Artificial Intelligence-Enhanced Network Modelling of ESG Risk in Global Supply Chains

1. Introduction

ESG risks are progressively understood as outcomes of system-wide interdependencies within global supply chains rather than as isolated firm-level deficiencies. Ecological harm, labour violations, and governance failures often originate outside focal firms and spread through production networks via reputational transmission, regulatory contagion, and operational disruption [1]. As a result, a firm’s ESG exposure depends not only on its internal practices but also on its relational position within networks of suppliers and customers. This perspective challenges conventional ESG assessment approaches that rely on static disclosures, aggregated ratings, and backwards-looking indicators. Such tools provide limited insight into how ESG risks accumulate and propagate across interconnected production systems, constraining their ability to detect emerging vulnerabilities before they materialise into noticeable events [2].

However, this study adopts a network-oriented approach and addresses two main questions. How does a firm’s position within supply-chain networks shape its ESG risk exposure? Does explicitly incorporating inter-firm network structure improve the accuracy and reliability of ESG risk prediction relative to firm-level models? Hence, recent advances in data availability and artificial intelligence enable these questions to be addressed more directly. Detailed supplier–customer connections, shipment-level trade data, and continuously updated news streams allow ESG exposure to be analysed as a dynamic and relational process [3]. At the same time, progress in network analytics, machine learning, and natural language processing strengthens the modelling of nonlinear interactions, inter-firm dependence, and early-stage risk signals [4].

Despite these developments, empirical ESG research remains fragmented, with network structure, ESG metrics, controversy records, and media sentiment typically examined in isolation. This study proposes an AI-network-oriented framework for analysing ESG risk in global supply chains. Thus, by integrating network econometric methods, graph-based learning, and transformer-based sentiment extraction, the framework captures the exposure, diffusion, and forecasting ability of ESG risk.

The study contributes by reframing ESG risk as a network-contingent phenomenon, demonstrating the predictive value of ESG-related media sentiment, and showing that models incorporating inter-firm network structure outperform traditional approaches in forecasting ESG risk. The remainder of the paper is organised as follows. Section 2 reviews the related literature, Section 3 presents the methodological approach, Section 4 discusses the results, Section 5 provides an in-depth discussion and outlines key policy implications, and Section 6 concludes the study.

2. Related Literature

2.1. The Concept of Network Modelling for ESG Risk

This study reviewed several empirical studies that focus on network modelling of ESG risk propagation across global supply chains in production and service systems, as shown in Table 1. The evidence shifts ESG risk from a purely firm-centric singularity to a relational and systemic exposure that can diffuse through inter-firm dependencies. A dominant stream uses industrial network data to model ESG spillovers at scale. Wei et al. [5] studied China’s industrial chain and reported that ESG performance shocks diffuse along supply chain links, with measurable impacts on downstream profitability and value-chain outcomes. Extending this logic, Tan et al. [6] introduce a directed graph neural network (GNN) that distinguishes upstream from downstream flows, demonstrating asymmetric propagation patterns that reveal where vulnerabilities accumulate and how network resilience is shaped by directionality. However, these studies provide strong empirical support for contagion-like ESG dynamics in production networks, illustrating the practical value of network-aware AI in identifying systemic exposure beyond a focal firm’s internal practices.

Bergier [7] reveals that in Brazil’s traceable beef export system, ESG risks such as deforestation can be embedded in indirect suppliers and remain invisible without relational infrastructures that connect ranches, intermediaries, and processors. This evidence supports a key limitation in conventional ESG monitoring. Angioni et al. [8] construct ESG knowledge graphs from news using NLP pipelines and demonstrate how ESG narratives evolve, providing a foundation for reputational monitoring and early warning systems. Brockmann et al. [9] focus on supply-chain link prediction under uncertainty using knowledge graphs extracted from web data, improving visibility across hidden tiers. Their approach is decision-relevant because accurate link inference enables earlier identification of where ESG exposure may reside, especially when supply-chain information is incomplete or noisy. To conclude, Cheng et al. [10] propose a large language model (LLM)-driven framework for schema induction and knowledge-graph construction in EV battery supply chains, illustrating how zero-shot and weakly supervised extraction can support disruption forecasting and ESG oversight in critical mineral contexts where supplier ecosystems are complex and rapidly evolving.

2.2. Hypothesis

2.2.1. ESG Risk as a Networked Phenomenon

Early ESG research conceptualised sustainability risk as an internal firm attribute, with a main focus on disclosure quality, ratings, and financial materiality [11,12]. This perspective underestimates the extent to which ESG risks are shaped by inter-organisational dependencies. Empirical evidence has shown that ESG incidents often originate from business partners and propagate through supply chain relationships [13]. Network-based literature suggests that ESG risk clusters among connected firms and that intermediaries or highly central actors face disproportionate exposure [14]. These findings align with theories of contagion and spillovers in economic networks, in which shocks diffuse along relational ties rather than remaining confined to originating nodes [15]. However, existing literature relies on static network representations or limited institutional settings, leaving open questions regarding temporal dynamics and predictive relevance. Based on the reviewed empirical studies, this study proposes the first hypothesis:

H1. 

Firms within supply-chain networks exhibit higher ESG risk exposure than other firms.

2.2.2. AI Network Modelling in ESG Analysis

Artificial intelligence has recently been applied to ESG and supply chain research, particularly for prediction, anomaly detection, and risk classification [16]. Hence, machine learning models capture nonlinear relationships, and graph neural networks explicitly encode relational dependence and multi-hop interactions [17]. Evidence from financial and supply chain networks suggests that network-aware models outperform firm-level approaches when outcomes are shaped by interdependence [18]. However, existing AI-based ESG studies often prioritise classification accuracy without sufficient attention to probability calibration, interpretability, or integration with ESG theory [19]. Moreover, ESG performance indicators, controversy metrics, and network structure is rarely modelled jointly, limiting insights into how relational exposure interacts with firm behaviour. Based on the reviewed research, this study proposes the second hypothesis.

H1. 

Firms within supply-chain networks exhibit higher ESG risk exposure than otherfirms.

2.3. Research gap

This study is motivated by gaps in the existing literature. Much of the industrial chain research focuses on a single national context and relatively stable institutional environment, which raises concerns about external validity when applied to global, multi-country supply chains operating under heterogeneous regulatory regimes. Moreover, existing studies often rely on comparatively static network snapshots, which may understate temporal rewiring, supplier substitution, and shock persistence within dynamic production and service ecosystems. In addition, prior literature demonstrates the separate value of network structure, such as directed graph neural networks (GNNs), and text-driven ESG signals, including news-based knowledge graphs and large language model (LLM) extraction. However, few studies explicitly integrate both data types into a unified predictive framework to assess whether combining relational supply chain structure with adverse ESG sentiment yields superior predictive accuracy and reliability. Addressing these limitations, this study builds on the empirical foundations presented in Table 4 to develop and evaluate network-aware AI models that jointly leverage supply-chain topology and ESG news signals to enhance the prediction of ESG risk and its propagation across global supply chains.

3. Research Design

3.1. Multi-Method Research Approach

To address the multifaceted nature of the ESG risk modelling paradigm, this study employs a structured multi-model framework. Traditional econometric specifications, including OLS, Poisson, negative binomial, and PPML, are used to verify the consistency of estimated network effects under different distributional settings and to address overdispersion and zero inflation [20]. Machine learning, such as Random Forest, Gradient Boosting, and XGBoost, is incorporated to improve the detection of relatively rare incidents and to accommodate nonlinear relationships that linear models may fail to capture [21]. Model performance is evaluated using both discrimination and calibration criteria, ensuring that predictive accuracy and probability reliability are jointly assessed. While SHAP analysis supports transparent interpretation of complex models. In addition, logit and survival specifications capture the timing and likelihood of ESG incidents. This design enhances robustness and delivers a balanced evaluation of ESG risk in networked production systems [22].

This study integrate data from different authoritative sources to capture supply-chain structures, cross-border trade activity, and ESG-related dynamics, as shown in Table 2. FactSet Revere provides firm identifiers, industry classifications, and validated supplier-customer relationships obtained through authenticated Application Programming Interface (API) queries. Panjiva (S&P) shipment data provide detailed logistics information, which is analysed using unsupervised anomaly-detected techniques, including autoencoder neural networks and Isolated Forest algorithms, to identify irregular trade patterns. Text-based ESG information is compiled from Global Database of Events, Language, and Tone (GDELT) using Google BigQuery filters, together with RepRisk incident data. These text sources are processed using transformer-based natural language processing models, including FinBERT, RoBERTa, and mBERT, to classify ESG-relevant content and track sentiment shifts. Country-level institutional conditions are captured using the Worldwide Governance Indicators (WGI). The sample periods are selected to ensure consistent reporting coverage and comparability over time.

3.2.. Data Volume

The datasets (Figure 1 and Figure 2) used in this study differ in scale and function, with GDELT and Panjiva serving as the core sources for ESG analysis. The GDELT Event database records tens of millions of news events annually, averaging approximately 63 million observations per year, which enables broad coverage of ESG-related narratives, sentiment, and reported incidents across countries and sectors [22,23]. Panjiva (S&P Global) provides shipment-level trade data at a million-record scale, with an average of roughly 25 million records per year, offering detailed evidence on cross-border trade flows and firm-level supply chain connections [24]. Table A3 in the Appendix presents the country’s most frequently associated with ESG-related events in the study sample, while Table A4 summaries the distribution of events across major media sources, illustrating both the geographical reach and the fragmented nature of global news coverage. The ten most frequently observed media sources account for 7.9% of all recorded events, whereas the remaining 92.1% are distributed across other outlets. GDELT and Panjiva are emphasised because their volume and granularity jointly support the examination of ESG information signals and supply-chain network structure; while the remaining datasets are primarily used for validation, benchmarking, and institutional context.

3.3. Sample Size Selection Procedure

The study sample was derived through a sequential filtering process designed to preserve firm traceability, network connectivity, and longitudinal coverage, which are important for supply-chain-based ESG analysis. The selection process is illustrated in Figure 3. Initially, the study combined 110,100 entity records from FactSet Revere, Panjiva, GDELT, and RepRisk. Due to differences in database structures and reporting units, firm identities were standardised using entity-matching procedures, which eliminated 42,300 duplicated or unresolved records and resulting in 62,500 firms with unique identifiers. Subsequent screening focused on data quality and network relevance.

Firms were removed if they had unstable identifiers (6,200), incomplete firm-level attributes (7,800), or unverifiable supply-chain relationships (7,500). This step yielded 41,000 firms with reliable identifiers and a basic relational structure. To ensure suitability for network-oriented ESG modeling, the study retained only firms with at least one confirmed supply-chain connection, observable ESG-related signals, and sufficient temporal information. This stage excluded 15,500 firms, resulting in a final analytical sample of 25,500 firms. Table 3 highlights the regional scope of the sample, capturing both major economic hubs and emerging markets with diverse regulatory contexts. For the retained firms, transaction-level and event-based observations were collected. Panjiva data produced approximately 2.5 million shipment observations, reflecting repeated supplier-customer interactions, while GDELT, yielded approximately 1.2 million ESG-relevant news events that met relevance and duplication criteria.

3.4. Network-Based ESG Dataset Structure

Table 4 outlines the construction of the network-based ESG dataset. Verified supply-chain relationships are translated into directed, weighted adjacency matrices, where link direction indicates product flows and weights reflect shipment intensity. Hence, all core results are robust to both binary and shipment -weighted adjacency indicators, yielding qualitatively identical coefficients. Graph computations are conducted using NetworkX and related libraries to generate firm-level structural attributes, including degree, betweenness, eigenvector centrality, and clustering measures.

Empirical evidence indicates that degree centrality capture ESG exposure, betweenness centrality identifies firms that transmit risk across supply chains, eigenvector centrality reflects amplified risk arising from influential partners, and clustering coefficients capture local connection of ESG incidents [26]. These indicators differentiate firm by connectivity, brokerages roles, and structural influence within the network. Network matrices are then integrating with yearly ESG incident counts, sentiment measures produced by transformer-based language model, anomaly scores capturing irregular trade behaviour, and country-level institutional indicators from the Worldwide Governance Indicators. Figure 4 shows the study dataset’s network-based patterns, it illustrates an adverse ESG event emerging upstream at $t₀$ and propagating downstream through direct and indirect supply-chain links. $\mathrm{F}$ donates the firm under analysis, while $\mathrm{B}$ represents a strategically positioned upstream supplier. Their high betweenness centrality indicates brokerage roles that channel indirect ESG exposure across connected firms and clustered production structures.

3.5. Measurement of Variables

Table 5 outlines the operational definitions and measurement approach for all study variables, which are grouped into dependent, independent, and control categories.

3.6. A Multilayer Network Econometric and AI Framework for ESG Modelling

The study adopts a multilayered analytical framework that combines network econometrics, supervised and unsupervised machine learning, graph-based models, anomaly detection methods, and natural language processing to capture the multifaceted nature of ESG risk in global supply chains. Each analytical layer is anchored in a clear statistical or algorithmic formulation, ensuring interpretability, internal coherence, and methodological rigour.

1.

Network-Based Baseline Model

$$R_{it}=\alpha +\rho (W{R}_t{)}_i+{\sum }_k{\beta }_k{C}_{k,i,t-1}+{\gamma }^{\mathrm{\text{'}}}{X}_{i,t-1}+{\mu }_i+{\tau }_t+{\epsilon }_{it},$$

(1)

$$R_{it}=\alpha +\rho (W{R}_t{)}_i+{\beta }_1{\mathrm{Deg}}_{i,t-1}+{\beta }_2{\mathrm{Betw}}_{i,t-1}+{\beta }_3{\mathrm{Eigen}}_{i,t-1}+$$

$${\beta }_4{\mathrm{Clust}}_{i,t-1}+{\gamma }^{\mathrm{\text{'}}}{X}_{i,t-1}+{\mu }_i+{\tau }_t+{\epsilon }_{it}.$$

(2)

Equations (1) and (2) specify a baseline model in which firm-level ESG risk $R_{it}$ is expressed as a function of network exposure $\left(W{R}_t{)}_i\right.$, lagged centrality measures $C_{k,i,t-1}$, and firm controls $X_{i,t-1}$, with firm and time fixed effects, capturing structural dependence of ESG risk across supply chain networks [27].

2.

Supervised and Graph-Based Learning

$$\mathrm{P}\mathrm{r}(y_{it}=1\mid X_{it})=\mathsf{\Lambda }({\beta }_0+{\beta }^{\mathrm{\text{'}}}{X}_{it}+{\mu }_i+{\tau }_t),$$

(3)

$${\mathrm{x}}_{it}=({\mathrm{ShipIrreg}}_{it},\hspace{0.25em} {\mathrm{SentNeg}}_{it},\hspace{0.25em} {\mathrm{NetMet}}_{it},\hspace{0.25em} {\mathrm{ESGHist}}_{it},\hspace{0.25em} {\mathrm{RegGov}}_{it})$$

(4)

The supervised ESG risk prediction model uses historical firm, network, and news data to predict future ESG outcomes. Graph-based models, particularly graph neural networks, are employed to represent inter-firm dependence and indirect exposure arising from multi-tier supply-chain linkages. Equation (3) estimates ESG event probability ($Pr$) based on $X_{it}$= shipment irregularities, sentiment, lagged incidents, network metrics, governance controls, and firm ($i)$, and time$\left(t\right)$ effects, providing a transparent benchmark for supervised learning [28].

$$y_{it}=\alpha +\rho (W{R}_t{)}_i+{\beta }^{\mathrm{\text{'}}}{X}_{it}+{\mu }_i+{\tau }_t+{\epsilon }_{it},$$

(5)

The graph-based model in equation (5) explicitly captures inter-firm dependence by incorporating a shipment-weighted network-lag term $\left(W{R}_t{)}_i\right.$. Equations (6)– (8) show Graph Neural Networks (GNNs), which generalise this spatial dependence in a nonlinear and high-dimensional framework. In Graph Attention Networks (GANs), neighbour information is aggregated through attention-weighted mechanisms [29].

a.

Attention coefficient (importance of neighbor $j$ for $i$):

$$e_{ij}^{\left(\mathcal{l}\right)}=\mathrm{LeakyReLU}\left({\mathrm{a}}^{\left(\mathcal{l}\right)\mathrm{\text{'}}}\right[W^{\left(\mathcal{l}\right)}{\mathrm{h}}_i^{\left(\mathcal{l}\right)}\hspace{0.25em} \parallel \hspace{0.25em} W^{\left(\mathcal{l}\right)}{\mathrm{h}}_j^{\left(\mathcal{l}\right)}\left]\right),$$

(6)

$${\alpha }_{ij}^{\left(\mathcal{l}\right)}={\displaystyle \frac{\mathrm{e}\mathrm{x}\mathrm{p}\left(e_{ij}^{\left(\mathcal{l}\right)}\right)}{\sum _{k\in \mathcal{N}\left(i\right)}\mathrm{e}\mathrm{x}\mathrm{p}\left(e_{ik}^{\left(\mathcal{l}\right)}\right)}},$$

(7)

b.

Attention-based message aggregation:

$${\mathrm{h}}_i^{(\mathcal{l}+1)}=\sigma ({\sum }_{j\in \mathcal{N}\left(i\right)}{\alpha }_{ij}^{\left(\mathcal{l}\right)}{W}^{\left(\mathcal{l}\right)}{\mathrm{h}}_j^{\left(\mathcal{l}\right)}).$$

(8)

Here, ${\alpha }_{ij}^{\left(\mathcal{l}\right)}$is an interpretable weight that tells you how strongly supplier $j$’s characteristics influence firm $i$’s ESG risk at layer $\mathcal{l}$. These attention weights identifying which upstream characteristics exert the strongest influence on downstream risk.

3.

Unsupervised Learning and Anomaly Detection

$$z_{it}\sim \sum _{k=1}^K{\pi }_k\hspace{0.17em}\mathcal{N}\left({\mu }_k,{\mathsf{\Sigma }}_k\right)\mathrm{with}{N}_{\epsilon }\left(z_{it}\right)$$

$$=\{\hspace{0.17em}{z}_{js}:\parallel z_{js}-z_{it}\parallel \le \epsilon \hspace{0.17em}\},$$

(9)

Equation 9 specifies unsupervised methods that identify latent ESG risk structures and abnormal behavior rather than direct predictions. Firm-time observations $z_{it}$ are assumed to arise from latent regimes $k$. while defining local density neighborhoods used for unsupervised clustering and anomaly detection [30].

$$\mathrm{IF}\_\mathrm{score} \left({\mathrm{z}}_{it}\right)=2^{-{\displaystyle \frac{h\left({\mathrm{z}}_{it}\right)}{c\left(n\right)}}},$$

(10)

Equation (10) estimate the anomaly detection using an Isolation Forest with $T$ random trees. The function $h\left({\mathrm{z}}_{it}\right)$ denotes the average path length of ${\mathrm{z}}_{it}$ across trees. Shorter paths lengths and greater isolability reflect more anomalous behaviour. where $n$ is the sample size and $c\left(n\right)$ are the average path length of unsuccessful searches in a binary tree. A High $\mathrm{IF}\_\mathrm{score}$ flags shipment or sentiment anomalies as early-warning ESG signals [31].

$$A\left(z_{it}\right)\hspace{0.25em} =\hspace{0.25em} \underset{\theta ,\varphi }{\mathrm{m}\mathrm{i}\mathrm{n}}\hspace{0.17em}\parallel z_{it}-g_{\varphi }\hspace{0.17em},\left(f_{\theta }\left(z_{it}\right)\right){\parallel }^2,$$

(11)

An autoencoder learns a low-dimensional representation via encoder $f_{\theta }$ and decoder $g_{\varphi }$, where $g_{\varphi }\left(f_{\theta }\right(z_{it}\left)\right)={\widehat{z}}_{it}$is the reconstructed observation and $A\left(z_{it}\right)$denotes the autoencoder-based anomaly score interpreted as an atypical pattern in ESG incidents, shipments, or sentiment [32].

4.

Natural Language Processing

$${\mathrm{T}}_{it}={\displaystyle \frac{1}{\mid {\mathcal{D}}_{it}\mid }}{\sum }_{d\in {\mathcal{D}}_{it}}{\widehat{\theta }}_d, S_{it}={\displaystyle \frac{\sum _{d\in {\mathcal{D}}_{it}}{w}_{dt}\cdot {\mathrm{tone}}_{dt}}{\sum _{d\in {\mathcal{D}}_{it}}{w}_{dt}}}$$

(12)

Text-based ESG indicators are integrated into panel regression models to retain interpretability. Topic exposure $T_{it}$ is derived from Latent Dirichlet Allocation, while $S_{it}^{-}$reflects adverse sentiment extracted through transformer-based models [33]. These NPL procedures enter the model as stochastic regressors, which allow marginal effects to be quantified from document-level $GDELT$ tone scores ${\mathrm{tone}}_{dt}\in [-\mathrm{100,100}]$, aggregated using relevance weights $w_{dt}$based on factors such as source prominence and content volume [34,35].

3.2. Test of Hypotheses

3.2.1. Test of H1: Firms Within Supply-Chain Networks Exhibit Higher ESG Risk Exposure than Other Firms.

The network-augmented regression (NAR) results reported in Table 8 provides evidence on the role of supply-chain structure in shaping firms’ ESG risk exposure. The estimated network-lag coefficient is positive and statistically significant ($WRᵢ$ = 0.351,$p$ = 0.011), indicating that ESG risk is not independent across firms but instead clusters within connected supply-chain networks. This finding supports the hypothesis that ESG risk propagates through inter-firm relationships. Centrality measures exhibit heterogeneous effects. Degree centrality is not statistically significant ($Degᵢ$= −0.018, $p$ = 0.262), suggesting that the sheer number of supply-chain connections does not, by itself, increase ESG exposure.

On the other hand, betweenness centrality ($Betwᵢ$= 0.094, $p$ = 0.016) and eigenvector centrality ($Eigenᵢ$= 0.157, $p$ = 0.011) are both positive and significant, implying that firms occupying brokerage positions or connected to influential partners face elevated ESG risk. Trade dependency is positive but insignificant ($TDᵢ$= 0.166, $p$ = 0.468), indicating that structural position dominates simple exposure measures once controls are included. Among the control variables, shipment volume significantly increases ESG risk, while stronger regional governance is associated with lower risk. Although the model explains a modest proportion of variation ($R²$ = 0.189), this result is consistent with ESG outcomes being driven by multiple interacting factors. Diagnostic tests reported in the Appendix (Table A1), including Moran’s I, LM-lag, and Wald tests, confirm the presence of spatial dependence and validate the network specification.

Figure 5 provides predictive evidence relevant to H1, which suggests a relationship between firms’ network positions and ESG risk exposure. Panels A and B show that predicted ESG risk increases monotonically with both the network lag term and eigenvector centrality; firms more strongly exposed to high-risk partners, or connected to influential actors, exhibit higher model-implied ESG risk. The association is smoother for network-lag, while eigenvector centrality exhibits greater dispersion, reflecting heterogeneous influence across supply-chain structures. Panels C and D present marginal effects by decile and reveal a consistent upward pattern, suggesting that network-related exposure intensifies toward higher deciles of connectedness and centrality. The partial dependence and marginal effects jointly support the hypothesis that supply-chain network structure contributes significantly to ESG risk beyond firm-level characteristics.

Table 9 reports robustness checks designed to test the hypothesis that ESG risk is shaped by inter-firm network dependence rather than firm characteristics alone. Three complementary specifications are applied: the Spatial Lag Model, the Spatial Durbin Model, and a panel fixed-effects model. Collectively, they test whether the findings are model-specific or structurally persistent. The term spatial follows the econometric convention and refers to network adjacency rather than physical geography. Across all specifications, the network-lag coefficient remains positive and statistically significant ($\mathrm{S}\mathrm{L}\mathrm{M}$= 0.314, $\mathrm{S}\mathrm{D}\mathrm{M}$ = 0.271, $\mathrm{F}\mathrm{E}$ = 0.284). This pattern indicates that ESG risk clusters systematically among connected firms, consistent with transmission through supply-chain linkages.

The $\mathrm{S}\mathrm{D}\mathrm{M}$ indicates that network effects extend beyond direct partners. Spatial lags of shipment volume (0.083), eigenvector centrality (0.067), and regional context (−0.212) suggest that firm scale, network prominence, and governance environments jointly influence ESG exposure. Thus, centrality effects remain stable across models, reinforcing the conclusion that ties to influential hubs increase risk. Fixed-effects estimates also confirm persistence, with lagged sentiment (0.021) and prior incidents (0.063) remaining significant. No causal interpretation is implied, due to observational design.

Table 10 reports results from graph-theoretic regressions examining whether firms’ positions within supply-chain networks are associated with variation in ESG risk. The findings provide strong empirical support for this hypothesis. Across environmental, social, and governance dimensions, betweenness and eigenvector centrality emerge as the most influential predictors. Firms occupying intermediary positions exhibit significantly higher ESG risk scores (Environmental: 0.472; Social: 0.286; Governance: 0.383), suggesting heightened vulnerability to operational disruptions and reputational spillovers transmitted through the network. Similarly, positive and sizable eigenvector centrality coefficients suggest that ties to influential or highly connected partners amplify ESG exposure. Thus, clustering shows a smaller but statistically significant effect, implying that dense local network structures can foster shared risk environments. In contrast, degree centrality indicates limited explanatory power, suggesting that the number of connections alone is less relevant than their strategic importance. Model fit statistics (R² ranging from 0.522 to 0.578) indicate substantial explanatory power relative to typical ESG applications. However, the results confirm that network position constitutes a significant determinant of ESG risk, beyond firm-level characteristics.

Table 11 shows the standardized coefficients indicate firms are positioned within the network matters more than the sheer number of connections they maintain. Across ESG dimensions, betweenness and eigenvector centrality consistently exhibit the strongest associations with ESG risk. Betweenness centrality shows the largest standardized effect for environmental risk (Std.$\mathsf{\beta }$ = 0.524) and a significant effect for governance risk (Std.$\mathsf{\beta }$ = 0.388), indicating that firms acting as intermediaries face elevated exposure to spillovers, regulatory attention, and disruption transmission. Eigenvector centrality exhibits its strongest relationship with social risk (Std.$\mathsf{\beta }$ = 0.518), suggesting that firms embedded within influential supplier networks experience greater reputational and labour-related pressures. Clustering coefficients contribute more modestly, particularly for governance outcomes, while degree centrality, although statistically significant in some specifications, yields comparatively smaller effects, implying that connection quality outweighs connection quantity. Therefore, the relatively high explanatory power of the models (R² between 0.52 and 0.58) underscores the substantive role of network structure in shaping firm-level ESG risk exposure.

3.2.2. Test of H2: Lagged AI-Extracted ESG News Sentiment has a Statistically Significant Relationship with Subsequent Firm-Level ESG Incident Risk

Table 12 reports a series of robustness checks assessing the association between lagged ESG news sentiment and subsequent ESG incidents. sentiment scores are standardised such that lower values indicate more adverse coverage. Across all model specifications, lagged sentiment enters with a negative and statistically significant coefficient, indicating that more adverse ESG-related media coverage is associated with higher future ESG incident risk. Therefore, negative coefficients imply increased ESG risk. In the fixed-effects OLS model with Driscoll–Kraay standard errors, the coefficient of −0.398 reflects a strong association after controlling for firm and year effects, as well as cross-sectional dependence. Comparable effect sizes are observed in count-based models. In the Poisson fixed-effects specification, the coefficient of −0.211 indicates that a one standard deviation improvement in sentiment corresponds to an approximate $19\%$ reduction in expected incident counts. Estimates from the negative binomial (−0.236) and PPML (−0.224) models are similar, confirming robustness to overdispersion and excess zeros. Temporal analyses indicate attenuation over longer horizons. When sentiment is lagged by two periods, the coefficient declines to −0.173, and in the distributed-lag specification, the short-run effect at $t-1$ (−0.311) exceeds the longer-run impact at $t-2$ (−0.089). Thus, heterogeneity analysis further shows stronger effects for firms occupying central network positions compared to peripheral firms. A placebo test using future sentiment yields no significant association, supporting correct temporal ordering. The results indicate that AI-extracted ESG news sentiment is a stable and informative predictor of future ESG incidents. However, all estimates are associational, and no causal interpretation is implied due to observational design, and unobserved confounding.

Table 13 assesses how extreme negative ESG sentiment in the preceding period relates to subsequent controversy outcomes across several model frameworks. In the logit results, a lagged sentiment shock is associated with a substantially higher probability of an ESG incident, with an estimated coefficient of 2.28 (p < 0.001), an odds ratio of 9.81, and an average marginal effect of +0.20. This indicates a pronounced increase in incident likelihood following highly adverse coverage. The Poisson specification yields comparable evidence, where the coefficient of 0.94 (p < 0.001) corresponding to an incident rate ratio of 2.56, implying a more than twofold increase in expected incident counts. Results from the Cox proportional hazards model further show that firms experiencing negative sentiment shocks enter an ESG controversy state more rapidly, with a hazard ratio of 1.51 (β = 0.41, p < 0.001). Network centrality remains positive and significant across all models, with estimated odds, rate, and hazard ratios of 1.69, 1.19, and 1.42, respectively, indicating elevated exposure for more centrally positioned firms. Firm size exhibits a consistently negative association, with odds ratios near 0.96 and a hazard ratio of 0.93. These findings demonstrate that extreme adverse ESG sentiment serves as a robust indicator of heightened future controversy risk.

Table 14 examines the robustness of ESG news sentiment across alternative NLP architectures in terms of alignment, inference, and prediction. Using FinBERT as a domain-adapted benchmark, cross-model correlations show strong agreement with RoBERTa (Pearson r = 0.84; Spearman ρ = 0.81) and weaker alignment with mBERT (r = 0.72; ρ = 0.69), indicating rising measurement noise with more general models. Despite this, inferential results are stable. Lagged sentiment is negative and significant in all logit models, with coefficients from −0.87 to −0.55 and odds ratios between 0.42 and 0.58; all confidence intervals exclude unity. Predictive tests using Random Forests show consistent gains from adding sentiment, with ΔAUC increases of 0.22 (FinBERT), 0.20 (RoBERTa), and 0.15 (mBERT). Smaller gains align with higher polarity flip rates (3%, 7%, 10%) and greater sentiment variance. Shock analyses confirm robustness, indicating bottom-decile sentiment raises incident odds by 10.3×, 8.9×, and 6.1×, respectively, with strongest effects for ESG-adapted models.

Figure 6 assesses the robustness of ESG sentiment across NLP architectures. Panel A evaluates predictive contribution, showing positive ΔAUC gains (0.22, 0.20, 0.15). Panel B examines inferential association, indicating a higher incident risk under sentiment shocks (OR = 10.3–6.1). Attenuation across models reflects measurement noise rather than structural instability.

Table 15 shows that models using only control variables perform little better than random guessing, with ROC–AUC values around 0.49–0.52 and balanced accuracy near 0.50, indicating that baseline firm characteristics have limited ability to predict ESG incidents. When ESG news sentiment is included, performance improves substantially across all models, with ROC-AUC rising to 0.66–0.79 and AUPRC increasing to 0.44-0.54. Balance accuracy and F1 scores also increase, reaching approximately 0.64–0.68 and 0.62–0.66, respectively, indicating that the improvements reflect genuine classification gains rather than threshold effects. Random Forest and XGBoost achieve the strongest results, with ROC–AUC of 0.79. Overall, ESG sentiment clearly enhances out-of-sample prediction performance without implying causality. Table A2 in the Appendix reports diagnostic tests supporting these findings. Delong statistics indicate significant$\mathrm{A}\mathrm{U}\mathrm{C}$ increases across models ($\mathrm{Z}$= 5.01–7.55, $\mathrm{p}$ < 0.001), while McNemar results ($\mathsf{\chi }²$ = 18.6–34.1) and strictly positive $\Delta \mathrm{A}\mathrm{U}\mathrm{C}$ confidence intervals further corroborate these improvements.

Figure 7. ROC and prediction-recall curves for predicting future ESG incidents Control +Sentiment Features.

Figure 7. ROC and prediction-recall curves for predicting future ESG incidents Control +Sentiment Features.

Table 16 shows a clear improvement in predictive performance as models progress from econometric approaches to machine-learning methods and then to graph-based AI. Baseline econometric models perform modestly. Logistic Regression and Poisson GLM record ROC-AUC values of about 0.69–0.71 and Brier scores above 0.21, indicating limited discriminatory power and weak calibration when ESG risk is inferred solely from-level characteristics. Including firm fixed effects improves performance slightly, with ROC-AUC rising to 0.73 and the Brier score falling to 0.207, although substantial variation in risk remains unexplained. Machine learning models provide stronger results. Random Forest and Gradient Boosting raise ROC–AUC to approximately 0.76–0.77 and reduce Brier scores to below 0.19. XGBoost further enhances performance, achieving a ROC–AUC of 0.81 and a Brier score of 0.168. These gains are accompanied by higher recall, increasing from 0.67 in the best econometric model to 0.72, indicating fewer missed ESG incidents rather than superficial accuracy gains. The largest improvements arise from the graph neural network, which attains a ROC–AUC of 0.87, the lowest Brier score at 0.149, and the highest precision and recall. The findings suggest that ESG risk prediction is strengthened by accounting for inter-firm connections. While the analysis is predictive rather than causal, it demonstrates that incorporating network structure yields more informative and reliable risk assessments than models based solely on isolated firm attributes. Figure 8 shows the comparative predictive performance of the models for ESG risk.

Table 17 compares the forecast accuracy and calibration of several machine-learning and graph-based models. The results show a clear performance gradient as models incorporate greater nonlinearity and structural information. Random Forest and Gradient Boosting deliver moderate performance, with Brier scores between 0.184 and 0.189, log-loss values above 0.53, and relatively higher MAE (0.287–0.296) and RMSE (0.417–0.423), indicating less precise probability estimates and larger average prediction errors. More advanced learners, including XGBoost and multilayer perceptions, further reduce forecast error. These models achieve lower Brier scores (0.168 and 0.161), improved log-loss values, and noticeable reductions in MAE (0.271 and 0.263) and RMSE (0.398 and 0.387), reflecting gains in both calibration and overall accuracy. The graph neural network (GNN) exhibits the strongest performance across all metrics. It records the lowest Brier score (0.149), log-loss (0.472), MAE (0.249), and RMSE (0.368), indicating the closest alignment between predicted probabilities and realized ESG outcomes. The narrow confidence interval around the Brier score suggests that these gains are stable rather than driven by sampling variation. The evidence underscores the values of network aware models for ESG risk forecasting, as incorporating inter-firm relationships enhances probability reliability and reduces prediction error.

SHAP (Figure 9) illustrates how individual feature contribute to predicted ESG incident risk across observations. Negative news sentiment in the prior risk period emerges as the most influential factor, with higher negative sentiment consistently pushing predictions toward higher risk. Network position also plays a major role. Firms with greater centrality exhibit positive SHAP values, indicating elevated exposure to ESG incidents through interconnected relationships. Environmental, social and governance risk indicators show similar patterns: higher values in each dimension are associated with increased predicted risk, although their marginal effects are smaller than those of sentiment and network position. Past ESG incidents contribute positively to current risk, conforming persistence in firm-level ESG exposure. Operational scale and context variables, such as shipment volume, regional context, and trade dependency, exert more modest and tightly distributed effects, suggesting that they act as background risk modifiers rather than primary drivers. Colours indicate feature magnitude. The vertical zero-line indicates a neutral effect. Positive SHAP values increase predicted ESG risk, while negative values reduce it.

Table 18 shows a detailed decomposition of classification outcomes on an independent test set, enabling direct comparison of how different models trade off missed ESG incidents against false alerts. Thresholds are selected by maximising the $\mathrm{F}1$ score on the validation set, ensuring that differences in performance reflect model capability rather than arbitrary cut-offs. XGBoost demonstrates a clear improvement, with recall rises to 0.66 and the false-negative rate falling to 0.34, while the false-positive rate declines slightly to 0.09. Gains in precision indicate a more balanced conversion of alerts into true incident detections. The graph neural network performs best overall, achieving the highest recall (0.74) and precision (0.54), along with the lowest false-positive (0.07) and false-negative (0.26) rates. These results suggest that incorporating network information improves the identification of ESG propagation, although false alarms remain present. This analysis enables a transparent assessment of predictive trade-offs in ESG risk monitoring. Figure 10 compares confusion matrices and error decomposition for the most efficient AI models used to predict ESG risk

4. Discussion and Implications

4.1. Discussion

This study confirms that ESG risk cannot be adequately understood as an isolated firm attribute. Instead, it is closely tied to the configuration and intensity of relationships that firms maintain within global supply chains. By applying network econometric techniques, machine learning models, graph neural networks, and NLP-based sentiment extraction, the analysis reveals that ESG vulnerabilities emerge and intensify through relational exposure, particularly for firms embedded in structurally influential positions. This network-based view aligns with recent scholarship emphasising interconnected ESG risk formation in production systems [36]. The persistent significance of network-lag effects and positional indicators, especially betweenness and eigenvector centrality, reveals that firms acting as bridges or connected to influential partners face heightened ESG exposure. These results are consistent with theoretical perspectives on contagion and spillover in complex networks whereby intermediaries transmit shocks more efficiently than peripheral actors [37,38].

These findings reveal that network structure conditions both the likelihood and persistence of ESG risk, even after accounting for firm characteristics, prior incidents, and institutional context. However, media-based ESG sentiment emerges as a significant anticipatory signal. Adverse news sentiment systematically precedes ESG incidents across a wide range of specifications, with diminishing effects over longer horizons and no detectable influence in placebo tests. This pattern supports an interpretation of sentiment as an early-warning mechanism rather than a contemporaneous reflection of realised events, consistent with prior evidence on text-based ESG indicators [39]. Sentiment signals are robust across multiple NLP architectures, and domain-adopted models yield more stable estimates, underscoring the value of contextualised language representations in ESG analytics. From a forecasting standpoint, the analysis highlights substantial gains from explicitly modelling inter-firm dependence. Graph neural networks outperform econometric and machine learning approaches across calibration and error-based metrics, producing probability estimates that more closely track realised outcomes. This is essential in ESG applications, where poorly calibrated predictions can delay intervention or misdirect oversight efforts.

4.2. Implications

The results indicate that firms should move beyond compliance-focused, firm-level ESG assessments toward risk management approaches that explicitly consider supply-chain structure. Firms occupying central or intermediary positions within production networks merit closer attention, as their exposure reflects both internal practices and risks transmitted through connected partners. Integrating real-time ESG news sentiment into monitoring systems can further strengthen early-warning capacity, enabling firms to respond before emerging issues escalate into formal controversies. The improved probability calibration achieved by graph-based models has directly operational relevance. More reliable risk estimates reduce unnecessary alerts while limiting overlooked exposures, allowing firms to allocate sustainability, auditing, and compliance resources more effectively across complex and geographically dispersed supply chains. For regulators and standard setters, the findings emphasized that ESG risk should be viewed as a systemic phenomenon embedded within supply networks. Oversight frameworks focused solely on direct corporate activities may underestimate exposure arising from upstream and downstream linkages. Incorporating network-informed indicators into supervisory and disclosure regimes could improve the identification of systemic ESG vulnerabilities, particularly in sectors characterised by dense cross-border production relationships. From a research perspective, this study demonstrates the value of integrating econometric analysis with AI-based prediction and interpretability tools.

This study is limited by its reliance on observational data, which limits causal interpretation.

Supply-chain relationship may be partially observed or static over time. ESG event records and sentiment indicators can contain measurement noise. Model effectiveness may differ across industries, regions, and alternative data environments.

5. Conclusion

The proposed framework captures how ESG vulnerabilities emerge, diffuse, and persist within interconnected production systems. The evidence shows that firms occupying structurally influential or intermediary positions face disproportionately higher ESG exposure, highlighting the importance of network topology in sustainability risk assessment. The analysis further establishes ESG news sentiment as a forward-looking signal that precedes the realization of ESG controversies. The consistent predictive power of lagged sentiment across multiple model classes and robustness checks supports its use as an early-warning indicator, while differences across NPL architectures emphasise the value of domain-adapted language models for ESG applications. Importantly, these insights remain descriptive and predictive and do not imply causal mechanisms.

From a modelling perspective, the results confirm that explicitly accounting for inter-firm dependence materially improves risk estimation. Graph neural networks deliver superior probability calibration and lower forecast errors than both econometric and convectional machine-learning approaches, which is particularly relevant in ESG contexts where inaccurate risk estimates can delay mitigation or misdirect oversight efforts. This study advances ESG research by reframing sustainability risk as a network-conditioned phenomenon and by demonstrating how AI-based, network-aware methods can enhance ESG monitoring at scale. The framework is designed to be adaptable across industries and regions, providing a foundation for more responsive and system-oriented ESG risk management. Future research can extend this work by incorporating alternative data sources, such as satellite imagery or audit records, and by developing causal designs to isolate specific transmission channels.