Introduction

The stock market occupies a central role in financial markets, and the predictability of its returns has long been a focal point of academic inquiry. Fama’s (1970) efficient market hypothesis posits that in an efficient market, security prices should reflect all available information, making price changes unpredictable and driven solely by new information. However, historical evidence—from the 1637 tulip mania to the 2020 COVID-19 pandemic—suggests that markets frequently deviate from these benchmarks due to sentiment (Bollen et al., 2011). Psychological biases, such as overconfidence and herd behavior, can push investors to act on collective mood rather than fundamentals, generating volatility.

Extensive research shows that investor sentiment plays a significant role in asset pricing, with a complex and state-dependent relationship between sentiment and stock returns (Chi et al., 2012). Sentiment fluctuations are often triggered by macroeconomic announcements, geopolitical events, and corporate disclosures. Traditionally, investors relied on mainstream financial news and reports, but digital platforms have transformed forums and social media into essential channels for information exchange. These platforms now disseminate company-specific signals at high frequency and wide scale (Breza and Chandrasekhar, 2019). While social media provides immediacy and breadth, information quality is more variable than in traditional media. Balancing these sources is therefore important for understanding temporal effects on market behavior (De Long et al., 1990).

In emerging sectors like the new energy vehicle (NEV) industry, these forces are amplified. The NEV sector, at the intersection of technological innovation and environmental objectives, is characterized by policy sensitivity, rapid product cycles, and dense supply-chain linkages. Traditional media provide depth and verification; social platforms transmit higher-frequency, attention-driven signals. NEV stocks respond strongly to announcements on technology, regulation, and actions by leading firms, making the sector well-suited to study the interaction between media and market outcomes. Understanding sentiment dynamics in this setting is relevant for both academic inquiry and practice. Beyond the general asset-pricing literature, sector-specific studies provide supporting evidence consistent with the dynamics observed in the NEV market. Research on clean and renewable energy equities shows that social-media sentiment, including measures based on platforms such as Twitter, predicts returns, volatility, and trading intensity, highlighting the role of short-horizon price pressure in energy-related themes (Reboredo et al., 2018). Complementary policy-oriented analyses document that news related to government interventions in new energy industries raises volatility through heightened investor attention, reinforcing the view that price formation in these markets is highly sensitive to sentiment (Su and Wang, 2022). In addition, studies focusing on key NEV inputs demonstrate that sentiment associated with lithium and battery markets contains predictive information for related financial assets, suggesting the existence of supply-chain transmission channels relevant to NEV equities (Jeong et al., 2024). Collectively, this body of evidence provides a sectoral foundation for integrating traditional news tone with platform-based sentiment measures when analyzing stock price dynamics in NEV markets.

Existing research frequently examines the unidirectional effect of media or platform sentiment on market behavior while abstracting from their joint dynamics. Identification and data constraints help explain this pattern. Investor-sentiment proxies in earlier work are noisy and typically available only at weekly or monthly frequency, for example, survey- or index-based measures and web-search attention (Brown and Cliff, 2005; Baker and Wurgler, 2006; Da et al., 2011). Media tone, by contrast, is often observed at the daily or intraday frequency (Fang et al., 2014). The resulting frequency mismatch, combined with classical measurement error, induces attenuation and timing misalignment when both series enter jointly (Griliches and Hausman, 1986). Credible mediation further requires overlapping observation windows, explicit temporal ordering, and designs that mitigate reverse causality and bad-control bias (Imai et al., 2010). Allowing for cross-firm spillovers introduces network interference that violates independence in single-layer designs and renders conditioning on the mediator non-ignorable (Manski, 1993; Bramoullé et al., 2009). For tractability, many studies therefore estimate reduced-form direct effects and treat media sentiment as approximately exogenous. In contrast, we model investor sentiment explicitly as a mediator within a dynamic multilayer network, align variables at the daily frequency, and employ a hybrid RNN-GNN architecture to recover lagged propagation within and across layers. This framework accommodates heterogeneous delays and spillovers, and links estimates to standard causal-mediation estimands. Advances in computing and internet data have expanded financial text analysis, although reliance on single sources can limit external validity and mechanism (Zhang et al., 2024). Sector-adjacent evidence for energy and climate-related assets underscores the need to consider news tone together with platform sentiment in policy-sensitive settings such as the NEV sector (Rogmann and Schreiber, 2024).

A substantial literature measures media and investor sentiment with complementary data and methods and links these measures to asset prices, trading activity, and volatility. Studies using professionally curated news construct tone measures through dictionary approaches or supervised text classification and evaluate market reactions in daily or intraday panels and event studies. Adverse news tone is followed by lower subsequent returns, higher volatility, and greater trading activity, and patterns of coverage mitigate information frictions and help explain cross-sectional variation in expected returns and liquidity (Tetlock, 2007; Shiller, 2014; Nassar et al., 2023). A complementary strand extracts investor sentiment from social media messages and platform activity using lexicons or machine learning classifiers and documents comovement with returns, trading volume, and, in some settings, realized volatility. Shifts in the salience of prominent news redirect retail attention, alter risk preferences, and affect trading intensity (Barber and Odean 2008; Yang et al., 2023). Additional proxies that combine market data, survey responses, and web activity, including sentiment indices, survey expectations, and search intensity, covary tightly with market conditions (Baker and Wurgler, 2006; Da et al., 2011). On measurement, lexicon methods coexist with classical machine learning and deep learning models, and recent work documents divergence across news sources (Pang et al., 2002; Cambria et al., 2020). Behavioral foundations, including overconfidence, loss aversion, information asymmetry, and social interaction, provide microeconomic channels through which sentiment affects trading and volatility, including in green and ESG settings (Gervais and Odean, 2001). Despite these advances, much of the literature relies on single sources or essentially static designs. Few studies integrate news tone derived from professional media with investor sentiment derived from platforms within sectoral production networks or in policy-sensitive contexts such as the NEV sector. The firm-level interaction of the two sentiment channels and the implications for volatility and liquidity remain open questions (Luo and Zhang, 2018; Han et al., 2020).

Against this backdrop, the research question and contribution are as follows. This paper addresses the econometric challenge of modeling sentiment transmission when the underlying mechanism is both time-varying and networked. Specifically, media sentiment (measured from traditional news sources) is hypothesized to affect stock prices indirectly through investor sentiment (measured from social-media activity), with persistence, lags, and cross-firm spillovers operating within and across the media-investor-price layers. Relative to static or single-layer approaches, our contribution is to cast the problem as a dynamic network mediation task and to implement it through a multilayer, heterogeneous architecture that integrates RNN and GNN. The RNN component parameterizes the lag structure through which traditional-media tone and social-media investor sentiment affect returns; the GNN component recovers intra- and inter-layer propagation paths and produces interpretable attention and centrality measures that identify sentiment hubs and quantify spillovers. By keeping measurement sources separate (traditional news for media sentiment; social media for investor sentiment), the mediated channel is empirically distinct, enabling mechanism-level inference on how sentiment shapes volatility in a high-uncertainty, tightly interconnected NEV sector. The choice to integrate RNN and GNN follows from the economics of the problem: sentiment transmission in NEV markets exhibits temporal persistence and lagged adjustment, together with rich cross-firm and cross-layer spillovers along production and attention networks. A specification restricted to RNN would implicitly impose conditional independence across nodes and thus fail to capture contagion or spillovers; a specification restricted to GNN would collapse dynamics into contemporaneous aggregation and therefore cannot represent the transfer function of sentiment shocks over time. The hybrid architecture furnishes an estimable dynamic-network representation with interpretable mechanisms, including attention weights and centrality measures, that isolate sentiment hubs. The sample is defined ex ante as a supply-chain representative set of listed NEV firms spanning upstream key materials and components, midstream manufacturers, and downstream charging and services, which provides the cross-sectional variation required to identify network propagation and cross-layer spillovers. The analysis period runs from January 2021 to April 2024, a phase marked by policy-sensitive shifts in NEV penetration and attention that furnish variation for dynamic transmission without post-treatment window selection. Consistent with monthly information and trading cycles, we adopt a 30-day baseline transmission horizon and report robustness to shorter and longer windows.

The remainder of the paper is organized as follows. Section “Literature review” reviews the relevant theoretical and empirical literature. Section “Research design” describes the data, sentiment measures, and implementation of the multilayer heterogeneous neural network model. Section “Empirical results and analysis” presents the empirical results, focusing on the dynamic interactions and network dependencies among media sentiment, investor sentiment, and stock prices. Section “Conclusion” discusses the policy and practical implications of the findings, highlights the limitations of the study, and outlines directions for future research.

Literature review

Media news and high-frequency price discovery

A substantial empirical literature uses professionally curated news at daily or intraday horizons to study how public signals are incorporated into firm-level prices. Editor verification and precise timestamps make this approach particularly informative for short-window reactions and reduce concerns about rumor contamination. Within this strand, negative business-press tone robustly predicts lower subsequent returns in both firm-specific and market-wide settings, consistent with belief updating and limits to arbitrage (Tetlock, 2007; Peress, 2014). Complementary evidence shows that adverse tone is followed by higher realized volatility and more intensive trading, which is consistent with risk reassessment and attention shocks (Loughran and McDonald, 2011; Garcia, 2013). Cross-sectional variation in coverage breadth and salience mitigates information frictions and helps explain differences in liquidity and expected returns, underscoring the intermediary role of the business press (Fang and Peress, 2009; Bushee et al., 2010). A parallel thread documents that slant and selection in coverage can reshape investor beliefs even when fundamentals are unchanged, reinforcing the view that news should be treated as a distinct economic object (Gentzkow and Shapiro, 2010; Dougal et al., 2012). The principal strength of this literature is credible identification at high frequency; its main limitation is that firms are often analyzed in isolation, which leaves the transmission through investor beliefs implicit and pays limited attention to propagation along sectoral input-output links. This observation motivates our decision to preserve source separation by treating curated news as a distinct firm-level input and to embed media tone within an industry network where cross-firm diffusion can be estimated.

Investor sentiment and trading responses

A complementary body of work infers investor sentiment from social-media text and platform activity, thereby capturing how audiences receive and reinterpret information at high frequency. Platform mood and discussion intensity co-move with same-day returns and trading at the firm level, consistent with attention-driven order flow on the retail margin (Bollen et al., 2011; Sprenger et al., 2014). Under conditions of elevated uncertainty, platform-based sentiment aligns with realized risk, including intraday volatility and tail sensitivity (Yang et al., 2025). Recent forecasting studies explicitly incorporate sentiment into machine-learning models for asset returns: Hamadou et al. (2023), for example, show that investor-sentiment features substantially improve the ability of tree-based and neural network algorithms to forecast Bitcoin returns relative to models using only technical and macroeconomic predictors, highlighting the incremental value of sentiment for return predictability. In parallel, Hamadou et al. (2025) use a quantile–time–frequency connectedness framework to document that investor attention generates state-dependent risk spillovers between clean and dirty cryptocurrencies, with spillover intensity increasing sharply in bearish and turbulent regimes. Within energy and clean-technology markets, Liu and Hamori (2021) and related work using time-varying parameter vector autoregressive (TVP-VAR) and connectedness measures find that investor sentiment interacts with clean energy stock returns and volatility, while Twitter-based and search-based sentiment proxies for renewable energy stocks exhibit predictive content for returns and volatility, especially around major policy and market events (Song et al., 2019; Herrera et al., 2022).

High-salience items generate asymmetric buy-sell imbalances, providing a microstructure channel through which sentiment can influence prices (Barber and Odean, 2008). Recent studies also highlight the role of bot activity, echo chambers, and participation selection in amplifying mood swings, producing nonclassical, state-dependent measurement error (Cinelli et al., 2021). The primary strengths of this literature are its timeliness and granularity, whereas its limitations include linguistic noise and imperfect synchronization with contemporaneous news. For our research question, these characteristics motivate modeling investor sentiment as an explicit mediator between media tone and prices, and aligning signals at the daily frequency so that the mechanism through which media tone affects prices via investor sentiment can be empirically estimated rather than remaining implicit.

Measurement, identification, and network propagation

Methodological literature clarifies how sentiment is constructed and why static aggregation can obscure transmission. Lexicon-based indices provide transparent mappings from words to polarity but are context sensitive and can understate semantic nuances, particularly regarding domain-specific negation and uncertainty (Loughran and McDonald, 2011; Cambria et al., 2020). Classical supervised learning improves predictive performance with moderate data requirements, although it faces domain-shift risks across corpora (McCallum and Nigam, 1998; Pang et al., 2002). Deep architectures capture context and long-range dependencies, at the cost of interpretability and the need for larger labeled datasets (Kim, 2014; Devlin et al., 2019). Closely related proxies such as search intensity co-move with prices and volatility, but they are typically weekly or monthly and polarity agnostic, limiting their usefulness for firm-day mediation (Da et al., 2011; Vlastakis and Markellos, 2012). Building on these foundations, recent studies have applied deep neural networks, transformer architectures, and ensemble learning techniques to large-scale financial text and social-media corpora. These methods demonstrate that sentiment indices derived from advanced machine-learning models can significantly enhance the forecasting accuracy of stock and cryptocurrency returns and volatility. They also reveal that the magnitude of predictive improvement depends critically on the informational domain and the prevailing market regime (Correia et al., 2022; Herrera et al., 2022).

In parallel, an expanding stream of research has developed multimodal and graph-theoretic forecasting frameworks. Examples include hierarchical graph neural networks, dynamically connected network representations, and sentiment-integration architectures that explicitly incorporate information diffusion across firms. These frameworks integrate textual and behavioral sentiment indicators with evolving stock-market networks. In most existing applications, however, sentiment is incorporated merely as an external predictive factor rather than being modeled as an endogenous mediator within an economic network (Du et al., 2024). Recent evidence indicates that sentiment effects differ systematically across news categories and textual frames, suggesting heterogeneous mappings from tone to market reactions and reinforcing the case for models that preserve cross-sectional structure rather than compressing signals into a single composite index (Shapiro et al., 2022). In our framework, heterogeneity across sources, frequencies, and error structures is addressed by representing signals as dynamic propagation objects, with an RNN component parameterizing the lag structure and graph attention and centrality summarizing intra- and inter-layer diffusion.

Behavioral finance provides the microfoundations and state dependence necessary to interpret these transmission patterns. Overconfidence and loss aversion generate asymmetric responses and excess turnover, implying that seemingly similar signals can elicit different trading behavior across states (Gervais and Odean, 2001; Kahneman and Tversky, 1979). Intermediation by the business press mitigates information frictions but can also amplify reactions through diffusion and salience channels, particularly under conditions of high uncertainty (Fang and Peress, 2009; Bushee et al., 2010). When macroeconomic or policy uncertainty increases, the sensitivity of prices to sentiment rises, making amplification more pronounced (Baker et al., 2016; Fang et al., 2023). Bidirectional feedback between media and investors further magnifies market swings during periods of stress, as emotionally charged content induces larger order flow and wider spreads, while trading itself influences subsequent coverage (Zhou and Huang, 2021). These characteristics suggest that sentiment transmission is inherently networked and state dependent, reinforcing the rationale for analysis in an industry context and for robustness windows aligned with economically meaningful information and trading cycles.

A rapidly growing literature examines how information and shocks propagate along real economic links, connecting sentiment to industry networks and supply chains. Customer-supplier relationships transmit earnings and news shocks across firms, affecting returns and volatility beyond direct exposures (Cohen and Frazzini, 2008; Barrot and Sauvagnat, 2016). Spillover indices and connectedness measures quantify directional risk transmission at higher frequencies, consistent with time-varying network centrality (Diebold and Yilmaz, 2012; Baruník et al., 2016). Extending this line of research, Liu and Hamori (2021) employ a connectedness framework grounded in a TVP-VAR model to demonstrate that investor sentiment contributes to dynamic spillovers among clean energy, technology, fossil energy, and broader financial markets. Similarly, Hamadou et al. (2025) utilize a comprehensive framework integrating the quantile, time, and frequency domains to document that investor attention induces regime-dependent risk transmissions between clean and high-emission cryptocurrencies. In addition, multilayer network analyses that explicitly distinguish an investor-sentiment layer from a stock-return layer reveal economically meaningful inter-layer connectedness, even though sentiment-driven linkages are generally weaker than those observed among returns. In sectors characterized by dense input-output relations and rapid product cycles, such as the NEV sector, these mechanisms imply that sentiment shocks originating at pivotal firms can cascade through upstream and downstream nodes, generating network effects that standard single-firm designs cannot capture. This evidence motivates modeling cross-firm diffusion explicitly and employing graph-based tools that provide interpretable measures of centrality and attention.

Finally, advances in causal design emphasize why mediation and interference must be addressed explicitly when sentiment is studied at high frequency. Identification of a mediator requires joint observation, temporal ordering, and the avoidance of post-treatment conditioning, or so-called bad controls (Imai et al., 2010; Angrist and Pischke, 2009). Network interference further complicates estimation because outcomes and treatments are not independent across units connected by links; peer-effects identification strategies highlight the importance of modeling the network structure itself (Manski, 1993; Bramoullé et al., 2009). In summary, organizing the evidence by research content establishes a direct correspondence between documented gaps and our design choices. First, studies of news tone identify price effects but rarely model mediation or cross-firm diffusion; accordingly, we preserve source separation and embed media tone within an industry network to estimate diffusion paths. Second, platform-based studies measure beliefs but seldom formalize their role as the central mediator; we therefore synchronize signals at the daily frequency and explicitly incorporate that channel into the model. Third, the measurement literature documents heterogeneity across sources, frequencies, and error structures; in response, we adopt a multilayer dynamic representation that transforms heterogeneous signals into testable propagation objects, with RNN capturing lag dynamics and GNN capturing intra- and inter-layer diffusion. Finally, behavioral and uncertainty evidence implies state dependence and networked transmission; consequently, we analyze a policy-sensitive NEV supply chain and evaluate robustness over a 30-day information and trading cycle of economic relevance. Collectively, these design choices synthesize dispersed findings into an estimable and interpretable dynamic network mediation framework while avoiding unnecessary complexity. Integrating recent advances in sentiment-driven risk spillovers, multi-layer sentiment-return connectivity, and multimodal graph-based prediction, this paper develops a multilayer heterogeneous graph network. The framework simultaneously models the interactions among media tone, investor sentiment, and NEV stock price volatility within a unified dynamic network mediation structure, thereby substantially extending the emerging literature.

Research design

Sample selection and data sources

This study selects 24 publicly listed companies in the NEV sector as the subjects of analysis. The full official names, stock abbreviations, and stock codes of these firms are summarized in Table A1 of Appendix A. The selection follows an evidence-based principle that aligns with our empirical mechanism: to study within-industry sentiment transmission, the sample must span upstream key materials and components, midstream vehicle manufacturing, and downstream charging and services so that cross-firm spillovers are observable on an economically meaningful network (Cohen and Frazzini, 2008; Barrot and Sauvagnat, 2016). These companies were chosen not only for their prominence in the industry but also due to the abundance of relevant news coverage, encompassing the entire supply chain from upstream key raw materials and core components, through midstream vehicle manufacturing, to downstream charging and aftermarket services (as shown in Fig. 1). In addition, the selected firms are characterized by sustained information intensity throughout the sample window, as reflected in high volumes of CSMAR news coverage and active investor discussions on the Eastmoney Guba forum. This ensures an adequate signal-to-noise ratio for daily-frequency sentiment measures and mitigates attenuation bias resulting from sparse textual data (Vlastakis and Markellos, 2012). Furthermore, all firms maintained continuous listing and trading availability during the sample period, without experiencing prolonged suspensions. This approach sustains a balanced panel structure essential for network-based identification and prevents data gaps that could obscure the analysis of dynamic transmission mechanisms. The representativeness and diversity of the sample enhance the accuracy and reliability of the multilayer heterogeneous GNN model, thereby mitigating potential data inconsistency issues that could arise from large datasets. Finally, focusing on a compact but tier-representative set is consistent with network-propagation designs that require sufficient edge density to identify diffusion while maintaining interpretability of learned centrality and hub roles (Diebold and Yilmaz, 2012; Baruník et al., 2016). This also improves the generalizability of the research findings, providing a reference for studies on the relationship between sentiment and stock prices in other companies.

Fig. 1: Key enterprises in the electric vehicle supply chain.
Fig. 1: Key enterprises in the electric vehicle supply chain.The alternative text for this image may have been generated using AI.
Full size image

This figure illustrates the major companies involved in the electric vehicle supply chain, including key raw materials and components from upstream, vehicle assembly from midstream, and recycling and charging services from downstream.

The data sources for this study include financial news, Guba forum comments, and stock price information. The financial news data were extracted from the news database. Considering the complexity of the news content, to reduce interference from irrelevant news information, this study used keywords such as “new energy vehicle” and “electric vehicle” as indices to filter out news reports highly relevant to the research. We define media sentiment as the editorial tone embedded in professionally curated articles; it represents how public information is framed and disseminated by verified outlets. The Guba forum comment data were also sourced from Eastmoney’s Guba forum, and the data collection period spanned from January 2021 to April 2024, covering the critical period when the NEV market share first exceeded 50%, to capture market performance and enhance the scientific validity of the empirical analysis. The chosen window also coincides with policy-sensitive adoption waves and rapid product cycles in NEVs, which are precisely the conditions under which sentiment transmission and cross-firm spillovers are most likely to be salient in networked markets (Barrot and Sauvagnat, 2016). By contrast, we define investor sentiment as the opinions and affect expressed by market participants on a public platform; it reflects how audiences receive, amplify, or contest information in real time.

The construction of the sentiment measures follows a structured and source-specific procedure. Media sentiment is derived from the CSMAR news database. We begin by selecting news articles that satisfy predefined keyword filters to ensure thematic relevance. Each retained article is then mapped to one or more listed NEV firms based on explicit mentions of company names or stock tickers in either the headline or the body text. To maintain data integrity, duplicate entries are removed, publication times are standardized to Beijing time, and articles are aggregated to the firm-day level according to the trading calendar. Sentiment polarity is subsequently assigned at the article level, and firm-day media sentiment is computed as the cross-article average of these polarity scores for firm \(i\) on day \(t\), using equal weights across articles. Investor sentiment is obtained from stock-specific discussion boards on Guba, where each post or comment is inherently associated with a particular listed firm. We conduct comprehensive text preprocessing by removing boilerplate content, URLs, emojis, machine-generated near-duplicates, and extremely short tokens, and standardize timestamps to align with the Chinese trading day. Each post and comment is then assigned a polarity label, and firm-day investor sentiment is computed as the equal-weighted average of comment-level labels for the corresponding firm-day pair. Because this content is user-generated and high-frequency, it provides complementary information to curated media coverage, and the two sources are intentionally kept separate to preserve source-specific dynamics.

Sentiment classification is conducted consistently across both corpora. A manually labeled subset of the data is used to fine-tune a Chinese pre-trained BERT model for three-class sentiment classification (positive = 1, neutral = 0, negative = −1). Positive sentiment corresponds to expressions that convey optimism, favorable expectations, or confidence regarding firms or the industry. Negative sentiment encompasses language reflecting pessimism, adverse expectations, or uncertainty. Neutral sentiment refers to factual statements without discernible emotional polarity. The training corpus was manually annotated by two coders using a unified annotation standard to ensure consistency. The resulting labeled dataset was then used to fine-tune the BERT model for sentiment classification in both media and investor domains. Model performance was evaluated using standard classification metrics, including precision, recall, and F1-score, separately for media and investor sentiment. As reported in Table 1, the model achieves a precision of 0.813, a recall of 0.779, and an F1-score of 0.764 for media sentiment, and a precision of 0.745, a recall of 0.709, and an F1-score of 0.708 for investor sentiment. These metrics indicate that the fine-tuned model provides a reliable and high-quality measure of daily sentiment dynamics, which serves as a key input for the subsequent empirical analysis. All firm-day sentiment series are merged with price and control variables by stock code and trading date to construct a synchronized panel dataset for subsequent empirical analysis. In terms of data processing, rigorous data cleaning and preprocessing were conducted, removing duplicates, meaningless content, and noisy samples. Ultimately, approximately 80,000 news texts and 4 million Guba comments were obtained. Other financial indicator data were sourced from the WIND database.

Table 1 Classification accuracy of media sentiment and investor sentiment.
Full size table

Model design

This study constructs a multilayer heterogeneous GNN model aimed at systematically analyzing the complex relationships between media sentiment, investor sentiment, and stock prices. In our framework, the multilayer structure consists of three distinct but interrelated layers representing media tone, investor sentiment, and stock prices, each evolving concurrently at the daily frequency. Heterogeneity arises from the fact that node and edge types, as well as cross-layer linkages, are defined by their underlying economic functions rather than being aggregated into a single undifferentiated network. The model consists of three primary layers: the media sentiment layer, the investor sentiment layer, and the stock price layer. Nodes in each layer represent the performance of a specific company in the respective sentiment or stock price layer, while the edges denote the intercompany influences. Within each layer, edges represent economically meaningful relationships, including input-output linkages and co-attention patterns inferred from news coverage and discussion intensity, which allow shocks to propagate across firms. Across layers, directed connections encode a sequential transmission mechanism from media to investors and ultimately to prices, with all signals aligned at the daily frequency to eliminate discrepancies arising from differences in temporal resolution. To accurately capture the dynamic cross-layer effects, this study integrates GCN, GAT, and LSTM to construct a deep learning model capable of dynamically capturing and representing features across multiple layers and time scales. GCN aggregates information within each layer’s neighborhood, smoothing firm-level signals over economically grounded graphs and mitigating idiosyncratic noise in individual firm inputs. GAT assigns attention weights to edges, allowing propagation to be asymmetric and firm-specific, which enables the model to identify sentiment hubs and heterogeneous spillover patterns instead of imposing uniform influence. LSTM parameterizes temporal dependence in terms of lags and persistence within the media-investor-price system, ensuring that delayed responses and multi-day adjustments are explicitly represented rather than compressed into contemporaneous effects. The combination of these components addresses both spatial and temporal dimensions of the research question. GCN and GAT learn the structure and direction of information flows across and within firms, while LSTM learns the temporal dynamics of these flows. Relying on only one class of models would either overlook cross-firm diffusion, as in the case of LSTM alone, or collapse dynamic adjustments into a single-period aggregation, as would occur with GNN alone.

The overall framework of the multilayer heterogeneous GNN model is illustrated in Fig. 2. Appendix B presents the formal notation, including adjacency matrices for each layer, directed cross-layer mappings that characterize the sequential linkage from media to investors and then to prices, as well as the training objective that connects predicted responses in the price layer to realized outcomes. Economic interpretations are provided alongside the corresponding symbols to facilitate clarity and theoretical grounding. Details are attached in Appendix B.

Fig. 2: The overall framework of a multi-layer graph neural network.
Fig. 2: The overall framework of a multi-layer graph neural network.The alternative text for this image may have been generated using AI.
Full size image

This figure presents a structure of the multilayer heterogeneous GNN model aimed at systematically analyzing the complex relationships between media sentiment, investor sentiment, and stock prices. Panels a-e describe the specific process of network construction. Detailed explanations for each panel are in Appendix B.

Empirical results and analysis

This section provides an empirical examination of the distinct yet interrelated roles of media and investor sentiment in shaping NEV stock prices. The evidence shows that price discovery reflects the interaction between a slow-moving information environment shaped by media and rapidly adjusting investor demand. The analysis proceeds in three steps. First, we establish data validity and document key stylized facts. Second, we characterize the contrasting temporal dynamics of media and investor sentiment using dynamic and nonlinear dependence analyses. Finally, we employ a multilayer network framework to visualize and quantify sentiment transmission across firms, revealing that sector-wide volatility is largely driven by shocks propagating through a small number of central nodes in the supply chain network.

Descriptive statistics and preliminary evidence

Descriptive statistics are reported in Table A2 of Appendix A and provide a detailed overview of the temporal properties of the key variables as well as their potential implications for financial market dynamics. The mean analysis reveals pronounced heterogeneity across the three dimensions, reflecting distinct temporal patterns in each indicator. The average levels of media sentiment and investor sentiment are broadly similar, indicating a general alignment in the evolution of these two series over the sample period. This pattern is consistent with an interactive relationship between media coverage and investor expectations during information transmission. The observed daily comovement between media tone and investor sentiment accords with previous evidence showing that news-based measures and investor beliefs tend to move together over short horizons (Nassar et al., 2023). These findings motivate the use of day-level signal alignment in the subsequent analysis and support treating media and investor sentiment as distinct but jointly evolving inputs into the price formation process. Volatility patterns further distinguish the series. Stock prices exhibit substantially higher standard deviations than both media and investor sentiment, indicating more pronounced fluctuations in market prices, which is consistent with theoretical expectations. The sentiment indicators display a smoother evolution, suggesting that although sentiment fluctuations influence the market, their rate of change is relatively gradual. Unlike weekly or monthly indices that mechanically smooth variation, our daily construction reveals that this smoother profile is an intrinsic feature of the media series rather than a consequence of low sampling frequency (Sharma et al., 2024). This contrast in dispersion anticipates the empirical findings below, where investor sentiment moves rapidly and with greater amplitude, while media tone evolves more slowly.

The distributional properties of sentiment and prices deviate markedly from normality, as indicated by significant skewness and excess kurtosis. Both sentiment indicators and stock prices exhibit varying degrees of asymmetry and leptokurtosis, reflecting the presence of fat-tailed distributions. Negative skewness points to the potential for extreme downside sentiment or unusually large price declines, while kurtosis values significantly exceeding three indicate a higher frequency of extreme observations than would be expected under a Gaussian benchmark. These features underscore the importance of accounting for tail risks and extreme events, such as abrupt market crashes or sharp sentiment shifts. The Jarque-Bera test results further confirm the rejection of normality for these variables. Consequently, the subsequent analysis employs statistics and models that remain valid and informative in the presence of heavy tails and asymmetry. Stationarity of all series is confirmed by ADF tests, which reject the null hypothesis of a unit root for all sentiment indicators and stock price time series. This result ensures that subsequent dynamic analyses are not confounded by stochastic trends and aligns the empirical setting with prior studies that examine the relationship between daily news and asset returns, thereby enabling direct comparability of transmission estimates (Tetlock, 2007). From an economic perspective, the pronounced skewness and excess kurtosis indicate that large sentiment shocks occur more frequently than would be expected under a Gaussian benchmark and that such shocks can be transmitted disproportionately through mechanisms such as inventory constraints and shifts in investor attention. Establishing stationarity ensures that the observed patterns reflect within-sample fluctuations rather than deterministic trends, allowing subsequent transmission estimates to be interpreted as behaviorally meaningful. This verification concludes the descriptive stage and lays the foundation for the time-series interaction analysis that follows.

Temporal dynamics of sentiment and price interactions

Figure 3 highlights the distinct temporal dynamics of media and investor sentiment and their interactions with stock prices. The blue line, which represents media sentiment, exhibits stable and persistent fluctuations, indicating that the media acts as a slow-moving information environment with sustained influence on market expectations. Similar persistence in daily news tone has been documented in earlier studies, and in our setting, it operates on a slower horizon than investor sentiment (Kaplanski and Levy, 2010). By contrast, the green line, corresponding to investor sentiment, displays pronounced volatility, reflecting heightened sensitivity to market signals. This short-run variability aligns with platform-based evidence on rapid market reactions reported by Sprenger et al. (2014). The clear divergence between these two series enables us to separate fast-moving belief updates from slower-moving informational conditions, which is essential for understanding the timing and intensity of sentiment transmission. Stock prices exhibit close synchronization with investor sentiment, particularly during the notable downturns of early 2021 and mid 2022. The red line, which traces stock price dynamics, shows pronounced comovement with investor sentiment during periods of heightened fluctuations. In both early 2021 and mid 2022, sharp shifts in sentiment coincided with substantial price declines, underscoring the immediacy of sentiment-driven market adjustments. Such episode-level comovement has been documented in prior work around salient, attention-grabbing events. Our sectoral evidence reveals similar patterns during NEV-specific episodes (Barber and Odean, 2008). These synchronous movements provide initial evidence that short-horizon demand shifts are rapidly incorporated into prices.

Fig. 3: A time series analysis of media sentiment, investor sentiment, and stock price.
Fig. 3: A time series analysis of media sentiment, investor sentiment, and stock price.The alternative text for this image may have been generated using AI.
Full size image

This figure shows a comparative analysis of media sentiment, investor sentiment, and stock price movements over time. Each line represents one of the three variables, highlighting trends and correlations across different periods.

The price formation process is shaped by the interaction between a slowly evolving media information environment and rapidly adjusting investor demand, with notable asymmetries on the downside. Media sentiment remains relatively stable, and its effect on stock prices operates primarily through lagged channels. A 1-day delay in the price impact of news tone is consistent with evidence from daily news and return studies, such as Tetlock (2007). This mechanism is particularly salient in the NEV sector, where technological uncertainties and policy dependencies amplify the influence of media sentiment on investor sentiment, making stock prices especially sensitive. From a price formation perspective, the blue series represents the slow-moving information environment that anchors expectations, while the green series reflects short-horizon demand shifts driven by risk-bearing capacity and retail order flow. Episodes in early 2021 and mid 2022, when the green and red series move in tandem, are consistent with belief revisions being rapidly incorporated into prices through order-flow pressure. In contrast, the smoother trajectory of the blue series reflects gradual adjustments in expectations. The asymmetry in downward movements, characterized by sharper declines in the red series when the green series is negative, is consistent with tighter funding and liquidity conditions during periods of stress, resulting in larger price elasticities with respect to adverse sentiment. This asymmetric response has also been documented in the literature when negative tone intensifies (Loughran and McDonald, 2011; Garcia, 2013), and our sector-level evidence displays a similar pattern.

The following section formalizes these timing differences using a nonlinear dependence metric. Figure 4 employs mutual information to quantify nonlinear dependencies and highlights a sharp contrast in the temporal influence of media and investor sentiment. To further examine these dynamics, we construct mutual information heatmaps to characterize the interrelationships among media sentiment, investor sentiment, and stock prices across different lag structures. This approach uncovers how the predictive power of sentiment variables varies systematically with the lag period. By using mutual information rather than linear correlation, we are able to detect dependencies that conventional linear measures may overlook, thereby providing a richer characterization of sentiment transmission (Sprenger et al., 2014). The heatmaps offer a quantitative representation of the qualitative patterns documented in Fig. 3, translating temporal co-movements into a structured measure of informational linkage. Media sentiment exerts an enduring, low-intensity influence on prices, whereas investor sentiment exhibits a sharp but short-lived peak within 0–3 days. The mutual information between media sentiment and stock prices remains relatively stable across different lag periods, reinforcing the view of media as a persistent channel of information dissemination (Kaplanski and Levy, 2010). By contrast, the mutual information between investor sentiment and prices peaks over short lags before gradually declining, indicating a significant yet transitory impact. This peak aligns with the short time windows emphasized in platform-based sentiment studies, while the flatter profile of media tone is consistent with the day-ahead predictability documented by Da et al. (2011) and Yang et al. (2023). These findings are in line with the noise trader theory proposed by Mian and Sankaraguruswamy (2012), which highlights how short-term sentiment shocks amplify volatility through non-fundamental trading. Motivated by these dynamics, the empirical specification below focuses on short transmission lags at the daily frequency.

Fig. 4: Heatmap analysis of lagged mutual information between sentiment variables and stock prices.
Fig. 4: Heatmap analysis of lagged mutual information between sentiment variables and stock prices.The alternative text for this image may have been generated using AI.
Full size image

This figure displays the nonlinear interrelationships among media sentiment, investor sentiment, and stock prices across varying lag periods. The intensity of the colors indicates the strength of the mutual information at each time lag.

The timing structure supports a transmission mechanism in which investor beliefs are incorporated into prices rapidly, whereas media tone influences expectations over longer horizons. The application of multilayer network analysis reveals that, within the NEV sector, sentiment transmission pathways display a high degree of complexity and nonlinearity. In particular, during episodes associated with policy changes or technological advancements, the impact of investor sentiment on market dynamics substantially exceeds the levels predicted by standard theoretical frameworks. This indicates that, in certain industry contexts, both the trajectory and the intensity of sentiment transmission are shaped by sector-specific features and exogenous shocks, thereby extending the relevance of noise trader theory to a more heterogeneous setting. The peak in investor-price mutual information at horizons of 0–3 days reflects rapid assimilation of belief-driven order flow by liquidity suppliers, followed by a gradual decline as the information content is incorporated into prices and contrarian capital arbitrages away transitory deviations. In contrast, the flatter media-price profile is consistent with a background information environment whose effects unfold gradually through expectation formation over multiple days. The timing structure presented in Fig. 4, therefore, provides support for aligning signals at the daily frequency and for modeling a short-lag transmission mechanism. Compared with linear Granger causality tests, the nonlinear decay observed here not only yields the same ordering of horizons but also provides additional information on the shape of the dissipation of investor effects over time (Sharma et al., 2024). This evidence forms a natural bridge from temporal dynamics to the network analysis that follows.

The distinct temporal profiles of media tone and investor sentiment reconcile seemingly conflicting findings in the literature regarding the persistence of sentiment effects. Figures 3 and 4 jointly reveal that investor sentiment exerts a large but short-lived impact, whereas media tone generates lower-amplitude effects that persist over multiple days. This temporal structure motivates the alignment of sentiment signals at the daily frequency and justifies the inclusion of short transmission lags in the model specification. By distinguishing between sentiment sources and aligning them temporally, our framework integrates both short-run and persistent effects, providing a coherent explanation for why some studies emphasize transitory influences while others document persistence (Tetlock, 2007). Having established the timing dimension, we next examine the robustness of the results to different observation windows.

These findings offer new insights into the feedback mechanisms linking media sentiment, investor sentiment, and market prices. The lagged effect of media sentiment on stock prices indicates that media not only facilitates information dissemination but also plays a central role in shaping market expectations. This mechanism is consistent with behavioral asset pricing theories, which posit that psychological factors such as overconfidence, loss aversion, and herding can generate deviations from the predictions of traditional models. Our results provide additional support for the view that media sentiment acts as a catalyst for these behavioral biases, amplifying market volatility through its influence on investor sentiment. At the same time, these findings challenge classical perspectives on cognitive bias and investor behavior, such as those proposed by Tversky and Kahneman (2013). Media has traditionally been regarded as a neutral information channel. However, in the NEV sector, media sentiment not only disseminates information but also serves as an emotional amplifier that shapes investor sentiment and expectations. This dual role suggests that media can exacerbate market inefficiencies, particularly in speculative environments where sentiment amplifies overconfidence and herding, leading to pronounced price fluctuations. These insights complement existing behavioral finance theories and highlight the importance of understanding and managing the potential destabilizing effects of media sentiment on emerging markets. From a market dynamics perspective, the coexistence of a persistent information environment associated with media and a volatile demand shifter associated with investors implies that price elasticities with respect to sentiment are state dependent. During periods of heightened uncertainty, identical news shocks produce larger and more rapid price responses due to stronger amplification on the investor side. This pattern is consistent with the heightened sensitivity documented in uncertainty-focused studies, such as Kaplanski and Levy (2010).

Robustness and sensitivity analysis

Sensitivity analysis confirms the robustness of the multilayer network model and identifies a 30-day window as the optimal horizon, striking a balance between capturing immediate and lagged sentiment effects. Figure 5 shows that prediction accuracy remains consistently above 65 percent across all tested windows, although it exhibits some sensitivity to the choice of horizon. The stability of accuracy across adjacent windows indicates that the results are not driven by a knife-edge parameter choice, reinforcing the model’s reliability. The 30-day horizon corresponds to natural information and trading cycles in the NEV sector, thereby avoiding the drawbacks of both under- and over-aggregation. In financial markets, sentiment transmission typically unfolds with a lag, and shorter horizons may fail to capture these dynamics, resulting in an incomplete representation of the transmission process. Our analysis shows that a 30-day window provides an optimal balance between immediate and lagged sentiment effects, effectively encompassing both short-term market fluctuations and longer-term trends. This peak at approximately 1 month aligns closely with the monthly frequency used in established sentiment and attention indices, offering a natural benchmark for the preferred horizon (Baker and Wurgler, 2006). When the window length is shortened to 10 days, predictive accuracy declines, indicating that shorter horizons do not fully capture the influence of sentiment on prices. Expanding the window to 30 days yields the highest predictive accuracy, both validating its econometric performance and supporting its economic interpretation.

Fig. 5: Model prediction accuracy with different window sizes.
Fig. 5: Model prediction accuracy with different window sizes.The alternative text for this image may have been generated using AI.
Full size image

The figure compares the prediction accuracy of the model across different window sizes. Each bar represents the performance of the model with a specific window size, showing how the choice of window size impacts prediction accuracy.

This result reflects a fundamental trade-off. Shorter windows fail to capture delayed adjustments, whereas longer windows dilute information by incorporating noise. In the NEV data, this pattern indicates that several short-lag investor effects accumulate within approximately a monthly horizon, while horizons that extend beyond this point weaken transmission by blending new information with stale signals. The 30-day window aligns closely with typical monthly investment cycles and patterns of information dissemination, thereby enhancing the model’s ability to capture the behavioral dynamics of market participants. Extending the window beyond 30 days does not yield additional predictive gains and may even reduce accuracy due to informational dilution over time, which obscures the immediate impact of sentiment variables on prices. Consequently, the 30-day window emerges as the most effective choice, ensuring that the model reflects the interaction between sentiment and prices with both reliability and interpretability. Figure 5 summarizes this economic trade-off: shorter horizons miss sequential transmission, while excessively long horizons mix new signals with mean reversion. The peak accuracy observed at 30 days is consistent with established patterns in the literature on information flows and investor attention (Vlastakis and Markellos, 2012), rather than reflecting ad hoc tuning.

To assess the incremental contribution of jointly modeling temporal and relational structure, we benchmark the proposed architecture against classical time-series models, including VAR and ARIMA, and against ablated neural-network baselines that isolate a single component. For each window size and model specification, we repeat the forecasting experiment multiple times with different random seeds. For every combination, we compute the mean and standard deviation of Accuracy, MSE, and MAE; the corresponding values are reported in Table 2, which complements Fig. 5 by providing the exact numerical performance of all models across prediction windows. The ablations consist of an LSTM that excludes graph structure and a single-layer GNN that omits temporal dynamics. At the 30-day horizon, the proposed model attains the highest directional accuracy (0.690), compared with 0.639 for the LSTM baseline and 0.660 for the GNN baseline, implying gains of approximately 5.1 and 3.0 percentage points, respectively. Relative to the linear benchmarks, accuracy is 0.561 for ARIMA and 0.526 for VAR, so the proposed specification improves performance by about 12.9 and 16.5 percentage points at the same horizon. Dispersion across repeated evaluations is modest at this window (standard deviations of 0.003 for the proposed model, 0.007 for the LSTM, and 0.006 for the GNN), indicating that the ranking of models is not driven by estimation noise. The ablation results, therefore, show that removing either the temporal component or the cross-firm relational component materially degrades predictive accuracy, while linear benchmarks that ignore nonlinearities and cross-sectional spillovers perform substantially worse. Taken together, these findings support the interpretation that media- and investor-sentiment signals propagate both intertemporally and across firm networks, and that explicitly modeling both channels yields economically meaningful and robust predictive gains.

Table 2 Predictive performance by model and window size.
Full size table

Multi-layer network analysis: unveiling the transmission mechanism

To examine the mechanisms underlying information transmission between media sentiment, investor sentiment, and stock prices, we develop a multi-layer heterogeneous graph network model. This framework captures the complex, nonlinear dynamics of sentiment propagation across media, investors, and price layers. Traditional single-layer network models are limited in their ability to address static or linear relationships and are therefore inadequate for characterizing the nonlinear and cross-layer interactions that shape sentiment dynamics. In contrast, the multi-layer graph network approach analyzes transmission pathways and feedback loops across layers, providing a more accurate and comprehensive representation of market behavior. This design enables us to attribute observed price dynamics to specific informational channels within the system. Node centrality, network density, and inter-layer coupling strength provide economically meaningful metrics for characterizing both the origins of sentiment signals and the system’s capacity to propagate them. Node centrality captures the influence of specific sentiment sources, such as media or investor nodes, within the transmission network. Highly central nodes function as hubs that shape the sentiment dynamics of other nodes, amplifying their impact on the broader system. Network density measures the overall connectivity of sentiment transmission pathways; higher density reflects tighter interconnections among market participants and is often associated with greater market sensitivity and volatility. Inter-layer coupling strength quantifies the degree of interaction between different sentiment layers, clarifying the extent to which media sentiment affects market behavior through its influence on investor sentiment. Compared with coverage-based measures of information frictions (Bushee et al., 2010), these network metrics offer a direct framework for identifying where signals originate and how effectively they propagate through the market. Each metric has a clear economic interpretation. Centrality reflects the potential for firm-specific signals to shape sector-wide beliefs. Density indicates the capacity of shocks to spread efficiently across firms. Inter-layer coupling captures the strength with which belief updates are translated into price revisions, linking informational shifts to observed market outcomes.

Figure 6 highlights the hierarchical structure of information flows, consisting of a sparse media layer, a contagion-prone investor layer, and a densely integrated price layer. Panel (a) displays the internal network of media sentiment, where nodes represent different sentiment sources and edge thickness reflects the intensity of influence among them. The network is relatively sparse, indicating that information propagation is concentrated around a few highly influential nodes. This concentration at hub firms amplifies the coverage-impact gradient documented at the firm level by Zhou and Huang (2024), underscoring the disproportionate role of central firms in shaping market reactions. Within this structure, NDS, BYD, and HYG occupy central positions, reflecting their dominant influence in the media sentiment layer. Coverage by these key outlets substantially magnifies investor sentiment, which in turn affects stock prices. The sparsity of the source layer implies that a small set of firms effectively governs the sector-wide flow of information. Sentiment contagion is most pronounced within the investor layer, where central firms such as NDS and BYD can trigger sector-wide belief revisions. Figure 6b depicts the internal structure of the investor sentiment network, revealing a clear network effect. A small number of nodes, including NDS, BYD, and GQJ, exhibit high centrality and exert substantial influence on overall sentiment dynamics. In contrast to market-wide comovement, the supply chain setting indicates that contagion is strongest between economically linked firms, underscoring the importance of network position (Fang and Peress, 2009). These dynamics become particularly salient during periods of heightened uncertainty or major informational shocks, when sentiment contagion amplifies market reactions, synchronizes sentiment fluctuations, and increases price volatility. This pattern suggests that systemic risk in the industry is shaped not only by fundamentals but also by collective shifts in investor sentiment (Wang and Li, 2022). Accordingly, the investor layer functions as the primary amplification channel for shocks originating from media sources.

Fig. 6: Structure and weights of intra-layer networks.
Fig. 6: Structure and weights of intra-layer networks.The alternative text for this image may have been generated using AI.
Full size image

This figure visualizes the structure of the intra-layer networks within the multi-layer graph neural network, panel a shows the network structure for media sentiment, panel b is for investor sentiment and panel c is for stock price. The weights of the connections between nodes are depicted, illustrating how information flows through the network at each layer.

The dense connectivity of the price layer indicates that shocks are transmitted rapidly across firms, with hub companies such as NDS and BYD exhibiting particularly high synchronicity. Within the investor sentiment network, NDS, BYD, and GQJ also occupy central positions, underscoring their pivotal roles in the NEV supply chain. These firms combine substantial market capitalization with high visibility, and sentiment shifts concerning them often have disproportionate effects on other market participants. For example, NDS’s dominance in the battery sector implies that changes in expectations for this firm can quickly propagate to both upstream suppliers and downstream manufacturers, exerting broad downward pressure on sectoral stock prices. Figure 6c illustrates the internal structure of the price network, which is more intricate and densely connected than the networks shown in earlier figures, reflecting a high degree of interdependence among individual stocks. This structure is consistent with the presence of common exposures. A key contribution of the analysis lies in identifying the sources of sentiment signals in the upper layers, thereby establishing a direct link between the information and pricing layers. The complex web of price co-movements suggests that shocks to individual firms can propagate swiftly through market mechanisms, amplifying volatility across the network. Given the collective behavior of investors, few firms are insulated from the systemic risks induced by broader market fluctuations. This configuration helps explain why volatility tends to cluster around firms that simultaneously dominate media coverage, investor attention, and network centrality. The coexistence of sparse information origins and dense propagation highlights the central role of hub firms in shaping sectoral dynamics. Stock prices of key companies such as NDS and BYD exhibit a high degree of synchronicity, reflecting strong integration and interdependence within the industry. In the NEV sector, firms are often embedded in tightly linked supply chains, so that shocks to the stock price of a central company can trigger corresponding movements in the prices of its partners. These resonance effects, transmitted through supply chain and exposure linkages, amplify firm-specific shocks into sector-wide fluctuations. The contrast between a sparse media layer and a dense investor-price layer further reveals an information structure in which a small set of sources generates signals that are subsequently magnified through investor interactions and incorporated into market prices. Monitoring the centrality of major hub firms, including leading battery producers and OEMs, provides critical information about sectoral vulnerability to sentiment-driven shocks. Consequently, network-based statistics offer useful early-warning indicators of potential instability at the sector level.

To visualize the structural relationships among nodes across different layers more intuitively, we construct a multilayer message-passing Sankey framework that illustrates how media reports affect stock prices through the investor sentiment channel. This visualization clarifies the roles of central firms within the NEV supply chain and provides empirical foundations for subsequent econometric analysis. Path thickness provides a visual representation of transmission intensity, allowing strong belief-updating channels to be distinguished from weaker, noisier links. Within the GNN framework, GAT is employed to examine the heterogeneous effects of sentiment on stock prices across firms. Variations in path thickness and color capture differences in the strength of sentiment transmission, highlighting how information propagates through the network.

Sentiment flows on the media layer are highly concentrated in a few nodes, most notably Media:BQL and Media:BYD, a pattern consistent with information production costs and editorial curation. Dominant outlets generate correlated signals across firms even when fundamentals are not fully shared, which raises the marginal value of monitoring core sources and fosters belief convergence around common news. This observation accords with prior evidence on media coverage, investor attention, and commonality in returns. In our three-layer visualization, the path structure shows that these sources bear the bulk of upstream shocks, providing a clear structural rationale for short-horizon comovement within the sector. The investor layer exhibits a bottleneck centered on PTL, with sizable connectivity also for BQL and BYD. This structure is consistent with limits to risk bearing and attention intermediation. When a small set of investor hubs absorbs and retransmits incoming signals, order-flow imbalances concentrate locally, and price pressure is largest along hub-routed paths. A direct empirical implication is cross-sectional dispersion in marginal price elasticity by path thickness: paths that traverse Investor:PTL should generate larger same-day and next-day returns, higher turnover, and stronger order imbalance than paths that rely on low-degree intermediaries. This prediction can be tested by regressing stock-level returns and trading activity on path-level measures such as cumulative thickness and interlayer coupling to the investor hub. On the price layer, the strongest aggregation occurs at TRD, with substantial but less concentrated inflows to NDS, PTL, and BYD, and secondary destinations including GQJ, HYG, and GXG. This pattern is consistent with common exposures and supply-chain or product-market linkages that facilitate pass-through, as in Cohen and Frazzini and related work on customer-supplier propagation and product similarity. The figure implies that media shocks originating at BQL or BYD and routed through Investor:PTL should be capitalized most strongly by TRD and by stocks with high coupling to the investor hub. A standard event-study design around salient media shocks at these sources should yield faster incorporation and larger magnitudes along the hub-routed cascades than along thin paths.

Heterogeneity in path thickness has direct economic content. Thick paths identify channels with high belief-updating intensity and limited immediate arbitrage, whereas thin paths indicate either weak information content or offsetting contrarian capital. The multilayer structure also rationalizes state dependence. When uncertainty or funding constraints are elevated, attention becomes more concentrated at investor hubs, and the marginal capacity of liquidity suppliers declines. In that state, an identical media shock produces larger transmission through Investor:PTL and sharper price adjustments at the main aggregation nodes. This mechanism complements short-lag timing evidence by linking state variables to network-amplified pass-through. The visualization further guides validation. Ablations that remove edges incident to Investor:PTL or that down-weight Media:BQL and Media:BYD should reduce mutual-information gains and out-of-sample accuracy more than analogous removals at peripheral nodes. Cross-sectional tests that interact path measures with proxies for uncertainty or funding conditions should deliver larger coefficients in stressed states, consistent with amplification. Together, these results map path geometry into economically meaningful predictions about where and when sentiment shocks translate into price revisions and trading intensity.

The evidence collectively supports a unified economic narrative in which persistent media information interacts with volatile investor demand within a hub-dominated network, giving rise to volatility clustering in the NEV sector. Figures 3–7 illustrate that this interaction operates through identifiable hubs and short-lag dynamics, linking firm-level sentiment shocks to sector-wide fluctuations. In contrast to studies that treat sentiment as an independent predictor (Da et al., 2011), our results demonstrate that the strength of transmission depends systematically on a firm’s network position and the thickness of specific transmission paths. This set of findings completes the empirical argument and provides the foundation for the subsequent policy discussion. Collectively, the evidence supports a narrative of persistent media information interacting with volatile investor demand on a hub-dominated network, explaining volatility clustering in the NEV sector. Taken together, the evidence from Figs. 3–7 supports a coherent economic narrative. A persistent information environment shaped by media interacts with a volatile demand shifter represented by investors on a network characterized by identifiable hubs. Short-lag effects, combined with hub-driven diffusion, help explain why volatility clusters around key firms and policy-sensitive intervals within the NEV sector. Relative to studies that treat sentiment as an independent predictor (Da et al., 2011), our results indicate that the strength of transmission depends on a firm’s position within the network and on the intensity of specific transmission paths in the data. This completes the empirical argument and sets up the policy discussion in the concluding section.

Fig. 7: Multi-layer network message passing: a Sankey diagram.
Fig. 7: Multi-layer network message passing: a Sankey diagram.The alternative text for this image may have been generated using AI.
Full size image

A Sankey diagram demonstrating the message passing process in the multi-layer network. It shows the flow of information between layers, with node connections representing the transfer of data and their respective weights.

Conclusion

This study employs a multilayer graph network model to examine how media sentiment and investor sentiment jointly shape stock price dynamics in the NEV sector. By combining recurrent and graph components, the framework captures both short-horizon temporal dependence and cross-firm network propagation. The empirical results reveal several robust patterns. Media tone evolves slowly and is highly persistent, whereas investor sentiment is volatile and exerts its strongest influence on prices within zero to three trading days. A 30-day window delivers the best out-of-sample performance among the horizons considered, indicating that shorter windows miss delayed effects while longer windows dilute informative signals. The estimated network is asymmetric: links are relatively sparse in the media layer, and denser in the investor and price layers, and transmission is concentrated in a small set of central firms. Price declines are steeper when investor sentiment turns negative, which is consistent with state-dependent liquidity conditions during periods of stress. Together, these findings support an interpretation in which slowly evolving media information shapes expectations over multiple days, rapid sentiment fluctuations generate transient pressures that are subsequently absorbed, and cross-firm connections aggregate firm-level signals into sector-wide movements when attention is focused on hub firms.

The results have practical implications for disclosure, market surveillance, and investor protection. Time bounded disclosure and rumor clarification can help ensure that persistent information is updated promptly and that short-run sentiment shocks are not amplified by uncertainty. For firms with high network centrality, exchanges and regulators may request timely responses to market-moving news and encourage predictable disclosure calendars. Market surveillance can concentrate on the first 3 trading days following salient events, integrating daily sentiment indicators, order imbalance, and realized volatility, and activating calibrated stabilizing measures only when well-specified thresholds are exceeded. Supervisory practice can align monitoring horizons with the 30-day window that performs best in prediction, and investor protection tools can be tailored to short lag dynamics, for example, through targeted risk alerts and simple guidance that reduce herding during the most elastic phase of sentiment adjustment.

The empirical analysis is based on China’s NEV market from 2021 to 2024, so external validity requires caution. The framework can be extended to other markets and industries when high-frequency measures of media tone and investor sentiment are available, an economically meaningful firm-to-firm network can be estimated, and disclosure rules and trading microstructure permit short lag transmission to be observed. These features imply that our conclusions should be interpreted as conditional on the Chinese NEV context and on the particular sentiment measures and network specification employed. Applying the framework in new settings involves re-estimating the relevant network, recalibrating the transmission horizon through out-of-sample validation rather than imposing a 30-day window, and adapting text models to local languages and platform conventions. Future research can extend this work in two directions. On the empirical side, applying the framework to other sectors and institutional environments would allow systematic assessment of generalizability and provide evidence on how disclosure regimes and microstructure features shape sentiment propagation. On the methodological side, advances in dynamic network modeling that improve robustness to noise and capture heterogeneous, state-dependent transmission mechanisms would deepen understanding of sentiment-driven price dynamics and broaden the applicability of the approach beyond the NEV sector.