiment, and used GPOMS to extract six different sentiment dimensions from the text content, thereby generating seven types of public sentiment by time series. Tsukioka et al. (2018) extracted investor sentiment toward Japanese IPO companies from local forums, using text mining and support vector machine classification to categorize the posts as bullish, neutral, or bearish. Yang et al. (2020) used FinBERT to classify the sentiment of stock market news and analyzed the impact of investor sentiment fluctuations on stock market performance. Ruan et al. (2025) combine financial sentiment extraction based on FinBERT with technical and statistical indicators to predict short-term stock price movements. The text mining index currently has a wide range of applications in financial reviews, media reports, and corporate financial reports because it employs sophisticated models such as BERT and FinBERT to enhance the accuracy of sentiment classification. Additionally, it can facilitate continuous and high-frequency sentiment analysis based on massive amounts of text data from multiple sources. Therefore, our study uses the text mining analysis to construct the investor sentiment index.
Measurement of fund market resilienceAfter the 2008 international financial crisis, financial resilience frequently appeared in domestic and foreign policy documents and research reports. As part of the resilience of the financial market, fund market resilience involves the ability of the fund market to effectively respond to, recover from, and maintain normal operation when encountering external or internal shocks (Li & Zhu, 2019). Regarding the connotation of market resilience, the mainstream academic community is divided into three views:
The first view focuses on understanding fund market resilience from the dynamic perspective of fund market liquidity. Foucault et al. (2013) proposed that market resilience can be measured by the speed at which the price difference returns to its original level before the next transaction after the market experiences a liquidity shock. Kim et al. (2019) further described resilience as the speed at which prices recover to their fundamental value from liquidity shocks. The faster the price returns, the faster the market recovers from short-term price shocks, thus showing stronger market resilience.
The second view holds that market resilience is a manifestation of multidimensional capabilities. Tang et al. (2022) believed that market resilience not only needs to withstand large external shocks but also needs to be able to recover quickly and adapt to market changes, which is reflected in the improvement of sufficient liquidity, reasonable market pricing, and risk diversification functions. These capabilities reflect the comprehensive response capabilities of the fund market in crises. Grosu et al. (2025) use slope functions for time series analysis, categorizing market resilience into phases, including the period before the event, during the event, and the recovery and adjustment phases after the event.
The third view is that market resilience is essentially the stability of the market dynamic network. Stocks tend to synchronize their price evolution, resulting in a high degree of dynamic network correlation between stock prices. Kauê Dal’Maso et al. (2012) used entropy-related measurements to quantify the stability of evolving financial market organizations. The study showed that the network topology changed dramatically during financial instability, and the stability of the network decreased. Leal and Napoletano (2019) demonstrated that the constant dynamic interaction between related entities is the underlying cause of systemic financial risks and market crashes. Yang et al. (2020) used entropy methods to quantify the stability of stock networks following market crashes and investigated the impact of government intervention on network stability as measured by entropy. Siudak (2025) employs a complex multi-layered network combination method to examine the statistical dependencies between each layer of the network and commonly used stock return networks in the financial market.
Although the definition of fund market resilience varies across academic contexts, most scholars regard financial market resilience as the overall capacity of a market to withstand and recover from crises. Building on this perspective, the present study adopts the third approach, constructing an entropy-based index of fund market resilience within a complex dynamic network framework. We adopt Shannon entropy as the proxy variable for financial resilience. Its core advantage lies in its sensitivity to capture the uniformity of Markov transition probability Pij distributions, which represents the path diversity of risk shock transmission across the market. High entropy suggests that price fluctuations are dissipated through a diffuse and redundant network structure, thereby enhancing the ability of the fund market to absorb and buffer shocks. By contrast, low entropy indicates path dependency and structural rigidity, a situation where disturbances spread predictably and induce cascading failures. By quantifying the informational diversity of risk transmission, our entropy-based metric provides a complementary lens to traditional risk measures, shedding new light on the systemic robustness of financial markets under different sentiments.
Theoretical framework: behavioral finance perspectivesBased on the aforementioned measurement of investor sentiment and the construction of fund market resilience indicators, this study proposes a theoretical mechanism based on behavioral finance:
(1) Overconfidence Theory: Daniel et al. (1997) proposed that investors tend to overestimate the accuracy of private information signals and underestimate their own prediction errors, thus overreacting to private information and underreacting to public information. Recent evidence has extended this framework to environments related to emotions. Nie et al. (2024) proposed that overconfident CEOs adapt their merger behavior based on the economic environment, with investor sentiment playing a dominant role in the decision-making process. Yeung et al. (2025) found that investors with lower portfolio values are more susceptible to the influence of past experiences, and negative experiences exacerbate behavioral biases, with this effect being asymmetrical. Specifically, when sentiment is high, investors tend to overestimate private information and underestimate prediction errors. This leads to increased trading frequency and price volatility, while during periods of low sentiment, public information continues to arrive, accuracy dominates, prices revert to fundamentals, investors more cautiously assess risks, reducing irrational trading, and thus enhancing market stability. The network becomes vulnerable due to concentrated risk exposure and liquidity mismatch caused by overconfident trading.
(2) Herding Effect: Banerjee (1992) proposed that people imitate the behavior of others rather than using the information they possess. Contrary to the common belief that optimism breeds herding behavior, recent evidence suggests a subtle relationship between the two. Yoon et al. (2022) found that herding behavior among retail, institutional, and foreign investors can negatively impact bullish sentiment—strong herding can cause crowding, thereby undermining confidence. Sheikh et al. (2025) showed that as investor optimism increases, the herding behavior in the Chinese market decreases, possibly because overconfidence suppresses imitative behavior (investors believe their private signals are superior). When market sentiment is high, the fund focuses on popular industries (such as artificial intelligence and clean energy), thereby reducing network diversity. When market sentiment is low, investors disagree on valuations (some believe that there are investment opportunities, while others believe that the market is falling), leading to diversified trading and thus increasing network diversity (high entropy).
(3) Prospect Theory: Kahneman and Tversky (1979) proposed that traders are more sensitive to losses than to gains, and this asymmetry is reflected in people’s tendency to avoid losses, which profoundly shapes redemption behavior in fund markets. When sentiment is high, excessively high expectations would amplify perceived losses (people believe the net worth is lower than expected). Cevik et al. (2022) confirmed that negative sentiment increases volatility and exacerbates market instability. Initial redemptions lead to a decrease in net asset value, which in turn triggers perceived losses, ultimately leading to more redemptions. In contrast, low sentiment regimes do not exhibit this amplification effect because the losses are anticipated. Sophisticated investors would delay redemptions to avoid realizing losses and thus stabilize cash flow. The research by Cui et al. (2025) suggests that bullish sentiment among retail investors foreshadows lower returns, which is consistent with the wave of redemptions overreaction. This mechanism thus explains our core empirical finding: high investor sentiment weakens fund market resilience, whereas low sentiment enhances it.
Based on this, we expect resilience to decrease during periods of rising sentiment and increase during periods of falling sentiment, which will be tested later using a text-mining-based sentiment index and an entropy-based resilience indicator.
The impact of investor sentiment on fund market resilienceInvestor sentiment and fund market resilience both play an important role in the financial field. There is a close and complex relationship between them, which together shape the dynamic pattern of the capital market. Racca et al. (2016) measured the knowledge dynamics and user behavior before and during the recent financial crisis, studied the impact of market uncertainty shocks on the group investor sentiment from online forums, and explored the impact of group sentiment on the resilience of online communities. The study showed that the crisis had a gradual impact on community groups, but still exhibited strong market resilience against external shocks. Ding et al. (2020) aimed to study the impact of market sentiment on stock market resilience during COVID-19. The study showed that the stock prices of various industries varied depending on the level of digital transformation of the industry. The industries with the most successful digital transformation showed strong financial market resilience to the negative market sentiment brought about by the epidemic. Yadav et al. (2023) used India’s consumer sentiment index to explore its impact on financial resilience and its potential association with macroeconomic variables. The results showed that consumers’ positive expectations of future credit availability would lead to an increase in credit limits, thereby bringing stronger financial resilience. Pernici et al. (2024) showed that after the impact of the COVID-19 pandemic in 2020, people became more optimistic about the capital market and paid more attention to the results and remedial measures related to the concept of financial resilience.
Most existing studies examine the relationship between investor sentiment and financial market variables from a single perspective, without providing an integrated view that links sentiment to fund market resilience. The literature remains sparse on how investor sentiment shapes the resilience of the fund market.
MethodologyLarge language model for sentiment analysisIn recent years, the rise of artificial intelligence for finance (AI4Finance) has profoundly transformed how financial markets process and interpret textual information. Large language models (LLMs) have become effective tools for extracting nuanced semantic patterns from large-scale financial text corpora, especially those with domain-specific architectures such as FinGPT and FinBERT. The original BERT model, which was pretrained on general-purpose and unsupervised datasets like Wikipedia, performs suboptimally when applied to highly specialized financial contexts. To overcome this limitation, Araci (2019) introduced FinBERT, a model built on the transformer encoder architecture of BERT. FinBERT adopts a two-stage learning process—pretraining and fine-tuning—on both general and domain-specific financial corpora. It enables FinBERT to capture subtle semantic meanings more precisely and enhances the effectiveness of deep learning methods in financial text analysis.
In this study, we employ the FinBERT model developed by Yang et al. (2020), which is pre-trained on a large-scale financial communication corpus containing approximately 4.9 billion words from company reports, earnings call transcripts, and analyst commentaries covering the period from 1994 to 2019. This model serves as the primary tool for analyzing investor comments in the Chinese fund market, collected from the Eastmoney online forum. As an open-source financial LLM, FinBERT provides a lightweight and cost-efficient framework that can adapt rapidly to changing market conditions, thereby addressing the persistent challenge of measuring investor sentiment in a comprehensive and scalable manner. The workflow of FinBERT is illustrated in Fig. 1.
The workflow of FinBERT consists of two main components. The first stage merges and preprocesses diverse textual sources—including company filings, conference call transcripts, and analyst reports—by removing HTML tags and tabular data. SentencePiece’s Unigram algorithm is then applied to estimate word probabilities and construct the Finvocab dictionary. The second stage involves encoding through a self-attention mechanism and a feed-forward neural network (FFNN). The self-attention layer computes attention weights between each token and all others in the sequence, while the FFNN refines the resulting representations using the Gaussian Error Linear Unit (GELU) activation function. Optimal parameters are obtained through iterative minimization of the loss function, and the resulting gradients are backpropagated to perform sentiment classification tasks.
Measuring investor sentimentThe core explanatory variable in this study is investor sentiment. Using web-scraping techniques, we collected more than 1.5 million user comments on open-ended partial equity and hybrid funds from the Eastmoney fund forum over the sample period. These text data were processed through the FinBERT model, which was pre-trained on financial corpora and subsequently fine-tuned using K-fold cross-validation and reverse transfer parameter adjustments to enhance classification accuracy. The model outputs sentiment categories at the comment level, classified into three groups: positive/high, negative/low, and neutral. Following Antweiler and Frank (2004), the daily sentiment index for fund i on day t is computed as:
where the number of posts with positive sentiment for fund i on day t is recorded as Npos,i,t, and the number of posts with negative sentiment is recorded as Nneg,i,t. When the index Sentiment is greater than 0, it is marked as positive/high sentiment. When the index is less than 0, it is marked as negative/low sentiment. When the index is equal to 0, it is marked as a neutral sentiment.Measuring fund market resilience
The core dependent variable of this study is the resilience of the fund market. Following the methodology of Kauê Dal’Maso et al. (2012) and Cerqueti et al. (2022), we employ a dynamic network approach that combines sliding window analysis with entropy-based measurements to quantify the structural stability of fund networks.
Network Construction: We compute dynamic correlations between all fund pairs within 30-day sliding windows that advance one day at a time. These correlations are transformed into distance metrics capturing co-movement patterns. Specifically, for any two funds i and j, the correlation coefficient is calculated as:
where Yi represents the unit net value of the i-th fund, Yj represents the unit net value of the j-th fund, and ρij represents the correlation coefficient between the two funds.
This correlation can then be converted to a distance measure, which can reflect the similarity of the evolution of the unit net value of the two funds:
where d(i,j) represents the dynamic distance between the two funds. This distance metric satisfies the mathematical axioms of a proper metric space (non-negativity, symmetry, and triangle inequality). The distances between N funds constitute a dynamic N*N matrix, from which we construct a weighted adjacency matrix:
Markov Chain Analysis: To capture the price transmission mechanism across funds, we model the network dynamics using a Markov chain framework. Through eigenvalue decomposition of the weight matrix W:
where λ denotes the dominant eigenvalue of the corresponding weight matrix W, representing the network’s overall correlation strength. v is the corresponding right eigenvector reflecting each fund’s network centrality. We then construct the transition probability matrix P of the Markov chain and steady-state distribution as the following expression:where wij represents the weight between i fund and j fund. vj represents the right eigenvector of the j-th fund. λ is the maximum eigenvalue of the corresponding weight matrix W. pij represents the probability of the i-th fund transferring to the j-th fund, capturing directional price contagion patterns. This probability satisfies the Markov chain normalization condition (∑jpij=1) by construction. ui represents the left eigenvector of the weight matrix W corresponding to the i-th fund. πi represents the value of the i-th fund reaching the steady-state distribution (at this time, there is no unit net value conversion between the i-th fund and the j-th fund).
Entropy-Based Resilience: We quantify fund market resilience using Shannon entropy—a measure from information theory that captures the diversity of a fund’s connections (Shannon, 1948). Higher entropy indicates more diversified linkages and thus greater resilience to shocks. The resilience of fund i is defined as:
where resiliencei represents the local entropy contribution of the i-th fund. This measure weights each transition probability pij by the fund’s steady-state importance πi, capturing both the diversification of price transmission paths and the fund’s centrality in the network. Higher entropy indicates greater structural diversity—when a fund maintains connections to multiple other funds with relatively balanced transition probabilities, it exhibits higher resilience to idiosyncratic shocks.
To facilitate cross-sectional and temporal comparisons, we normalize the resilience index using z-scores to ensure the robustness of the corresponding resilience values of individual funds:
where Resiliencei,t represents the normalized local entropy contribution of the i-th fund on the day t. resiliencei,t represents the local entropy contribution of the i-th fund on the day t, u¯i and σi denote the time-series mean and standard deviation of the entropy of the i-th fund over the sample period. This standardization removes fund-specific scale effects while preserving within-fund temporal variation.
To avoid conceptual ambiguity, it is important to distinguish clearly between fund-level and market-level resilience in our study. Our empirical analysis is conducted at the fund level, where Resiliencei,t measures the resilience of the i-th fund. Specifically, it captures the fund’s ability to absorb shocks while maintaining its relative position within the return-based fund network. Market-level resilience is not directly estimated in the regression framework, but is instead defined as an aggregate indicator. Resiliencet‾=1N∑i=1nResiliencei,t represents aggregate market resilience—the overall stability of the fund market under systemic stress. This aggregation is used solely for robustness tests, rather than for causal inference. From a network perspective, entropy is a system-level property. Higher entropy reflects a more diversified configuration of network connections and greater feasible state-transition paths, which enhances the market’s capacity to absorb shocks. Conversely, lower entropy indicates a more concentrated network structure, increasing vulnerability to systemic contagion. Importantly, this interpretation applies to the structure of the fund network as a whole, rather than implying that any individual fund is inherently more robust in isolation. For readers interested in detailed technical derivations, including the mathematical properties of the distance metric, convergence properties of the Markov chain, and the information-theoretic foundations of Shannon entropy, we provide comprehensive documentation in Appendix B.
Entropy as the proxy of resilience: validation with traditional metricsTo address concerns that high entropy may reflect chaotic disorder rather than resilience, we distinguish between two interpretations. Under the disorder interpretation, high entropy arises from random, uncorrelated shocks that create structural noise and market instability. Under the risk dispersion interpretation, high entropy reflects uniform distribution of returns across network nodes, reducing concentration and contagion risk—consistent with network theory where maximum entropy minimizes cascading failure probability.
We examine the relationship between entropy and four traditional metrics: market-level risk is measured by the theoretical daily volatility calculated by fund managers. Fig. 2 Panel A shows a significant negative correlation between entropy and market volatility: volatility is significantly lower during periods of high entropy. According to the fitted line, a 1-standard-deviation increase in entropy corresponds to a 0.15-unit decrease in volatility (approximately 30% of the mean). This pattern confirms that entropy captures stability rather than disorder. We also adopt net asset value (NAV) change ratio as the proxy of redemption pressure, where negative changes indicate selling pressure. Fig. 2 Panel B reveals the negative correlation: funds in high-entropy states face lower redemption pressure. This is consistent with the predictions of risk diversification theory, which suggests that higher entropy can reduce panic outflows.
Panels C and D examine liquidity proxies including fund size and portfolio turnover, which show no significant relationship with entropy. We attribute this to institutional features of China’s mutual fund market: open-end funds trade exclusively over-the-counter with fund companies as counterparties, lacking secondary markets. Consequently, standard liquidity measures such as bid-ask spreads and trading volumes are inapplicable, and turnover reflects management strategy rather than redemption liquidity.
Two main observations are obtained from the validation exercises: First, the negative correlation between entropy and conventional risk indicators (volatility and redemption pressure) confirms that entropy captures the stability dimension identified by these traditional indicators. Second, the zero correlation results regarding liquidity reflect more the characteristics of market institutions than a problem with the effectiveness of entropy. We argue that if entropy merely represents disorder, it should exhibit a positive correlation with all risk indicators. However, entropy is negatively correlated with volatility and redemption pressure but not with market liquidity measures. These findings imply that entropy is not only theoretically grounded but also empirically robust, providing a conceptually distinct and empirically validated measure of fund market resilience.
Data sources and variablesWe construct the econometric equation with investor sentiment as the core independent variable and fund resilience as the core dependent variable serving as our main econometric model. The control variables are the fund establishment years, fund custody fee rate, fund management fee rate, fund cumulative net value, annual return, maximum drawdown during the tenure, maximum return during the tenure, and price fluctuation. The following regression equation is constructed:
where Resiliencei,t represents the fund resilience corresponding to the day t of the i-th fund, Sentimenti,t represents the investor sentiment corresponding to the day t of the i-th fund, εi,t represents the random error term, and the rest are control variables. Age means the fund establishment years. Cus and Mana are the fee ratios charged by the custodian bank to the fund for safekeeping and managing the fund assets. Leverage is equal to the total fund size divided by the fund’s net assets. Stockfive stands for the percentage of the fund’s portfolio held in its top five stocks. Accunet is equal to asset unit net value plus accumulated unit dividends. Annualreturn is the annual return of the corresponding fund. Maxdrawdown is the max drawdown of the corresponding fund. Maxreturn is the max return of the corresponding fund. Change is the fund price amplitude, which is equal to the closing price of the day minus the closing price of the previous day, then divided by the closing price of the previous day.
After correlation analysis and excluding the influence of multicollinearity, this study compares the results of using random effects or fixed effects. According to the Hausman test results, the fixed effects specification is better than the ordinary least squares (OLS) model, the random effects model is superior to OLS, and the fixed effects model is preferred over the random effects model. Therefore, this study estimates the model using fixed effects. Since the selected control variables fund establishment years (Age), maximum drawdown during tenure (Maxdrawdown) and maximum return during tenure (Maxreturn) already include the heterogeneous performance of individual funds, we only use time fixed effects to show the heterogeneous performance of funds in the time dimension. Meanwhile, considering the heteroskedasticity problem, we adopt the robust standard error.
This study scrapes open-end partial equity fund and hybrid fund user comments from Eastmoney (www.eastmoney.com), and the sample period is set from June 1, 2023 to January 1, 2025. We utilize Spyder to scrape more than 1.5 million user comments and manually label them as the training set. After cleaning the data, 351,904 samples are finally obtained. Following Yang et al. (2020), we adopt the FinBERT model for further classification tasks to construct an investor sentiment index. The daily unit net value of the fund is obtained from the Choice financial database. The weight matrix and network distance are calculated by using the complex dynamic network, and the transition probability under the steady state is calculated by using the Markov chain, ultimately yielding the dynamic entropy of the individual fund as the proxy index of fund market resilience. Control variables are obtained from the Wind database. Their summary statistics are shown in Table 1. Overall, the variables exhibit substantial cross-sectional and time-series variation, providing a suitable basis for econometric identification.
Summary statistics.
To further assess potential multicollinearity concerns, Appendix C (Fig. C1) presents the pairwise Pearson correlation coefficients among the main variables. Several correlations are statistically significant due to the large sample size. However, their magnitudes are generally modest. In particular, the correlation between investor sentiment and fund resilience is positive but economically small, indicating that the baseline relationship is unlikely to be driven by simple linear dependence. Among control variables, fund age and accumulated net value exhibit a relatively high correlation, reflecting the natural accumulation process of fund growth over time. Nevertheless, no pairwise correlation exceeds conventional thresholds associated with severe multicollinearity. Consistent with this observation, variance inflation factor (VIF) diagnostics with an average VIF of 1.339 remain well below the critical threshold of 10, suggesting that multicollinearity does not materially affect the regression estimates.
Empirical analysisMain resultsTable 2 shows the baseline regression results of investor sentiment on fund market resilience after gradually adding control variables. Specifically, the systematic reduction in sentiment coefficients from 0.048 to 0.034 across specifications reflects the incremental explanatory power of control variables while preserving the core relationship’s statistical and economic significance. The regression coefficient of Column 1 is 0.048 and is significant at the 1% level, indicating that for each 1-unit increase in investor sentiment, fund resilience increases by 0.048 units. Column 2 adds fixed fund characteristics such as fund manager tenure, management fees, custody fees, leverage ratio, and percentage share of top 5 stocks to the regression. The regression coefficient becomes 0.044 and is still significant at the 1% level. Column 3 further adds fund return characteristics, such as cumulative returns and annualized returns, to the regression, and the results remain roughly the same. Notably, the coefficient on Annualreturn shifts from negative in Column 3 to positive in Column 4 upon the inclusion of downside risk measures (Maxdrawdown, Maxreturn, and Change). This reversal indicates that once maximum drawdown and maximum return are held constant, the residual component of annual returns reflects fund-level performance quality rather than risk exposure, thereby shifting its association with resilience from negative to positive. This finding underscores the importance of jointly controlling for both performance and risk dimensions when modeling fund resilience. Column 4 contains all the control variables, and the regression coefficient is reduced to 0.034 but is still significant at the 1% level. The results are more robust and further validate that investor sentiment has a significant positive effect on fund market resilience.
Benchmark regression results of investor sentiment on fund market resilience.
| (1) Resilience | (2) Resilience | (3) Resilience | (4) Resilience | |
|---|---|---|---|---|
| Sentiment | 0.048*** | 0.044*** | 0.044*** | 0.034*** |
| (28.851) | (26.523) | (26.354) | (20.226) | |
| Age | -0.004*** | 0.002*** | 0.002*** | |
| (-19.724) | (6.711) | (7.221) | ||
| Cus | 0.437*** | 0.412*** | 0.549*** | |
| (8.017) | (7.569) | (10.033) | ||
| Mana | -0.276*** | -0.278*** | -0.316*** | |
| (-31.078) | (-31.359) | (-35.129) | ||
| Leverage | -0.027** | -0.028** | -0.014 | |
| (-1.976) | (-2.080) | (-1.059) | ||
| Stockfive | -0.022** | -0.023** | -0.023** | |
| (-2.178) | (-2.280) | (-2.264) | ||
| Accunet | -0.034*** | -0.017*** | ||
| (-25.582) | (-12.707) | |||
| Annualreturn | -0.001*** | 0.000*** | ||
| (-9.248) | (4.780) | |||
| Maxdrawdown | -0.004*** | |||
| (-60.067) | ||||
| Maxreturn | -0.000*** | |||
| (-30.177) | ||||
| Change | -0.034*** | |||
| (-31.211) | ||||
| Constant | 0.011*** | 0.304*** | 0.327*** | 0.232*** |
| (10.042) | (16.578) | (17.799) | (12.631) | |
| TimeFE | YES | YES | YES | YES |
| N | 351904 | 351904 | 351904 | 351904 |
| R2 | 0.665 | 0.667 | 0.667 | 0.673 |
| ΔR² | 0.000*** | 0.000*** | 0.000*** |
Note: Robust t-statistics in parentheses; ***, **, and * represent significance at the 1%, 5%, and 10% levels, respectively.
A 1-unit increase in investor sentiment is associated with a 0.034-unit increase in normalized fund market resilience. Scaling by the standard deviation of sentiment in Table 1 (SD = 0.564), a 1-standard-deviation increase in sentiment is associated with a 0.019-unit increase in normalized resilience—equivalent to approximately 1.9% of 1 standard deviation of the resilience distribution. The statistical significance of changes in model explanatory power (ΔR²) across specifications is assessed using Wald tests on newly added variable blocks. Although the ΔR² increments are small in absolute magnitude, the Wald tests confirm that each block of added variables contributes statistically significant explanatory power, consistent with the large sample size (N = 351,904) providing high statistical power to detect small but real improvements in model fit. The results indicate that each model extension leads to a statistically significant improvement in explanatory power, confirming that fund characteristics, performance indicators, and risk-related variables provide incremental explanatory content beyond investor sentiment alone. These results provide strong evidence for the causal impact of investor sentiment on fund market resilience, supporting the behavioral finance theory that investor sentiment significantly influences market stability.
Furthermore, we regress the full samples of resilience into different sentiment types (high, low, and neutral sentiment), as shown in Table 3. The regression coefficient of Column 1 is -0.092 and is significant at the 1% level, indicating that when investor sentiment is high, market resilience will be significantly reduced. This further reflects that investors may engage in irrational behaviors such as overconfidence when their emotions are high, which leads to reduced market stability. The regression coefficient of Column 2 is 0.109 and is significant at the 1% level, indicating that when the investor sentiment is low, fund market resilience increases. The main reason may be that investors in low-sentiment markets are more cautious about their capital, thus preferring lower price value but more stable investments, and fund market resilience is also strengthened. Together, these two findings demonstrate the asymmetric effect of investor sentiment on fund market resilience. Column 3 is the neutral sentiment sample with the regression coefficient of 0. Under these circumstances, the market is relatively stable, and marginal sentiment changes do not exert a significant impact.
Benchmark regression results of different investor sentiments on fund market resilience.
| (1) High sentiment Resilience | (2) Low sentiment Resilience | (3) Neutral sentiment Resilience | |
|---|---|---|---|
| Sentiment | -0.092*** | 0.109*** | 0.000 |
| (-9.104) | (26.787) | (.) | |
| Age | -0.002*** | 0.001*** | -0.000 |
| (-2.593) | (2.946) | (-0.315) | |
| Cus | 0.583*** | 0.933*** | 0.376*** |
| (4.613) | (10.106) | (4.966) | |
| Mana | -0.021 | -0.053*** | -0.457*** |
| (-0.894) | (-3.363) | (-37.939) | |
| Leverage | 0.108*** | 0.080*** | 0.056*** |
| (28.901) | (39.497) | (28.265) | |
| Stockfive | -0.038*** | -0.040*** | -0.046*** |
| (-12.640) | (-24.653) | (-32.382) | |
| Accunet | -0.001 | -0.006*** | -0.011*** |
| (-0.410) | (-3.328) | (-8.009) | |
| Annualreturn | 0.001*** | 0.000 | 0.001*** |
| (3.719) | (0.524) | (7.097) | |
| Maxdrawdown | -0.002*** | -0.003*** | -0.003*** |
| (-9.970) | (-22.875) | (-41.707) | |
| Maxreturn | -0.000*** | -0.000*** | -0.000*** |
| (-4.183) | (-11.817) | (-18.633) | |
| Change | -0.030*** | -0.011*** | -0.025*** |
| (-11.787) | (-7.027) | (-17.105) | |
| Constant | -0.001 | -0.070*** | 0.392*** |
| (-0.037) | (-3.368) | (28.362) | |
| TimeFE | YES | YES | YES |
| N | 36259 | 123365 | 192280 |
| R2 | 0.697 | 0.703 | 0.678 |
Note: Robust t-statistics in parentheses; ***, **, and * represent significance at the 1%, 5%, and 10% levels, respectively. An undefined t-statistic results from the inability to estimate the coefficients of the neutral sentiment observations since Sentiment = 0 for all neutral group data (the variance of the regression variable is zero).
Similarly, we regress the full samples of resilience according to different quantiles (top 10%, top 25%, top 50%, top 75%, top 90%), as shown in Table 4. The regression coefficients increase monotonically with the quantile. To provide a comprehensive understanding, we synergize the parametric regression results with non-parametric visualizations, as shown in Fig. 3. To mitigate the influence of extreme values on the non-linear relationship, the distribution plot uses kernel density estimation (KDE) (see Appendix A) to display the centroid of the sample. It represents the joint distribution of standardized investor sentiment (measured in standard deviations) and normalized Shannon entropy as market resilience. The overlaid curves show quantile regression fits (10th, 25th, 50th, 75th, 90th percentiles) using restricted cubic splines with four knots at default percentiles (5th, 35th, 65th, 95th), demonstrating heterogeneous sentiment effects across the resilience distribution. The shading intensity indicates the bivariate probability density, with darker regions representing higher data concentration. The five equally thick fitted lines in the distribution plot demonstrate that the fund’s response pattern to investor sentiment is highly consistent across different quantile levels. As the percentile increases, the slope of the fitted curve becomes steeper in the high-sentiment region, which is consistent with the increasing absolute values of the coefficients, suggesting that funds with higher resilience may exhibit incremental resilience decline under extreme sentiment shocks. The quantile regression results can further demonstrate the universality of this nonlinear impact. This spatial distribution characteristic also provides solid empirical support for the significant asymmetric correlation results shown in Table 3, demonstrating that high sentiment weakens resilience is not driven by a few outliers.
Quantile regression results of investor sentiment on fund market resilience.
| (1) Top 10% Resilience | (2) Top 25% Resilience | (3) Top 50% Resilience | (4) Top 75% Resilience | (5) Top 90% Resilience | |
|---|---|---|---|---|---|
| Sentiment | 0.013*** | 0.014*** | 0.025*** | 0.031*** | 0.035*** |
| (2.719) | (5.166) | (12.628) | (17.911) | (20.873) | |
| Age | -0.001 | -0.001** | -0.000 | -0.000 | -0.001** |
| (-0.968) | (-2.175) | (-1.049) | (-0.878) | (-2.035) | |
| Cus | 0.237* | 0.365*** | 0.510*** | 0.673*** | 0.690*** |
| (1.930) | (4.729) | (8.223) | (12.149) | (13.007) | |
| Mana | -0.124*** | -0.204*** | -0.262*** | -0.283*** | -0.277*** |
| (-6.247) | (-16.657) | (-26.043) | (-30.609) | (-31.473) | |
| Leverage | 0.052*** | 0.054*** | 0.061*** | 0.074*** | 0.077*** |
| (15.459) | (25.683) | (40.289) | (56.476) | (59.275) | |
| Stockfive | 0.001 | -0.008*** | -0.023*** | -0.038*** | -0.046*** |
| (0.569) | (-5.304) | (-19.288) | (-36.284) | (-45.243) | |
| Accunet | 0.002 | 0.003** | 0.000 | -0.002** | -0.004*** |
| (0.701) | (2.169) | (0.135) | (-2.125) | (-3.593) | |
| Annualreturn | 0.000** | 0.001*** | 0.001*** | 0.001*** | 0.000*** |
| (2.081) | (8.549) | (12.205) | (9.119) | (6.506) | |
| Maxdrawdown | 0.000 | -0.001*** | -0.002*** | -0.002*** | -0.003*** |
| (0.114) | (-9.974) | (-25.067) | (-34.444) | (-41.022) | |
| Maxreturn | -0.000*** | -0.000*** | -0.000*** | -0.000*** | -0.000*** |
| (-3.981) | (-14.518) | (-23.012) | (-23.989) | (-21.690) | |
| Change | -0.015*** | -0.014*** | -0.020*** | -0.025*** | -0.025*** |
| (-5.177) | (-7.750) | (-15.026) | (-22.852) | (-25.241) | |
| Constant | 1.835*** | 1.375*** | 0.939*** | 0.572*** | 0.343*** |
| (87.582) | (97.572) | (78.040) | (51.457) | (32.270) | |
| TimeFE | YES | YES | YES | YES | YES |
| N | 35172 | 87973 | 175951 | 263928 | 316714 |
| R2 | 0.784 | 0.660 | 0.569 | 0.566 | 0.609 |
Note: Robust t-statistics in parentheses; ***, **, and * represent significance at the 1%, 5%, and 10% levels, respectively.
To further examine whether the omission of explicit sentiment intensity biases our baseline findings, we classify sentiment observations into four regimes based on both direction and extremeness: high-extreme, high-mild, low-extreme, and low-mild, with |Sentiment|>0.5 as “extreme”, 0<|Sentiment|≤0.5 as “mild”. The corresponding regression results are reported in Appendix C (Table C1). Across all four regimes, the estimated effects of investor sentiment on fund resilience remain statistically significant and economically meaningful. Notably, the magnitude of the sentiment coefficient is largest under the high-extreme specification, consistent with the expectation that extreme sentiments exert stronger effects on resilience. These findings suggest that although intensity weighting may refine the sentiment index, its omission would not materially bias our core estimates.
Robustness testTo verify the reliability of the benchmark regression results and to ensure the robustness of the empirical conclusions, we perform a series of robustness checks. Specifically, we re-estimate the models after applying 1% and 5% winsorization to all variables, after excluding all observations from 2023, and after removing both the abnormal impact periods and their corresponding extended spans. The results are reported in Table 5.
Robustness test results for the baseline regression.
| (1) 1% winsorize | (2) 5% winsorize | (3) Exclude 2023 | (4) Exclude shock | (5) Exclude Extended shock | |
|---|---|---|---|---|---|
| Resilience | Resilience | Resilience | Resilience | Resilience | |
| Sentiment | 0.042*** | 0.041*** | 0.024*** | 0.039*** | 0.037*** |
| (24.909) | (22.743) | (10.790) | (22.765) | (21.021) | |
| Age | 0.001* | 0.001*** | -0.003*** | 0.000 | 0.000 |
| (1.901) | (3.988) | (-6.478) | (0.557) | (1.199) | |
| Cus | 0.681*** | 0.675*** | 0.958*** | 0.608*** | 0.657*** |
| (11.829) | (7.379) | (13.694) | (10.788) | (11.227) | |
| Mana | -0.395*** | -0.319*** | -0.295*** | -0.340*** | -0.302*** |
| (-41.750) | (-35.659) | (-25.142) | (-36.274) | (-30.812) | |
| Leverage | 0.066*** | 0.074*** | 0.117*** | 0.069*** | 0.073*** |
| (49.908) | (51.707) | (62.516) | (50.522) | (50.527) | |
| Stockfive | -0.045*** | -0.050*** | -0.040*** | -0.046*** | -0.048*** |
| (-44.685) | (-50.828) | (-29.252) | (-43.922) | (-43.328) | |
| Accunet | -0.012*** | -0.016*** | 0.006*** | -0.009*** | -0.009*** |
| (-9.107) | (-11.098) | (4.103) | (-8.455) | (-8.278) | |
| Annualreturn | 0.001*** | 0.001*** | 0.000 | 0.000*** | 0.000*** |
| (7.572) | (6.950) | (1.338) | (6.566) | (6.673) | |
| Maxdrawdown | -0.003*** | -0.003*** | -0.003*** | -0.003*** | -0.003*** |
| (-55.928) | (-53.807) | (-35.534) | (-52.461) | (-48.972) | |
| Maxreturn | -0.000*** | -0.000*** | -0.000*** | -0.000*** | -0.000*** |
| (-24.981) | (-21.101) | (-13.977) | (-23.960) | (-21.814) | |
| Change | -0.026*** | -0.034*** | -0.010*** | -0.020*** | -0.025*** |
| (-23.977) | (-27.811) | (-8.128) | (-20.034) | (-23.695) | |
| Constant | 0.276*** | 0.193*** | 0.303*** | 0.249*** | 0.248*** |
| (24.278) | (12.505) | (22.167) | (21.944) | (20.741) | |
| TimeFE | YES | YES | YES | YES | YES |
| N | 351904 | 351904 | 206981 | 336036 | 309210 |
| R2 | 0.653 | 0.646 | 0.684 | 0.668 | 0.644 |
Note: Robust t-statistics in parentheses; ***, **, and * represent significance at the 1%, 5%, and 10% levels, respectively.
To identify abnormal market periods, we first cluster episodes of excessive market fluctuations using the daily mean rate of change. We then apply a two-standard-deviation rule to detect days of extreme volatility, defined as abnormal impact periods. Each identified period is subsequently expanded by one day before and after to obtain the extended impact period. The definitions are as follows:
where ΔXt represents the volatility of the corresponding indicator on the day t, σ is the corresponding standard deviation, the logical disjunction symbol ∨ represents “or” meaning that the condition holds if at least one inequality in the set is satisfied. Moreover, Shockt equals 1 when the event At is satisfied, indicating an abnormal impact day, and 0 otherwise. Similarly, ShockExtendt equals 1 when the event Bt holds—i.e., if day t or either of its adjacent days (t−1 or t+1) is identified as an abnormal impact day—and 0 otherwise. I(·) represents the indicator function defined as:
The abnormal period identification results are shown in Fig. 4. The main abnormal impact days are 15 days, accounting for 3.9% of the total days. The total sample size corresponding to these 15 days is 15,868, and these observations are subsequently removed. Furthermore, the period before and after the impact period is taken as the extended period, corresponding to a total sample size of 42,694 individuals, and robust regression is performed to account for sample removal.
Based on the above identification of abnormal shock periods, Table 5 shows the robustness test regression results. Columns 1 and 2 report results after winsorizing all continuous variables at the 1% and 5% levels. The regression coefficients of 0.042 and 0.041 show that investor sentiment has a positive impact on the fund market resilience and are still significant at the 1% level. Column 3 excludes 2023 and uses a one-year period for regression. The reduced coefficient in the shorter post-2023 window may reflect a relatively calmer sentiment environment in 2024, consistent with the stimulus-driven market stabilization observed during this period. Columns 4 and 5 mainly exclude the above abnormal shock period and the extended abnormal shock period samples for regression. The regression results of both columns are still robust and significant at the 1% level, which are roughly consistent with the benchmark regression result coefficient of 0.034.
Heterogeneity testIn the heterogeneity analysis, we further investigate how the relationship between investor sentiment and fund market resilience varies with fund manager educational background, investment style, and geographic location. The results reveal systematic differences across these dimensions, indicating that the influence of sentiment on resilience is contingent on managerial characteristics, portfolio structure, and regional context. These findings are consistent with the notion that behavioral channels operate heterogeneously rather than uniformly across different institutional and market environments. Full regression results are reported in Appendix C Tables C2-C4.
Fund manager educationThe educational background of fund managers provides an important proxy for their analytical capacity and their ability to process market information under uncertainty. Prior studies show that more highly educated investors generally possess stronger information acquisition and analytical skills, which help them filter market noise and mitigate cognitive biases more effectively (Calvet et al., 2009). By contrast, less educated investors tend to rely more on heuristics and are more prone to emotional reactions and herd behavior, which may increase the likelihood of suboptimal decision making (Baker & Wurgler, 2006).
We therefore partition the sample according to managers’ highest academic degree, distinguishing between PhD, master’s, and bachelor’s degree holders. The results reveal an unexpected pattern. The estimated sentiment coefficient is largest for managers with a bachelor’s degree (β1=0.108,p<0.01), approximately three times larger than for PhD (β1=0.041,p<0.01) or master’s degree holders (β1=0.039,p<0.01). One possible interpretation is that highly educated managers are more likely to rely on proprietary information, quantitative models, and historical data when making investment decisions, which allows them to filter out short-term sentiment fluctuations. However, their superior training may paradoxically foster a degree of cognitive overconfidence, making them more inclined to rely on their own judgment and engage in contrarian trading during periods of extreme sentiment, thereby weakening the direct transmission of sentiment to portfolio decisions. In contrast, managers with bachelor’s degrees may have more limited access to sophisticated analytical tools and thus depend more heavily on publicly observable sentiment signals, leading to a tighter coupling between investor sentiment and fund behavior. Detailed regression results are reported in Appendix C Table C2.
Investment styleFund investment style reflects heterogeneity in asset allocation strategies and risk appetite, potentially moderating sentiment transmission. Kumar and Lee (2006) documented systematic differences in investment philosophy, portfolio structure, and liquidity characteristics across value, growth, and balanced funds. Brown and Cliff (2005) showed that growth funds, investing in high-volatility assets, may be more sensitive to sentiment fluctuations, whereas value funds, whose investments typically have more stable fundamentals, may exhibit greater resilience to sentiment shocks.
Classifying funds by balanced, growth, and income investment style, we find that balanced funds display the highest sentiment sensitivity (β1=0.064,p<0.01), indicating that the balanced-style fund may be more likely to change investment strategies when sentiment fluctuates, because they have flexible asset allocation strategies which make them more responsive to different market sentiment cycles. Growth funds show moderate effects (β1=0.031,p<0.01), indicating that investor sentiment has a relatively weak positive impact on the resilience of growth funds. Income funds exhibit the weakest and marginally significant response (β1=0.029,p<0.1), consistent with their long-term investment focus, which renders them less susceptible to short-term sentiment shifts (Appendix C Table C3).
Geographic locationCoval and Moskowitz (1999) emphasize that regional differences reflect variations in institutional environments, information infrastructure, and local investor preferences, all of which may shape investment behavior and market outcomes. Funds located in first-tier cities such as Beijing, Shanghai, and Guangdong generally benefit from more developed financial ecosystems, stronger information channels, and deeper talent pools, which tend to enhance professional decision-making and risk management capabilities. By contrast, funds based in lower-tier cities may face greater resource constraints and more limited access to high-quality information, potentially leading to distinct patterns of resilience.
We classify registration locations into three tiers, labeling "Beijing," "Shanghai," and "Guangdong" as first-tier cities, "Tianjin," "Zhejiang," "Fujian," "Xi’an," "Shandong," and "Chongqing" as second-tier cities, and "Tibet," "Guangxi," and "Xinjiang" as third-tier cities, and uncover striking regional asymmetries. For first-tier cities, the estimated sentiment effect is positive and statistically significant (β1=0.036,p<0.01), indicating that more mature financial environments allow institutional investors to rely on proprietary research and professional judgment to mitigate retail sentiment noise, thereby contributing to system stability. In contrast, funds in second-tier cities exhibit a significant negative sentiment effect (β1=−0.026,p<0.01). One plausible explanation is that these funds depend more heavily on internet-based public sentiment while lacking the same level of private information as first-tier institutions, which amplifies procyclical behavior during both market booms and downturns. For third-tier cities, the sentiment coefficient is not statistically significant (β1=0.018), which may reflect the fact that local investors and institutions in these regions tend to display lower risk tolerance and less engagement with online financial platforms, thereby reducing the transmission of digital sentiment to fund decisions. Detailed regression results are reported in Appendix C Table C4.
Mechanism testAlthough the baseline results indicate a significant positive association between investor sentiment and fund market resilience, the relationship is clearly asymmetric. In particular, elevated sentiment tends to weaken resilience, whereas subdued sentiment is associated with stronger resilience. Clarifying the underlying transmission channels is therefore essential for both theoretical interpretation and policy implications. To shed further light on the mechanisms at work, we focus on two primary channels through which sentiment may influence fund market resilience: price information dynamics and fund capacity constraints. These channels capture how sentiment interacts with market signals and institutional decision-making processes to shape resilience outcomes. Detailed regression specifications and results for the mechanism tests are reported in Appendix C Tables C5-C8. Moreover, Fig. 5 illustrates the estimated slopes of the moderating effects across different groups, providing a visual summary of how the strength of the sentiment-resilience relationship varies under alternative conditions.
Price information mechanism
Opening and closing prices represent two distinct stages of information aggregation within a trading day. The opening price reflects the market’s initial clearing outcome after incorporating overnight news, global market movements, and macroeconomic developments. Following Admati and Pfleiderer (1988), price formation can be viewed as an information aggregation process in which informed traders and noise traders interact under asymmetric information. During the opening auction, information asymmetry is typically high, and prices largely reflect the market’s collective interpretation of accumulated signals since the previous close. When the opening price is already high, it suggests that optimistic expectations have been substantially incorporated, leaving limited room for additional sentiment-driven upward adjustments.
The closing price represents the final valuation of the trading session and summarizes the cumulative impact of intraday information flows and trading activity. In the spirit of Kyle (1985), it can be interpreted as the market’s consensus price, where heterogeneous information is gradually incorporated through continuous trading. Compared with the opening price, the closing price embeds richer information because it reflects not only public news but also institutional rebalancing and order flow throughout the day. A strong closing price may indicate sustained positive sentiment, but it may also signal potential overvaluation, which can weaken the subsequent marginal effect of sentiment on resilience.
From an information asymmetry perspective, the impact of investor sentiment depends on how far the prevailing price deviates from its informational equilibrium. When opening or closing prices are high, bullish expectations are largely priced in, reducing the incremental role of sentiment in enhancing resilience. In contrast, when prices are low, the market may not have fully absorbed favorable information, allowing sentiment to play a more prominent role in stabilizing resilience.
Results confirm negative interaction effects for both closing prices (βinters=−0.005,p<0.01) and opening prices (βinters=−0.005,p<0.01), indicating diminishing marginal effects of sentiment at higher price levels (Appendix C Tables C5-C6). Group regressions further show that the sentiment coefficient is 0.042 in high-closing-price markets compared with 0.033 in low-closing-price markets, with similar patterns observed for opening prices (0.043 versus 0.033). This discrepancy possibly stems from a "sentiment saturation" effect: when the opening price reflects optimistic expectations, the marginal contribution of intraday sentiment to resilience is limited, as bullish signals have already been substantially incorporated into market prices. Fig. 5 (Panels A-B) visualizes these moderating dynamics, illustrating that despite higher baseline resilience in high-price markets, the slope of resilience in response to sentiment fluctuations is perceptibly suppressed. The low-price groups show stronger resilience recovery potential in the low sentiment range, indicating that there is more room for sentiment-driven stabilization in a low-valuation environment.
Fund capacity mechanismWhile market price variables reflect the broader information environment, the capacity of a fund to capture upside gains and limit downside losses serves as a critical determinant of its micro-level risk management efficacy and sentiment exposure. We examine two complementary dimensions of fund capacity: the up-market capture ratio, which quantifies the excess returns of a fund relative to its benchmark during bullish regimes, characterizing aggressiveness and market timing ability, and the down-market capture ratio, which evaluates capital preservation during bearish episodes, reflecting defensive capabilities and risk buffering capacity. These metrics reveal asymmetric moderation patterns which illuminate the heterogeneous transmission of investor sentiment across funds with varying operational characteristics.
High-profitability funds typically possess superior information advantages and active management capabilities (Cremers & Petajisto, 2009). During market expansions, these funds can accurately identify high-quality assets and adjust their portfolios promptly, thereby amplifying returns amidst buoyant investor sentiment. This performance advantage further attracts capital inflows, which in turn reinforce the fund’s centrality and resilience within the network. Conversely, the ability to withstand market downturns directly relates to a fund’s survival capacity under stress, aligning closely with the core concept of network resilience. Brunnermeier and Pedersen (2009) demonstrated that during market downturns, the liquidity spiral effect intensifies, and funds with strong resilience can mitigate the erosion of investor confidence caused by negative sentiment by holding low-volatility assets, maintaining sufficient cash reserves, or employing hedging strategies.
Regression results reveal a striking asymmetry between upside profitability and downside resilience moderation patterns (Appendix C Tables C7-C8). For up-market capture, the interaction coefficient is negative and significant (βinters=−0.008,p<0.01), indicating that profitability attenuates sentiment’s impact on resilience. Subgroup analysis yields a counterintuitive pattern: high-profitability funds exhibit a sentiment coefficient of 0.022, substantially lower than the 0.057 coefficient for low-profitability funds. In contrast, down-market capture exhibits a positive moderation effect (βinters=0.005,p<0.01), with high-resilience funds demonstrating a sentiment coefficient of 0.058 versus 0.029 for low-resilience funds—precisely the opposite pattern.
This divergence stems from distinct amplification effects. Highly profitable funds, already occupying network cores, possess elevated baseline resilience levels, generating a ceiling effect. When sentiment’s marginal contribution encounters a numerical ceiling, there is limited room for further improvement when resilience is already high. Fig. 5 (Panel C) illustrates this constraint: the high-profitability group maintains a consistently higher intercept, but its slope (0.022) is significantly flatter than the slope of low-profitability group (0.057). However, the two trajectories exhibit a converging trend: the resilience gap narrows when sentiment reaches extreme optimism (Sentiment = +2), suggesting that profitability serves as an important resilience support under normal conditions, but its incremental effect diminishes during periods of heightened emotional volatility.
Conversely, downside resilience generates a multiplicative amplification effect. Funds with strong downside protection effectively maintain investor confidence during negative sentiment periods, reducing panic redemptions and preserving stable network connections. When market sentiment turns positive, this established safety premium attracts capital inflows from risk-averse investors, further strengthening resilience. However, low-resilience funds experience amplified losses during downturns. Even when sentiment improves, it is not easy to restore investor confidence, leading to capital outflows and network isolation. Fig. 5 (Panel D) visualizes this amplification: the slope of high-resilience group (0.058) is significantly steeper than the low-resilience group’s (0.029). Critically, the two trajectories display a divergent trend: the resilience gap is largest when sentiment is extremely negative (Sentiment = -2) and further expands when sentiment becomes extremely positive (Sentiment = +2), indicating that downside resilience creates self-reinforcing momentum rather than encountering diminishing returns.
Endogeneity testTo address the endogeneity problem of investor sentiment, we use the percentage change in the value of heavily weighted stocks as an instrumental variable. Our identification strategy relies on the premise that such fluctuations serve as exogenous sentiment shocks. The validity of the instrumental variable rests on two conditions. First, the relevance condition: fluctuations in the price of a fund’s most heavily weighted stock directly influence the expected trajectory of the fund’s net asset value, thereby generating significant and measurable variation in investor sentiment. This ensures a strong and statistically significant first-stage correlation between the instrument and the endogenous variable. Second, the exclusion restriction: price movements in a single heavily weighted stock are primarily driven by firm-level fundamentals and industry-specific factors, rather than by the structural stability of the broader fund market network. Accordingly, the instrument affects fund market resilience exclusively through the sentiment channel, and exerts no direct influence on resilience independent of this pathway.
As reported in Table 6, the first-stage regression results (Column 2) reveal a highly significant positive correlation between the IV and the endogenous sentiment variable. A negative R² in the 2SLS second stage is expected when the IV-predicted values of the endogenous regressor fit the dependent variable less well than the unconditional mean. This is a standard feature of 2SLS estimation and does not indicate model misspecification. The F-statistic of 540.19 significantly surpasses the Stock-Yogo critical value of 10 (Stock & Yogo, 2002), effectively ruling out weak instrument concerns. The second-stage estimates in Column 3 demonstrate that investor sentiment exerts a significant negative impact on fund market resilience (significant at the 5% level). Notably, the IV coefficient is not only larger in magnitude than the baseline estimate (0.034) in Table 2 but also exhibits an opposite sign (-0.472), suggesting that the OLS estimate substantially attenuates the true destabilizing impact of sentiment by averaging across heterogeneous sentiment regimes. The OLS coefficient captures the average partial correlation between sentiment and resilience across the full distribution of sentiment states, encompassing both mild positive and negative sentiment regimes in which sentiment generally exerts a stabilizing influence (as documented in Table 3). By contrast, the instrumental variable generates exogenous variation that is disproportionately concentrated in episodes of extreme sentiment surges, since large-cap stock movements predominantly spike during bullish market conditions. Therefore, the 2SLS estimate isolates the causal effect of exogenous, high-sentiment shocks on resilience, capturing the destabilizing states identified in Table 3, Column 1. This finding reinforces the asymmetric impact of sentiment in Table 3. Exuberant sentiment fosters momentum trading and investment concentration, thereby heightening market fragility and undermining resilience. By contrast, pessimistic shocks induce cautious diversification, which paradoxically strengthens the fund market resilience. Taken together, the OLS and 2SLS results are mutually reinforcing rather than contradictory: both confirm that the net effect of sentiment is regime-dependent, consistent with the core asymmetry thesis of our study.
Instrumental variable method based on the price changes of heavily weighted stocks.
| (1) OLS | (2) First Stage | (3) Second stage | |
|---|---|---|---|
| Sentiment | 0.034*** | -0.472** | |
| (8.534) | (-2.521) | ||
| Age | 0.002*** | -0.000 | 0.002*** |
| (4.400) | (-0.321) | (4.235) | |
| Cus | 0.549*** | -0.253*** | 0.421*** |
| (7.138) | (-4.615) | (4.493) | |
| Mana | -0.316*** | -0.130*** | -0.382*** |
| (-9.068) | (-14.322) | (-8.588) | |
| Leverage | -0.014 | 0.012 | -0.008 |
| (-1.006) | (0.920) | (-0.501) | |
| Stockfive | -0.023** | -0.013 | -0.029** |
| (-2.299) | (-1.356) | (-2.581) | |
| Accunet | -0.017*** | 0.024*** | -0.005 |
| (-7.573) | (16.203) | (-0.973) | |
| Annualreturn | 0.000** | 0.003*** | 0.002*** |
| (2.349) | (25.051) | (3.380) | |
| Maxdrawdown | -0.004*** | -0.003*** | -0.005*** |
| (-21.464) | (-31.317) | (-9.635) | |
| Maxreturn | -0.000*** | -0.001*** | -0.001*** |
| (-13.913) | (-63.018) | (-4.737) | |
| Change | -0.034*** | 0.061*** | -0.000 |
| (-4.390) | (19.938) | (-0.003) | |
| IV | 0.005*** | ||
| (10.675) | |||
| Constant | 0.232*** | -0.082*** | |
| (4.584) | (-4.457) | ||
| TimeFE | YES | YES | YES |
| N | 351904 | 351904 | 351904 |
| R2 | 0.673 | 0.059 | -0.206 |
Note: Robust t-statistics in parentheses; ***, **, and * represent significance at the 1%, 5%, and 10% levels, respectively.
Our baseline OLS estimate documents a positive average association between investor sentiment and fund market resilience: a 1-unit increase in sentiment is associated with a 0.034-unit increase in normalized resilience (β=0.034, p<0.01). This average effect, however, masks a pronounced asymmetry across sentiment regimes. During high-sentiment periods, excessive optimism drives speculative trading and portfolio concentration, which ultimately undermines market stability and weakens resilience (β=−0.092, p<0.01). Conversely, low-sentiment periods foster cautious and diversified investment behavior, thereby strengthening resilience (β=0.109, p<0.01). The endogeneity-corrected 2SLS estimate (β=−0.472, p<0.05) further corroborates this destabilizing channel by isolating the causal impact of exogenous high-sentiment shocks, which are disproportionately associated with episodes of extreme market optimism captured by the instrumental variable. Additional research indicates that this asymmetric pattern is systematic rather than incidental, varying with fund characteristics, pricing mechanism, and fund capacity constraints. Overall, our findings suggest that sentiment is stabilizing in pessimistic markets but destabilizing when optimism becomes exuberant, highlighting a nonlinear role of investor psychology in shaping market resilience.
Our findings carry significant implications for financial regulation, risk management, and investment strategy. For regulators, we recommend incorporating sentiment asymmetry into market stabilization and disclosure frameworks. During periods of extreme optimism, regulatory tools such as circuit breakers or risk-warning thresholds could be tightened to curb speculative behavior, while during pessimistic phases, these mechanisms may be adjusted to support market liquidity and prevent overcorrection. In addition, a standardized disclosure regime reporting funds’ sentiment sensitivity could enhance market transparency and improve risk monitoring. For fund managers, advances in natural language processing tools, such as FinBERT, can allow them to develop real-time systems for monitoring investor sentiment. When sentiment fluctuations exceed predefined thresholds, mandatory stress tests should be enforced, and asset risk exposures could be dynamically adjusted. For managers with bachelor’s degrees (who have higher sentiment sensitivity), targeted behavioral finance training should be provided to help offset psychological biases arising from limited quantitative and analytical expertise. For investors, our results highlight the importance of aligning fund selection with individual psychological risk tolerance. During bull markets, it is recommended that investors use the heterogeneity of fund features to achieve risk hedging by increasing the portfolio allocation of low-sensitivity, high-resilience funds to maintain stability against sentiment reversals. Overall, by aligning regulatory frameworks, managerial conduct, and investor behavior with the underlying sentiment-resilience dynamics, the financial ecosystem can foster a more sustainable and robust equilibrium.
Despite the systematic empirical analysis conducted above, this study has several limitations worth exploring. First, our sentiment measure relies on textual data from a single online platform (East Money Forum), which may not fully capture comprehensive investor beliefs across different communication channels and could introduce platform-specific biases. While FinBERT demonstrates strong performance in financial text analysis, it still suffers from certain precision biases, resulting in discrepancies in the classification tasks of the index construction. Future research could validate our findings using alternative sentiment proxies or multi-source textual data. Second, the construction of return-based fund networks may be affected by common risk exposures across funds, which could generate relatively sparse network connections that do not fully reflect underlying interdependencies. While more advanced techniques such as partial correlation networks could help mitigate this issue by filtering out common factor effects, data dimensionality and computational constraints limit its implementation in our study. We therefore acknowledge this as an important limitation and leave it for future research. Third, our heterogeneity analysis identifies significant variation across manager education, investment style, and fund location, yet the underlying behavioral mechanisms remain incompletely understood. Future research could take other potential factors such as manager’s work experience into account. Finally, we select the opening and closing prices as the price information mechanism to examine the causal relationship. The use of opening and closing prices as proxies for the price information mechanism is limited by their high intraday correlation and their inability to fully capture the rich dynamics of intraday order flow. Future studies could address this by employing tick-level transaction data or microstructure noise measures to decompose price discovery more precisely.
CRediT authorship contribution statementWeijie Chen: Writing – review & editing, Writing – original draft, Methodology, Funding acquisition. Bingqing Lu: Writing – review & editing, Writing – original draft, Methodology, Formal analysis, Data curation. Jiasen Tian: Software, Methodology, Data curation. Yafen Ye: Writing – review & editing, Supervision, Funding acquisition.
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
This work is supported by the National Natural Science Foundation of China (No. 12101552, 12271131, 11871183 and 61603338), the Natural Science Foundation of Zhejiang Province (No. LY21F030013), Philosophy, the Social Sciences Leading Talent Training Project of Zhejiang Province (No. 21YJRC07-1YB), Social Science Foundation of Zhejiang University of Technology (GB202303001) and the Zhejiang Provincial Department of Education Research Project (Y202455759).
To visually characterize the intricate relationship between investor sentiment and fund market resilience, we employ a visualization framework that integrates joint density estimation with quantile regression based on restricted cubic splines (RCS). The technical specifications are as follows:
A.1. Joint Density Estimation (KDE)
To visualize the sample distribution of 351904 observations without the visual clutter of a traditional scatter plot, we utilize Bivariate Kernel Density Estimation (KDE) (Parzen, 1962) to estimate the joint probability density function f(x,y):
where K is the Gaussian kernel and h is the bandwidth matrix optimized via the plug-in method to balance bias and variance. The shaded contours represent the estimated joint probability density function of Sentiment(x) and Hd(y). The density illustrates that the observations are predominantly concentrated near the origin, with a notable shift toward the second and fourth quadrants, providing intuitive evidence for the baseline correlation identified in Table 3.
A.2. Nonlinear Quantile Specification (Restricted Cubic Splines)
Instead of assuming a simple quadratic form (e.g., Sentiment2), we model the nonlinear conditional mean and quantiles using Restricted Cubic Splines (RCS) to allow for maximum flexibility while maintaining smoothness. The model is specified as:
where Bj denotes the spline basis functions. Qτ(·)istheτ−thconditionalquantileofHd.Hd denotes the entropy-based resilience. We utilize four knots (df=4) placed at the default percentiles, which ensures the curves are sufficiently smooth in the center while remaining linear at the boundaries to prevent overfitting in the tails.
A.3. Quantile Regression (QR)
While Ordinary Least Squares (OLS) only estimates the conditional mean, the impact of sentiment may vary across funds with different stability profiles. We therefore estimate the RCS model across the conditional distribution of Hd. For each quantile τϵ{0.10,0.25,0.5,0.75,0.90}, we minimize the check-loss function:
A.4. Data Precision and Visualization Integrity
To ensure the visualization is representative of the true economic signal, we apply winsorization at the 1% and 99% levels. The vertical axis is truncated at the 0.5th and 99.5th percentiles to focus on the dense data regions where the econometric identification is strongest, effectively filtering out noise without loss of generality.
As a highly complex evolutionary system, the linkage relationship between various funds in the financial system can be represented by dynamic topological networks (Emmert-Streib et al., 2014). Complex dynamic networks can be analyzed from the perspective of graph theory and regarded as the interactive system between independent individuals and the whole. In this model, each independent individual is represented as a node, and the interactions between them can be reflected by edges. The entire network can thus be viewed as a set of points and a set of edges connecting these points.
B.1. Dynamic Network Construction
We construct dynamic networks using a sliding window approach with length Δt=30 days (sufficient observations N=30 for stable correlation estimation) and step size δt=1 day. The n-th network is constructed from returns spanning days t1n=1+(n−1)·δt and ends on day t2n=t1n+Δt. This overlapping design captures continuous shock propagation: a disruption on day 15 appears across 30 consecutive networks, revealing its full transmission dynamics. Non-overlapping windows would observe it only once, missing the propagation entirely. Each network is represented as:
where V is the set of N fund nodes,Et denotes edges between funds at day t, and Wt=[wij,t] is weighted adjacency matrix measuring co-movement strength between funds i and j. W is symmetric since wij=wji by construction.
B.2. Network Weight Construction
For each window, we compute pairwise correlations from daily returns Yi,t:
where Yi,t denotes the net value return rate of the i-th fund on the t-th day within the window. ρij,t represents the correlation coefficient between fund i and fund j on day t. We then transform correlations into connection weights using:
This exponential form maps highly correlated pairs (ρi,j=1) to strong connections (Wi,j=1) and weakly correlated pairs (ρi,j=−1) to weak connections (Wi,j=e−2), consistent with empirical correlation decay patterns in financial networks.
B.3. Markov Chain Price Transmission
We decompose the weight matrix through eigen-analysis:
The maximum eigenvalue λmax measures aggregate connectivity strength—it spikes during crises when correlations surge. The dominant right principal eigenvector v*=[v1*,…,vN*] (normalized to sum to 1∑ivi*=1) captures each fund’s structural centrality. Funds with large v1* occupy hub positions where shocks cascade broadly; those with small v1* are peripheral.
We construct transition probabilities governing shock propagation:
This satisfies ∑jρij,t=1(∑jpij,t=1λvi*∑jWij,tvj*=1λvi*(Wv*)i=1λvi*·λvi*=1), forming a valid Markov chain. The asymmetry in Pt=[pij,t] [P=1λWV−1(V=diag(v*))] captures directional influence: central funds (high vi*) disproportionately affect peripheral funds, reflecting empirical price leadership patterns. The steady-state distribution π solving πTPt=πT is:
This measures long-run systemic importance—high funds are critical for price discovery and risk propagation. High πi means fund i is central to the price discovery process—its shocks have persistent market-wide impact. Low πi means fund i is peripheral—its fluctuations remain localized.
B.4. Entropy-Based Resilience
We measure each fund’s resilience as the normalized entropy of its transmission channels:
We standardize z-score within each period for regression analysis:
Aggregate resilience weights individual fund entropy by systemic importance:
Benchmark regression results of investor sentiment on fund market resilience under different intensity and direction.
| (1) High-extreme Resilience | (2) High-mild Resilience | (3) Low-extreme Resilience | (4) Low-mild Resilience | |
|---|---|---|---|---|
| Sentiment | -0.191*** | 0.206* | 0.149*** | -0.292*** |
| (-16.691) | (1.689) | (34.323) | (-3.235) | |
| Age | -0.001 | -0.011*** | 0.002*** | -0.007*** |
| (-1.405) | (-3.603) | (3.651) | (-3.455) | |
| Cus | 0.556*** | 0.585 | 0.898*** | 0.842** |
| (4.141) | (1.479) | (9.403) | (2.474) | |
| Mana | -0.021 | 0.060 | -0.054*** | 0.063 |
| (-0.826) | (0.733) | (-3.355) | (0.970) | |
| Leverage | 0.106*** | 0.092*** | 0.079*** | 0.098*** |
| (26.772) | (8.415) | (37.971) | (12.869) | |
| Stockfive | -0.038*** | -0.023** | -0.039*** | -0.022*** |
| (-12.201) | (-2.306) | (-22.942) | (-3.203) | |
| Accunet | -0.004 | 0.026** | -0.006*** | 0.009 |
| (-1.282) | (2.234) | (-3.814) | (1.110) | |
| Annualreturn | 0.001*** | -0.000 | 0.000 | 0.000 |
| (3.462) | (-0.145) | (0.520) | (0.924) | |
| Maxdrawdown | -0.002*** | 0.001 | -0.003*** | 0.000 |
| (-9.551) | (1.450) | (-21.463) | (0.033) | |
| Maxreturn | -0.000*** | -0.000** | -0.000*** | -0.000 |
| (-2.805) | (-2.055) | (-9.853) | (-0.936) | |
| Change | -0.029*** | -0.021*** | -0.011*** | -0.015*** |
| (-10.762) | (-2.759) | (-6.654) | (-2.855) | |
| Constant | 0.095*** | -0.292*** | -0.016 | -0.483*** |
| (2.958) | (-2.751) | (-0.728) | (-5.397) | |
| TimeFE | YES | YES | YES | YES |
| N | 32963 | 3294 | 117372 | 5993 |
| R2 | 0.702 | 0.704 | 0.705 | 0.712 |
Note: Robust t-statistics in parentheses; ***, **, and * represent significance at the 1%, 5%, and 10% levels, respectively.
Benchmark regression results of investor sentiment on fund market resilience under different educational backgrounds.
| (1) PhD Resilience | (2) Master Resilience | (3) Bachelor Resilience | |
|---|---|---|---|
| Sentiment | 0.041*** | 0.039*** | 0.108*** |
| (9.361) | (21.181) | (6.121) | |
| Age | 0.001 | -0.000 | 0.013*** |
| (1.108) | (-0.647) | (3.158) | |
| Cus | 1.614*** | 0.507*** | 0.056 |
| (9.451) | (8.595) | (0.112) | |
| Mana | -0.595*** | -0.350*** | 0.023 |
| (-14.329) | (-37.152) | (0.234) | |
| Leverage | 0.059*** | 0.069*** | -0.005 |
| (15.771) | (48.358) | (-0.340) | |
| Stockfive | -0.063*** | -0.041*** | -0.080*** |
| (-20.835) | (-37.348) | (-6.992) | |
| Accunet | 0.008* | -0.008*** | -0.160*** |
| (1.819) | (-8.005) | (-8.726) | |
| Annualreturn | 0.001*** | 0.000*** | 0.008*** |
| (6.781) | (3.619) | (7.306) | |
| Maxdrawdown | -0.002*** | -0.004*** | -0.004*** |
| (-11.443) | (-53.662) | (-7.228) | |
| Maxreturn | -0.000*** | -0.000*** | -0.000*** |
| (-12.626) | (-21.450) | (-4.170) | |
| Change | -0.021*** | -0.021*** | -0.015* |
| (-7.816) | (-19.238) | (-1.656) | |
| Constant | 0.381*** | 0.253*** | 0.094 |
| (9.513) | (21.640) | (0.979) | |
| TimeFE | YES | YES | YES |
| N | 44850 | 301580 | 5474 |
| R2 | 0.692 | 0.679 | 0.625 |
Note: Robust t-statistics in parentheses; ***, **, and * represent significance at the 1%, 5%, and 10% levels, respectively.
Benchmark regression results of investor sentiment on fund market resilience under different investment style backgrounds.
| (1) Balanced Style Resilience | (2) Growth Style Resilience | (3) Income Style Resilience | |
|---|---|---|---|
| Sentiment | 0.064*** | 0.031*** | 0.029* |
| (3.736) | (18.449) | (1.676) | |
| Age | 0.011*** | 0.003*** | 0.014*** |
| (4.239) | (9.297) | (3.375) | |
| Cus | 0.000 | 0.370*** | 0.879*** |
| (.) | (6.564) | (5.280) | |
| Mana | 0.199 | -0.154*** | -0.155*** |
| (1.612) | (-15.689) | (-4.878) | |
| Leverage | 0.080 | -0.012 | 0.280*** |
| (0.600) | (-0.863) | (3.239) | |
| Stockfive | 0.255** | -0.025** | -0.070 |
| (2.402) | (-2.489) | (-1.130) | |
| Accunet | -0.142*** | -0.015*** | -0.152*** |
| (-11.140) | (-11.166) | (-5.523) | |
| Annualreturn | 0.020*** | 0.000*** | -0.016*** |
| (10.100) | (3.275) | (-6.779) | |
| Maxdrawdown | -0.009*** | -0.004*** | -0.006*** |
| (-9.961) | (-68.135) | (-6.548) | |
| Maxreturn | -0.001*** | -0.000*** | 0.001* |
| (-4.523) | (-28.077) | (1.890) | |
| Change | -0.032** | -0.035*** | -0.023 |
| (-2.467) | (-31.671) | (-1.081) | |
| Constant | -0.120 | 0.054*** | -0.181* |
| (-0.610) | (2.796) | (-1.803) | |
| TimeFE | YES | YES | YES |
| N | 3868 | 334897 | 13139 |
| R2 | 0.683 | 0.686 | 0.546 |
Note: Robust t-statistics in parentheses; ***, **, and * represent significance at the 1%, 5%, and 10% levels, respectively.
Benchmark regression results of investor sentiment on fund market resilience under different fund domicile backgrounds.
| (1) First-tier city Resilience | (2) Second-tier city Resilience | (3) Third-tier city Resilience | |
|---|---|---|---|
| Sentiment | 0.036*** | -0.026*** | 0.018 |
| (21.166) | (-2.674) | (0.893) | |
| Age | 0.002*** | -0.000 | 0.003 |
| (7.027) | (-0.049) | (0.636) | |
| Cus | 0.309*** | 0.972*** | -0.402 |
| (4.989) | (7.525) | (-0.781) | |
| Mana | -0.316*** | -0.225*** | 0.197** |
| (-33.430) | (-6.557) | (2.151) | |
| Leverage | -0.030** | 0.173** | 0.419*** |
| (-2.218) | (2.218) | (3.610) | |
| Stockfive | -0.021** | 0.009 | -0.147 |
| (-2.099) | (0.151) | (-1.643) | |
| Accunet | -0.018*** | -0.007 | 0.011 |
| (-12.813) | (-0.739) | (0.478) | |
| Annualreturn | 0.001*** | -0.001** | 0.011*** |
| (6.032) | (-2.204) | (6.426) | |
| Maxdrawdown | -0.004*** | -0.002*** | -0.001 |
| (-59.915) | (-6.910) | (-1.272) | |
| Maxreturn | -0.000*** | -0.000 | -0.001*** |
| (-30.524) | (-1.409) | (-7.732) | |
| Change | -0.035*** | -0.012** | -0.013 |
| (-31.363) | (-2.126) | (-0.882) | |
| Constant | 0.296*** | -0.155* | -0.464*** |
| (15.533) | (-1.692) | (-3.081) | |
| TimeFE | YES | YES | YES |
| N | 334499 | 13160 | 4245 |
| R2 | 0.676 | 0.634 | 0.744 |
Note: Robust t-statistics in parentheses; ***, **, and * represent significance at the 1%, 5%, and 10% levels, respectively.
The moderating results of investor sentiment on fund market resilience based on closing prices.
| (1) Centralized Resilience | (2) Moderating Effect Resilience | (3) High Closing Price Resilience | (4) Low Closing Price Resilience | |
|---|---|---|---|---|
| C.Sentiment | 0.040*** | 0.041*** | 0.042*** | 0.033*** |
| (24.000) | (24.109) | (18.442) | (13.703) | |
| C.spj | 0.012*** | 0.012*** | -0.007*** | 0.259*** |
| (6.142) | (6.104) | (-2.602) | (31.347) | |
| Age | 0.001*** | 0.001** | 0.006*** | -0.006*** |
| (2.617) | (2.573) | (13.451) | (-8.938) | |
| Cus | 0.563*** | 0.564*** | 0.152** | 1.140*** |
| (10.223) | (10.241) | (1.969) | (14.405) | |
| Mana | -0.355*** | -0.355*** | -0.259*** | -0.449*** |
| (-38.561) | (-38.550) | (-24.159) | (-26.333) | |
| Leverage | 0.068*** | 0.068*** | 0.067*** | 0.077*** |
| (50.498) | (50.523) | (34.890) | (39.996) | |
| Stockfive | -0.046*** | -0.046*** | -0.047*** | -0.049*** |
| (-44.577) | (-44.588) | (-33.064) | (-32.519) | |
| Accunet | -0.017*** | -0.017*** | -0.012*** | 0.006* |
| (-10.094) | (-10.047) | (-5.569) | (1.884) | |
| Annualreturn | 0.000*** | 0.000*** | -0.000 | 0.001*** |
| (6.463) | (6.454) | (-1.493) | (5.956) | |
| Maxdrawdown | -0.003*** | -0.003*** | -0.004*** | -0.003*** |
| (-54.340) | (-54.326) | (-41.116) | (-35.212) | |
| Maxreturn | -0.000*** | -0.000*** | -0.000*** | -0.001*** |
| (-25.415) | (-25.599) | (-20.556) | (-15.237) | |
| Change | -0.021*** | -0.021*** | -0.024*** | -0.016*** |
| (-20.742) | (-20.743) | (-17.887) | (-10.635) | |
| inters | -0.005*** | |||
| (-3.219) | ||||
| Constant | 0.258*** | 0.258*** | 0.166*** | 0.448*** |
| (23.313) | (23.291) | (11.478) | (24.391) | |
| TimeFE | YES | YES | YES | YES |
| N | 351904 | 351904 | 175950 | 175954 |
| R2 | 0.677 | 0.677 | 0.647 | 0.708 |
Note: Robust t-statistics in parentheses; ***, **, and * represent significance at the 1%, 5%, and 10% levels, respectively.
The moderating results of investor sentiment on fund market resilience based on opening prices.
| (1) Centralized Resilience | (2) Moderating Effect Resilience | (3) High Opening Price Resilience | (4) Low Opening Price Resilience | |
|---|---|---|---|---|
| C.Sentiment | 0.040*** | 0.040*** | 0.043*** | 0.033*** |
| (24.002) | (24.093) | (18.648) | (13.412) | |
| C.kpj | 0.012*** | 0.012*** | -0.006** | 0.262*** |
| (6.381) | (6.337) | (-2.318) | (31.756) | |
| Age | 0.001*** | 0.001*** | 0.006*** | -0.006*** |
| (2.710) | (2.668) | (13.550) | (-8.843) | |
| Cus | 0.563*** | 0.564*** | 0.147* | 1.142*** |
| (10.219) | (10.236) | (1.903) | (14.446) | |
| Mana | -0.355*** | -0.355*** | -0.259*** | -0.448*** |
| (-38.546) | (-38.537) | (-24.142) | (-26.287) | |
| Leverage | 0.068*** | 0.068*** | 0.066*** | 0.077*** |
| (50.506) | (50.528) | (34.766) | (40.219) | |
| Stockfive | -0.046*** | -0.046*** | -0.047*** | -0.049*** |
| (-44.581) | (-44.591) | (-32.940) | (-32.652) | |
| Accunet | -0.018*** | -0.017*** | -0.012*** | 0.006* |
| (-10.290) | (-10.241) | (-5.816) | (1.815) | |
| Annualreturn | 0.000*** | 0.000*** | -0.000 | 0.001*** |
| (6.458) | (6.449) | (-1.417) | (5.896) | |
| Maxdrawdown | -0.003*** | -0.003*** | -0.004*** | -0.003*** |
| (-54.334) | (-54.322) | (-41.240) | (-35.034) | |
| Maxreturn | -0.000*** | -0.000*** | -0.000*** | -0.001*** |
| (-25.480) | (-25.641) | (-20.592) | (-15.345) | |
| Change | -0.021*** | -0.021*** | -0.024*** | -0.016*** |
| (-20.613) | (-20.613) | (-17.386) | (-10.580) | |
| inters | -0.005*** | |||
| (-2.893) | ||||
| Constant | 0.258*** | 0.258*** | 0.167*** | 0.448*** |
| (23.337) | (23.316) | (11.576) | (24.393) | |
| TimeFE | YES | YES | YES | YES |
| N | 351904 | 351904 | 175950 | 175954 |
| R2 | 0.677 | 0.677 | 0.647 | 0.708 |
Note: Robust t-statistics in parentheses; ***, **, and * represent significance at the 1%, 5%, and 10% levels, respectively.
The moderating results of investor sentiment on fund market resilience based on the up-market capture ratio.
| (1) Centralized Resilience | (2) Moderating Effect Resilience | (3) High Up-capture Resilience | (4) Low Up-capture Resilience | |
|---|---|---|---|---|
| C.Sentiment | 0.040*** | 0.040*** | 0.022*** | 0.057*** |
| (24.085) | (24.073) | (9.643) | (23.863) | |
| C.upcatch | -0.005*** | -0.005*** | -0.027*** | 0.020*** |
| (-4.754) | (-4.828) | (-10.233) | (12.104) | |
| Age | 0.000 | 0.000 | 0.000 | 0.000 |
| (0.240) | (0.249) | (0.846) | (0.641) | |
| Cus | 0.569*** | 0.569*** | 0.686*** | 0.415*** |
| (10.336) | (10.322) | (9.035) | (5.285) | |
| Mana | -0.358*** | -0.358*** | -0.392*** | -0.369*** |
| (-39.116) | (-39.111) | (-30.312) | (-28.540) | |
| Leverage | 0.067*** | 0.067*** | 0.058*** | 0.066*** |
| (50.275) | (50.238) | (30.877) | (33.644) | |
| Stockfive | -0.046*** | -0.046*** | -0.042*** | -0.045*** |
| (-44.500) | (-44.445) | (-30.124) | (-29.830) | |
| Accunet | -0.009*** | -0.009*** | -0.010*** | -0.009*** |
| (-8.491) | (-8.481) | (-7.222) | (-5.807) | |
| Annualreturn | 0.000*** | 0.000*** | 0.001*** | 0.000*** |
| (6.599) | (6.599) | (6.454) | (3.933) | |
| Maxdrawdown | -0.003*** | -0.003*** | -0.004*** | -0.003*** |
| (-54.545) | (-54.562) | (-45.202) | (-31.986) | |
| Maxreturn | -0.000*** | -0.000*** | -0.000*** | -0.000*** |
| (-25.082) | (-25.083) | (-20.445) | (-17.211) | |
| Change | -0.020*** | -0.020*** | -0.023*** | -0.038*** |
| (-20.344) | (-20.352) | (-17.235) | (-21.288) | |
| inters | -0.008*** | |||
| (-5.552) | ||||
| Constant | 0.251*** | 0.251*** | 0.146*** | 0.415*** |
| (22.722) | (22.734) | (9.244) | (26.959) | |
| TimeFE | YES | YES | YES | YES |
| N | 351904 | 351904 | 175952 | 175952 |
| R2 | 0.677 | 0.677 | 0.670 | 0.686 |
Note: Robust t-statistics in parentheses; ***, **, and * represent significance at the 1%, 5%, and 10% levels, respectively.
The moderating results of investor sentiment on fund market resilience based on the down-market capture ratio.
| (1) Centralized Resilience | (2) Moderating Effect Resilience | (3) High down-capture Resilience | (4) Low down-capture Resilience | |
|---|---|---|---|---|
| C.Sentiment | 0.042*** | 0.042*** | 0.058*** | 0.029*** |
| (24.639) | (24.579) | (25.816) | (11.910) | |
| C.downcatch | 0.015*** | 0.016*** | 0.008*** | 0.024*** |
| (14.737) | (15.237) | (5.089) | (7.969) | |
| Age | 0.000 | 0.000 | 0.001* | -0.000 |
| (0.508) | (0.509) | (1.828) | (-0.983) | |
| Cus | 0.579*** | 0.578*** | 0.754*** | 0.478*** |
| (10.327) | (10.314) | (9.633) | (6.234) | |
| Mana | -0.361*** | -0.361*** | -0.215*** | -0.432*** |
| (-38.725) | (-38.724) | (-14.962) | (-35.413) | |
| Leverage | 0.065*** | 0.065*** | 0.055*** | 0.063*** |
| (47.868) | (47.843) | (29.511) | (30.868) | |
| Stockfive | -0.045*** | -0.045*** | -0.053*** | -0.037*** |
| (-42.979) | (-42.961) | (-37.923) | (-23.951) | |
| Accunet | -0.009*** | -0.009*** | -0.006*** | -0.010*** |
| (-8.194) | (-8.197) | (-4.566) | (-6.734) | |
| Annualreturn | 0.001*** | 0.001*** | 0.001*** | 0.001*** |
| (7.551) | (7.543) | (6.969) | (5.351) | |
| Maxdrawdown | -0.003*** | -0.003*** | -0.004*** | -0.003*** |
| (-52.935) | (-52.928) | (-47.324) | (-30.467) | |
| Maxreturn | -0.000*** | -0.000*** | -0.000*** | -0.000*** |
| (-25.992) | (-25.994) | (-18.638) | (-18.902) | |
| Change | -0.021*** | -0.021*** | -0.004** | -0.013*** |
| (-20.230) | (-20.190) | (-2.307) | (-9.551) | |
| inters | 0.005*** | |||
| (4.431) | ||||
| Constant | 0.242*** | 0.242*** | 0.072*** | 0.310*** |
| (21.502) | (21.522) | (4.021) | (21.571) | |
| TimeFE | YES | YES | YES | YES |
| N | 351904 | 351904 | 175952 | 175952 |
| R2 | 0.678 | 0.678 | 0.683 | 0.689 |
Note: Robust t-statistics in parentheses; ***, **, and * represent significance at the 1%, 5%, and 10% levels, respectively.
