Beyond linear assumptions: neural network tracing of climate-driven financial contagion in firms

1 Introduction

The financial implications of climate change with respect to investors, regulators, and corporate strategists have moved rapidly beyond conversations about the environment (Demekas and Grippa, 2021). Extreme physical shocks such as extreme heat waves, flooding, and wildfires, as well as transition impacts from volatile carbon prices, environmental policies, and the move towards cleaner technologies, bring significant uncertainties to firms (Schult et al., 2024). Uncertainties associated with climate change are no longer hypothetical, but realities that affect balance sheets, firm-level asset values, credit spreads, and stock price volatility (Bell and Van Vuuren, 2022).

The Task Force on Climate-related Financial Disclosures (TCFD), the Network for Greening the Financial System (NGFS), and other large financial actors have all issued statements making it clear that climate risk should be viewed as financial risk in regulatory systems, investment strategies, and corporate risk management procedures (Network for Greening the Financial and S, 2022). Although there exists strong agreement on the significance of climate risk, the how remains poorly understood (Li et al., 2024). Evaluative frameworks that aggregate risk at a sectoral or national scale miss the intricate, rich diversity within industries: firms differ in their geographic exposure, carbon intensity, regulatory sensitivity, and operational structures (Bressan et al., 2024; Mandel et al., 2025; Sun et al., 2024).

More recent research has attempted to construct climate measures at the business level. One of them illustrates that, using North American conference call disclosures, corporate vulnerability to climate risks is inversely proportional to corporate risk-taking behavior, revealing sophisticated internal attitudinal firm responses to climate uncertainty (Rahman and Bose, 2025). Other research uncovers novel metrics relating operational productivity to firm-specific climate factors, revealing critical micro-level differences as well. Conventional financial risk models based on linear regression, factor analysis, or econometric stress tests aren't equipped to analyze climate shocks, specifically because climate shocks are multi-dimensional, conditional, and lagged events (Schötz et al., 2025).

Conventional financial models use climate variables as inputs in the model and assume some form of linearity, and that the effects of the climate shocks are stable. These models do not allow for an endogenous firm response to the climate shocks or include the possibility of non-linear interactions that impact firm outcomes (Schötz et al., 2025; Tabachová et al., 2025). Thus, they miss out on scenarios where a climate policy shock leads to deterioration of supply chain stability, changes investor sentiment, or enhances a competitor’s technology position. To fill this methodological gap, we provide a machine learning framework, based on Artificial Neural Networks (ANNs), to track and quantify the transmission pathways from both physical and transition climate shocks to corporate financial risk outcomes. ANNs provide a robust toolbox for uncovering latent structures in rich, high-dimensional datasets structures that cannot easily be explained through simple additive relationships (Cashman et al., 2025; Gu and Dunson, 2023; Guo et al., 2025).

With event datasets on climate as well as characteristics of firms like emissions intensity, exposure to supply chain risk, geographical footprint, and financial health, our models will not just ascertain whether climate risk matters but diagnose how it matters, and, importantly, under what conditions it emerges as quantifiable financial stress (Tabachová et al., 2025; Fialkowski et al., 2026; Li, 2025).

Time-series recordings of environmental occurrences and policy changes, as well as novel indices of firm-level climate exposure, are among the many recent data sources we use. We aim to create a predictive model which we expect to be able to predict not only when firms’ probability of default increases due to the climate shock but also to conduct robustness tests on the model with related risk metrics such as credit spreads and equity volatility. In summary, we want to move from traditional risk metrics to a situated, conditional interpretation of vulnerability where year-over-year differences in changing climate transitions or policy changes can create non-linear developments on firm finances depending on firm specific exposures - this is further supported by recent examples of this at the individual firm level.

In summary, this study contributes to three areas: firstly, it proposes a new perspective on the transmission of climate risk at the firm-level while emphasizing the intermediate variables between these processes, secondly, it illustrates analogy ability of ANNs to adapt to the patterns within the data to a wider range of explanatory variables than conventional modelling would suggest, particularly in the ability to learn from the data without imposing strong structural assumptions, and lastly, the study identifies firm-specific risk factors - such as regulatory frameworks, location and emissions profile - that increase vulnerabilities and give actionable intelligence to financial institutions and corporation management to respond to risks when developing disclosures, investment, and resilience plans. This paper delves into a real and relatively unexplored space with the central question in particular, how do climate risks flow through to firm-level financial outcomes across a range of scenarios?

2 Literature review

The literature surrounding climate-related financial risk has changed rapidly in the last few years, particularly as it has become more obvious how the effect of environmental factors can impact company level financial outcomes. Two main kinds of climate risks have been targeted: (1) physical risks (impact of extreme weather and gradual environmental change), and (2) transition risks (impact from changing regulatory regimes, technology, disruptive technologies and changing market state of mind).

Tangible climate-related risks have been related to declines in corporate performance and increased volatility. A study from 1990 to 2017 reports that major climate news events enhanced co-movement across financial markets, were networked, and led to a rise in systemic risk (Mao et al., 2023). When considering the outputs of the strategy, most of these analyses stop at the macro level or at the market level and little has been done on the firm level. Policy gains, like carbon price and green subsidies, are received in the form of top-down stress tests by the Bank of England, ECB and IMF on the transition front. These stress tests focus on climate risk but are subject to theory-based assumptions on which they are based, and a lack of granular, disaggregated information at the firm level (Radzewicz-Bak et al., 2024). In addition, network-based methods have demonstrated how macro-prudential propagation (Battiston et al., 2017) through the financial system can be helpful to complement our firm-level focus, as it has been observed how climate-policy shocks can propagate through a financial system at the macro-prudential level.

All the frameworks written as part of the DNB, BIS and BoE contain climate stress tests in addition to the regular financial risks. These frameworks are of prime importance not just for visualizing exposure and systemic risks, but on a broad basis. Many of these, however, rely on stylized balance‐sheet scenarios or industry averages, and miss the nuance of what could be driving the results: company‐specific characteristics like carbon intensity or supply chain fragility. Meanwhile, AI and machine learning are coming to the forefront in the field of risk modeling for finance. Combined with SSA, LSTM models could accurately forecast credit risk and carry out scenario analysis according to the transition, and their reliability and nonlinearity are hard to observe with econometric models (Wang and Huang, 2023). In the interim, neural methods have been developed to uncover signals for climate risk in textual disclosures, like earnings calls, with high predictive power for firm volatility and investment activity (Kim et al., 2023). Likewise, central banks are testing AI tools such as the BIS’s Gaia to categorize climate-related disclosures in multi-jurisdictional datasets, enhancing the harmonization of this information (Moreno and Caminero, 2023).

Despite progress in climate-finance modelling, most of the approaches available do not currently have the capacity to combine physical and transition climate risks and the financial consequences of climate change in one detailed framework. Climate risk is generally perceived as an external cost, and its linkages to other specific characteristics of firms and sectoral transmission channels are relatively weak. We aim to fill this observed gap in relation to integrating climate risk with risk to firm-level financial outcomes by proposing an Artificial Neural Network (ANN) based model that is able to represent climate shocks and their resultant financial risks in a data-driven, firm-specific, effective way. The approach that we propose provides documentation of how firms are exposed to climate risks and how climate shocks affect financial outcomes for firms, based on examples of three ANN-based credit risk and disclosure-based models. We propose ANN-based integrated models that are more flexible, interpretable, and scalable, and can be extended to existing models, thereby increasing the understanding of the exposure of corporate climate risk.

Our focus is also located in and builds upon two adjacent literature. Structural and statistical models of firm-level default, including the Merton (Merton, 1974) distance to default approach that we follow for the construction of PDs (Supplementary Appendix A2), and discriminant/logit-based models, are still widely used in corporate credit risk assessment, with recent research working on integrating these models with market-based and ESG-related indicators (Postiglione et al., 2025). In non-climate credit-risk contexts, a new body of literature has also started to systematically benchmark machine-learning strategies against these classical benchmarks (Braunsberger and Aschauer, 2025), and thus far, typically at the margin of accuracy achieved and at the expense of interpretability. Second, the above reviewed literature on ML-for-climate-finance has a positive view on the benefits of non-linear data-driven methods, while at the same time a parallel stream of literature has raised concerns regarding the deployment of black-box models in credit and financial-stability applications, such as the challenge of auditing model logic for supervisory purposes (Černevičienė et al., 2024). The second literature has motivated us to develop our SHAP-based attribution framework (Section 6), which we believe is not a resolution of all of the concerns about interpretability.

3 Theoretical framework

This experiment goes beyond simply descriptive analyses and static predictive models, and this experiment offers a systematic analytic framework that explains the channel of transmission through which climate shocks become financial shocks, or risk exposures. The framework uses a systematic, dynamic conceptualization of the channels of transmission that extend from the immediate disruption of operations to the extended adjustment of markets, representing the core of our inquiry.

Our approach here is predicated on a bifurcation that embraces both immediacy and complexity in respect to climate-related risks. The first channel, direct climate transmission, includes sudden, tangible impacts such as flooding that, for example, inundates a production facility, heat waves that stress an electrical grid, or wildfires that suspend supply chains. Each of these events results in immediate financial implications related to revenue, operating costs, and liabilities (e.g., insurance). Moreover, firms with concentrated physical exposures such as manufacturing plants or other operations dependent on localized geography and commodities are specifically vulnerable to direct shocks.

By contrast, the indirect route implies more complex, and often lagged, impacts. For example, regulatory measures such as a carbon tax or an increase in investor preferences for sustainable assets can affect firm value, even if the firm has not experienced any physical damage to it. It is important to recognize that these regulatory and market-driven processes can change access to capital, alter consumer behavior, and impact credit ratings. Similarly, if the firm is perceived to be irresponsible in its treatment of the environment, there is the potential for reputational risks to exacerbate this financial impact. A growing body of academic research is now recognizing this dual-path transmission structure.

Figure 1 illustrates the multifaceted transmission mechanism through which climate shocks impact corporate financial risk. At the initial stage, climate inputs encompass both physical events such as floods and wildfires and transition-related developments like carbon pricing and regulatory reforms. These inputs propagate through distinct transmission pathways, the direct channel manifests as operational disruptions and asset impairments, causing immediate financial strain, whereas the indirect channel involves more nuanced effects, including regulatory repercussions, shifts in investor sentiment, and reputational damage. Ultimately, these processes culminate in measurable financial outcomes, reflected in credit risk indicators, fluctuations in stock price volatility, and pressures on corporate liquidity. This layered framework provides a structured lens to understand the complex pathways linking climate phenomena to firm-level financial vulnerability.

Creating a conceptual framework is just the beginning, the main challenge is to model the complex, nonlinear properties of climate-related financial risks. Many default econometric models do not have enough flexibility to appropriately account for direct and indirect effects, firm heterogeneity, temporal lags, and generally rely on too many assumptions that ignore significant dynamics. Artificial Neural Networks (ANNs) are a good alternative as they can process many inputs, including climate hazard data, geographical vulnerability indicators, emissions and policy exposures, and rapidly changing key financial data inputs associated with complex multi-dimensional interactions.

Recent evidence has demonstrated that AI approaches are good at predicting multi-hazard outcomes and mapping systemic financial risks, while still being interpretable with tools like SHAP. Using the model (Figure 2) with multiple hidden layers, activation functions chosen for a specific purpose, normalization, and regularization, we extract latent risk signals and ultimately other granular predictive indicators linked to financial risk, like credit spread, variability, and liquidity issues. SHAP analysis helped to outline the contribution of specific features to the outputs of a model and will help to conceptualize interpretability in practice.

4 Data and variables

This study utilizes a multi-layered dataset that integrates climate shock indicators with firm-level financial and geographic data to trace the transmission of climate events into corporate financial risk metrics. Careful data harmonization, temporal alignment, and spatial matching were applied. Physical risks are assessed by EM-DAT disaster records, and additional remote sensing data from NASA GPM and ERA5 (precipitation, temperature, wind speed) were also utilized. This will allow geo-located identification of extreme weather events related to firm headquarters or facilities. Transition risks come from IEA and WRI data (regulatory changes, carbon pricing) amplified by ESG indicators provided by Refinitiv and MSCI (carbon disclosure, carbon emission intensity, exposure to policy). Financial data were gathered from Bloomberg, Refinitiv Eikon, and CompStat Global datasets, the data collected illustrate volatility, debt-to-equity ratios, firm-level VaR, CDS spreads, and credit ratings.

All these data records were collected quarterly so that the time series would align with climate events, and lag analysis could be calculated. Preprocessing included normalization, fissurization, and spatiotemporal interpolation of missing climate data. Firms were spatially linked to climate shock zones using geo-coordinates. VIF analysis and principal component transformation were used to manage multicollinearity, and financial metrics were lagged to reduce simultaneity bias. The final balanced panel consists of 1232 publicly listed firms (2014–2023) from North America, the EU, and Asia-Pacific. Tables 1 and 2 summarize the variables and descriptive firm characteristics by sector, size, and region.

TABLE 1. Summary of variables and data sources.

PD, is the primary outcome variable used throughout the study. VaR, CDS, spread, and annualized volatility are used exclusively as alternative outcome variables in the robustness checks reported in Section 6.3, and are not part of the primary predictive model.

TABLE 2. Sample description of firms used in the study (N = 1,232).

5 Methodology

To model the complex and nonlinear transmission of climate shocks to firm-level financial risk, we develop an artificial neural network (ANN) framework able to accommodate structured financial variables alongside high-frequency climate variables. Our dataset has both cross-sectional observations and temporal observations, so we use a multilayer perceptron (MLP) framework as our primary model, while we also model some parts of the overall analysis using Long Short-Term Memory (LSTM) networks to accommodate.

5.1 Multi-layer perceptron (MLP)

A conventional MLP (Multilayer Perceptron) architecture consists of an input layer, output layer, and hidden layers. All layers are completely linked. There are a number of neurons in each layer, which corresponds to neurons in the human brain. The input to each neuron (most basic element of the network) is connected to the previous layer via a set of weights, and the weighted sum of those inputs is added to a bias, and then passes through an activation function, which is used to compute the output of each neuron. The AF is therefore crucial to the overall performance of the entire neural network and especially during the training of the network. The action of the neuron may be mathematically characterized as Equation 1 (Goodfellow et al., 2016):where F denotes the AF, b is the bias, x j is the input of the neuron, wj is the corresponding weight, and m is the number of inputs (Goodfellow et al., 2016).

y = F b + ∑ j = 1 m w j . x j (1)

For ease of understanding, we recap the entire modeling process below. The main variable of interest in this study is one quarter ahead PD, which is a continuous variable (in bps) and is trained using MSE loss. The accuracy of the model on this continuous target is reported using RMSE, MAE, R ² and MAPE. The predicted continuous PD is also thresholded at 100 bps to obtain a binary default indicator and report the AUC and the confusion matrix for interpretability and comparison with classification based baseline. To assess the robustness of this main pipeline, without changing the pipeline, we extend it with 3 sets of robustness checks: (i) alternative pipeline, different ANN architectures, LSTM and 1D CNN, with the same continuous PD target, (ii) alternative pipeline, different classification baselines, logistic regression (logit) and a random forest (RF) classifier, with the same thresholded PD, and (iii) 2 models independently trained with different outcomes, idiosyncratic equity volatility and the Altman Z score, to test the generalization of the transmission mechanism beyond the PD definition. For both the PD prediction and the volatility robustness checks, the GARCH(1,1)-X model is implemented extending the original model with exogenous regressors representing climate features.

5.2 Long Short-Term Memory

LSTM network is a type of RNN that aims to solve the vanishing gradient problem (Hochreiter and Schmidhuber, 1997). LSTMs have demonstrated promising performance when used for forecasting nonlinear and long-range sequences. The forward LSTM processes the input sequence in chronological order from left to right. During each time step, the hidden state ht and cell state ct are computed according to Equations 2–7 (Zhang et al., 2023; Zhao et al., 2024):

i t = σ W i x X t + W i h h t − 1 + b i (2)

f t = σ W f x X t + W f h h t − 1 + b f (3)

O t = σ W o x X t + W o h h t − 1 + b o (4)

c t ∼ = tanh ⁡ W c x X t + W c h h t − 1 + b c (5)

c t = f t ⊙ c t − 1 + i t ⊙ c ∼ (6)

h t = O t ⊙ tanh ⁡ c t (7)

The mathematical formulation involves several key parameters: W_fx denotes the weight matrices associated with the forget gate, W_ix, shows the weight matrices associated with the input gate, and W_ox, represents the weight matrices associated with the output gate, which connect the input and hidden layers. The gate activations at time step t are referred to with the symbols f_t, i_t, and o_t for the forget, input, and output gates. The gate bias terms for the forget, input, and output gates are b_f, b_i, and b_o. The gates utilize the sigmoid function σ as their activation function, and tanh is typically used to constrain the value of input, c, to the memory cell.

The input layer consisted of climate risk indicators (e.g., flood severity index, carbon tax exposure), and specific attributes of each firm (e.g., leverage ratio, size, sector dummy variables). These were concatenated into a single feature vector. The output layer generated just one single target, the probability of default (PD) of the firm computed at one quarter ahead, as basis points (bps) derived from a Merton (Merton, 1974) distance-to-default framework. Since PD is bounded within [0,1] a sigmoid output activation has been employed. Alternative financial risk metrics (Value-at-Risk (VaR), idiosyncratic equity volatility, and Altman Z-score) were only used for robustness checks in Section 6.3, and were not part of the primary training objective.

More formally, the model learns a function f: X → Y, where X is the joint climate-firm feature space.

Hidden layers were optimized iteratively (through various trial and error), with two to five fully connected layers with ReLU activation functions, while dropout regularization applied at each layer (0.2–0.5 dropout), where intended to prevent overfitting, and standard dropout rates were experiment based. Moreover, batch normalization was employed at every layer to stable the training process. The last output layer employed either a linear or sigmoid activation function, depending on the nature of the financial risk metric being perturbed.

After removing observations with insufficient temporal coverage (fewer than 10 complete quarters of climate and financial data per firm), the usable panel consisted of 12,320 firm-quarter observations. This panel was split into training (70%, N = 8,624), validation (15%, N = 1,848), and testing (15%, N = 1,848) subsets using stratified sampling by industry sector, geographic region, and climate exposure category.

The split was actually done at the firm-quarter level, not the firm level, so that this can result in different quarters showing up in different subsets. The potential for information leakage as a consequence of this design choice is discussed in Section 7.1.

This was done to ensure equal distribution of organizations across sectors and event exposure types during training and in the evaluations. Mean squared error (MSE) and mean absolute error (MAE) represented our primary loss functions during training, which also included L2 regularization in the loss term. Early stopping was employed during the training phase with a patience of 10 epochs using validation loss to prevent overfitting during training. To optimize the ANN, hyperparameter tuning was performed using a grid search over the hyperparameters: batch size (32,64,128), learning rates of (0.001,0.0005), hidden layer neurons (64,128,256), and dropout rates (0.2,0.3,0.5). The descriptions of the configurations that were evaluated with respect to hyperparameter tuning are included in Table 3.

TABLE 3. Hyperparameters and Configurations tested.

Table 4 shows the complete model training path, such as pretraining and configuration, optimizer (Adam), and summary of model training. We tested the accuracy of the ANN by comparing it with four baseline models from different methodological traditions in financial risk prediction. Logistic regression is the most common linear method of binary default classification used in credit-risk practice. The linear prediction of continuous PDs is given by linear regression. Random forest was used as a non-linear, non-neural benchmark, which enabled us to see the specific effect of the non-linear, gradient-based feature interactions of the ANN in comparison to the non-linear, tree-based interactions, as well as to assess it in both the classification and continuous prediction tasks. GARCH(1,1)-X, which was extended using climate features as exogenous regressors, was then tested on both tasks and on the specific dedicated volatility robustness check. To evaluate all models, each was tested with a common held-out data set and tested using a consistent set of metrics appropriate to the particular task, and reported together in Table 5.

TABLE 4. Description of model training procedure.

TABLE 5. Hyperparameter tuning and validation-based model selection for comparison models.

To make all the baseline models comparable, they were optimized through grid search on the validation set using logistic regression on the L2 penalty strength, random forest on number of trees, maximum depth, minimum leaf size and climate-shock lag length, and GARCH(1,1) on error distribution and climate-shock lag length.

All baseline models were fine-tuned using a grid search on the validation data set for the logistic regression L2 penalty strength, the random forest number of trees, maximum depth, minimum leaf size, and feature subsampling rate, while the GARCH(1,1)-X was fine-tuned for error distribution and climate-shock lag length.

Overall, the ANN models excelled at modelling the interaction terms and the nonlinearities associated with climate and transition risks. This was particularly evident in the downward, tail-side predictions, such as the Value at Risk (VaR) metrics. Furthermore, saliency mapping was employed post hoc to analyses the features most influential in making the model’s predictions, improving model transparency and aligning them with regulatory requirements for explainability.

6 Results

6.1 Quantified transmission using SHAP

When applied to the commonly held-out evaluation sample, and after tuning all the models, the ANN model achieved the best overall predictive performance. The ANN delivered an AUC of 0.840 and an RMSE of 29.0 bps for default discrimination and continuous prediction of the PD, respectively, while a logistic/linear model had an AUC of 0.724 and an RMSE of 39.7 bps, a tuned GARCH(1,1)-X model had an AUC of 0.751 and an RMSE of 36.4 bps and a tuned Random Forest model had an AUC of 0.806 and an RMSE of 32.1 bps. Fair tuning only marginally reduced the difference between the ANN and the two non-linear baselines, when compared to our initial default-configuration comparison, which suggests the ANN’s success is not a result of a disproportionate amount of tuning effort. Model rankings were uniform across discrimination (AUC), prediction-error (RMSE, MAE) and calibration (Brier score) measures.

Because both models were tested on the same held-out test set, we used a DeLong test (DeLong et al., 1988) for the difference between correlated ROC curves to formally assess the statistical significance of the difference between the ANN and logistic discrimination. The bootstrap confidence intervals reported in Table 6 also do not overlap, and the ANN’s AUC (0.840) significantly outstrips the logistic baseline (0.724, difference = 0.116, z = 5.214, p < 0.000001) as expected given the results from Table 6, showing that the discrimination advantage reported is not due to sampling variation.

TABLE 6. Consistent out-of-sample performance comparison across benchmark models (test set, all models independently tuned via grid search on the validation set).

AUC, accuracy, Sensitivity, Specificity, and Brier Score reflect the binary default-classification task (threshold = 100 bps, N = 1,300). RMSE, MAE, and R ² reflect the continuous PD-prediction task (N = 1,848). All baseline hyperparameters were selected via grid search on the validation set (Table 6) using the same protocol as the primary ANN, ensuring a fair comparison.

After training the multi-layer perceptron (MLP) with two hidden layers (64 and 32 neurons, ReLU activation, dropout 0.2) on the training set (70% of the panel), we achieved an out-of-sample AUC of 0.84 for predicting one-quarter-ahead probability of default (PD). This is much better than the linear logit model (AUC 0.72), and it supports the idea of non-linear transmission mechanisms (Hypothesis H1). We go beyond predictive accuracy and quantify the impact of a climate shock on risk by using SHAP (SHapley Additive exPlanations). SHAP values basically break down the prediction of the model into individual contributions of each input feature at the level of each firm-quarter. In a specific prediction, the SHAP value of a feature is its contribution to the prediction (marginal contribution) averaged over all possible orders of features. This method retains the non-linearities and interactions learned by the ANN, and offers a human-interpretable attribution (Lundberg and Lee, 2017).

We stress here that SHAP values pertain to the given model fitted to the ANN and are meant to measure the contribution (i.e., explainability) of each input feature to the model prediction, rather than the effect that the feature would have on default risk in the real world. The use of the words effect or ‘impact’ or ‘amplification’ in the rest of this section means the size and direction of the learned attribution in the model, and not necessarily a causal effect, as the discussion of causal identification and its limitations is dealt with separately in Section 7.

The global SHAP summary (beeswarm plot) is shown in Figure 3. On the y-axis are all input features, ordered by their mean absolute SHAP value. Also on the x-axis is the SHAP value, which is the impact of each feature on the predicted PD (in bps). The individual data points are for individual observations of the firms in a single quarter, and the color represents whether the feature value is high (red) or low (blue) for that observation. The figure shows that the drought index (11.2 bps), followed by firm leverage (14.7 bps) and flood depth (21.3 bps) have the largest average absolute SHAP. Perhaps most importantly, the high leverage values (red points) are concentrated to the right on the SHAP axis for both flood depth and drought, suggesting that for high leverage values, an excess of one unit of climate shock has a significant impact on the reported default risk. On the other hand, the SHAP values of weak leverage (blue points) are close to zero or even negative, indicating that more financially sound companies should be able to withstand moderate shocks. Moreover, more than 99% of the SHAP values for flood depth are positive, indicating that even small flood events tend to raise the default risk of most firms in the sample.

Figure 3 illustrates the global importance, the actual transmission function the marginal effect of the climate shock on the risk - is represented in the SHAP dependence plot. Figure 4 shows the SHAP value for flood depth as a function of the flood depth value (in m), given the other features’ actual values for each instance. This is overlaid with a locally weighted scatterplot smoothing (LOWESS) curve. The nonlinear pattern is evident in the plot: for flood depths less than 0.2 m, the SHAP values are close to zero, and the mean contribution is just 0.04 bps, suggesting that there is not significant extra default risk that increases with flood depth. SHAP values are positive between 0.2 and 0.5 m, but not very high, with an average of 0.25 bps. The shallower the slope, the greater the mean SHAP value, as it rises sharply beyond 0.5 m to 10.02 bps for the depth range of [0.5, 1.0). In this area, for every meter rise (from 0.6 to 0.7 m), the predicted PD rises by about 2.0 bps (from 6.9 to 8.9 bps), and the increase in PD is consistently greater with increasing depth. This is just one of the kinds of processes that linear models (which require a constant slope) cannot explain. We take Figure 5. To be the magnitude of the transmission mechanism between a physical climate shock and corporate credit risk quantified.

To understand the reasons for the increase in risk with the flood depth, we break down the overall SHAP contribution of each climate shock into three transmission channels: (i) direct asset damage (flood depth × property, plant, and equipment density), (ii) supply chain disruption (flood depth × supplier concentration HHI), and (iii) expectations/revaluation (policy shock × analyst coverage). The SHAP values of the interaction terms for each channel are computed for each firm-quarter, averaged over firms, and divided by the SHAP value of the raw climate shock term. This decomposition is shown in Table 7 for three types of shock.

TABLE 7. Decomposition of total climate shock effect into transmission channels.

A greater proportion of total impact (58%) for flood events is indirect asset damage flood causes damage to physical capital, which lowers the value of the collateral and raises the default risk. The supply chain channel makes up 32%, underlining the relevance of production network disruptions even for companies without their own assets directly inundated. For physical shocks, expectations are relatively small (10%) and presumably from a local perspective, as flood risk is already at least partially priced. In the case of drought, both direct and supply chain channels are equally important (45% each), as drought affects agriculture and water-intensive industries, and its impact is also felt upstream. This is a sharp contrast to a carbon policy shock (e.g., an unforeseen rise in the EU ETS carbon price), where the expectations channel is by far the most dominant (60%). This aligns with investors adjusting their future earnings and discount rates at the time of policy announcements, and physical damage to assets being small. The results confirm Hypothesis H2: there is systematic variation of the relative importance of transmission channels by shock type.

The metrics for each evaluation include RMSE, MAE and R ² for prediction of the continuous PD target and AUC, accuracy, sensitivity, specificity and Brier score for the thresholded binary default classification, reflecting the two different prediction tasks defined in Section 5.1. We present both sets together, not a single summary statistic, since a model with good discrimination (high AUC) but poor calibration (high Brier score) is as bad for stress-testing and capital-planning uses as a model with good calibration but poor discrimination, and conversely, a model with good calibration but poor discrimination is not as good as a model with good discrimination and good calibration.

In addition to the statistical significance obtained from the above-described bootstrapping confidence intervals and DeLong test, the size of the ANN’s gain is economically significant for stress-testing applications. The ANN outperforms the logistic model relative to the logistic baseline in terms of the continuous PD prediction error, in that it reduces the positive difference (RMSE) from 39.7 to 29.0 bps, or 27% fewer PDs, and in terms of default-classification sensitivity, the ANN correctly classifies about 14% more of the truly defaulting firms in our test sample than the logistic model, at a similar false-positive rate. This represents a material reduction in the number of default events that are not identified by those monitoring hundreds, if not fewer, of risks at the portfolio level, and we believe this is the most practical benefit of the ANN over and above its interpretive benefits using SHAP in this context.

6.2 Heterogeneity across firms and industries

The results above are presented as averages, but there is significant variation across sections. First, we study the effect of firm leverage on the transmission of flood shocks. The dependence plot in Figure 4 is reproduced in Figure 5, with the color gradient indicating the firm’s leverage ratio (total debt/total assets). Yet, even in the case of firms which are conservatively financed (blue points, in the bottom quartile), the SHAP curve is still very shallow, and the default risk is estimated to rise by 1.16 bps per meter for all moderately severe floods in the range of 0.2–0.7 m, which is still small. Conversely, for high-leverage firms (red points–top quartile), the SHAP curve is sharply steeper, with a slope of 7.91 bps per meter a 6.79-fold higher SHAP-attributed slope than for low-leverage firms. Such interaction is learned without a pre-specified multiplicative term, it is just there in the ANN.

Another important source of heterogeneity is industry affiliation. Figure 6 shows a bar chart of the mean absolute SHAP value for flood depth across 5 industry groups: Utilities, Manufacturing, Retail, Real Estate, and Technology. The highest mean |SHAP| (18.0 bps increase in PD per 0.5 m flood) is associated with utilities, which have an asset-intensive infrastructure (power plants, transmission lines) and infrastructure that is geographically concentrated. Manufacturing is close behind (14.0 bps), driven by the complexity of the supply chain and physical capital requirements. Real Estate is No. 3 (12.0 bps), and property values are extremely responsive to flood risk. Retail has the lowest, but significant impact (8.0 bps), partly due to disruption of the supply chain. The most secure are the technology companies, at 3.0 bps, a sign of geographic diversification of data centers, intangible asset dominance, and agile supply chains. The ratio of vulnerability between Utilities and Technology is 6.0, pointing to significant differences between industries in their climate risk exposure. This ranking serves as a yardstick for a specific sector’s climate risk assessment.

After obtaining the SHAP value for flood depth of each firm quarter, we conduct a quantile regression of the SHAP values on a set of firm characteristics at the median. The coefficients are presented in Table 8.

TABLE 8. Firm-level characteristics that moderate transmission.

Economically and statistically significant all the coefficients. An increase in leverage from the 25th to the 75th percentile is associated with an 8.2 bps larger SHAP attribution for flood depth. An increase in the amount of cash that firms hold has a negative impact on transmission, decreasing by 4.3 bps, presumably because cash reserves help firms to repair damage or maintain operations without immediate default. Supplier concentration measured by the Herfindahl index of the firm’s supply chain has a similar effect on transmission, with an increase of 6.7 bps for every increase in concentration (e.g., one key supplier increases the SHAP). Last but not least, firms reporting the presence of property insurance show a 9.1 bps smaller SHAP attribution for flood depth, a pattern consistent with risk transfer mechanisms mitigating the model-attributed pass-through of physical shocks to credit risk. These results support Hypothesis H3.

6.3 Robustness and falsification tests

To make sure that our quantified transmission mechanism isn't a model artefact or a correlation by chance, we run a few tests. To begin with a falsification (placebo) test: For each company, climate shocks (flood depth, drought index) are measured from geographic locations with a distance of at least 500 km from all the locations of the firm’s reported assets. If the main results were driven by model artifacts or omitted spatially correlated confounders rather than genuine exposure to climate shocks, this placebo shock should retain similar predictive power. The distribution of SHAP values from these placebo shocks is plotted in Figure 7 for all firm quarters. The distribution has a mean of 0.053 bps, a standard deviation of 1.24 bps, and a t-test does not reject the null hypothesis that the true mean is zero (p-value = 0.1726). Furthermore, the number of bps for which the absolute SHAP values are greater than 2, is below 11.0% for the placebo and more than 99% for the true flood depth. The outcome is in line with transmission being driven by genuine level-specific exposure, and not by spurious spatial correlation. This test, however, only considers a specific threat: spatial confounding, and does not provide complete causal identification.

Second, to test robustness against alternative architectures of ANNs. Table 9 compares the key features of the main MLP (2 hidden layers) with two other architectures: an LSTM using three lags of climate shocks (to account for dynamic propagation), and a 1D convolutional neural network (CNN) over a window of 4 quarters. All models are trained with the same data and tested with the same test set.

TABLE 9. Robustness to alternative ANN architectures.

The LSTM also has a slightly lower RMSE (27 bps) and SHAP rank correlation of 0.98 for the feature of flood depth, indicating that the importance of this particular transmission channel is the same in both architectures. The same results are obtained by the 1D CNN. This robustness check only involved the overall predictive accuracy and the rank ordering of a single feature’s attribution, the channel decomposition, industry/regional heterogeneity (Sections 6.2, 6.5), and temporal dynamics (Section 6.4) were not re-estimated here for the LSTM or CNN architectures. Thus, this is reported as evidence that our main quantitative results are not architecture-dependent, but not as a statement that all downstream analyses are architecture-independent.

Third, we explore the relationship between financial risk and transmission mechanism when we adopt a new definition of financial risk. Table 10 presents results for two alternative risk measures: idiosyncratic equity volatility (as measured by a Fama-French three-factor model) and the Altman Z-score (an inverse measure of distress, with higher scores representing higher risk). The ANN’s predictive power is still very good (the R ² is 0.62 for volatility and the AUC is 0.79 for the Z-score distress).

TABLE 10. Sensitivity to alternative financial risk definitions.

Perhaps most importantly, there are differences in the ordering of the top three transmission channels for each risk metric. In the case of idiosyncratic volatility, the carbon policy shock is the most important component, followed by flood depth and supply chain interactions, in line with the notion that the announcement of a policy introduces uncertainty which quickly impacts equity markets. In the Z-score, the main shock is drought instead of flood depth, due to the high sensitivity of working capital and liquidity to extended droughts. In all three specifications, however, the non-linear, channel-decomposable transmission is still apparent and the direction of effects (e.g., leverage inducing shocks) has no changed. We also demonstrate that our results are not driven by our target definition by re-estimating the model using two alternative independently-trained financial risk measures.

6.4 Temporal dynamics of transmission

The previous subsections defined the cross-sectional and channel specific impacts of climate shock, but the temporal aspects of the transmission of climate shock are key for risk management. We use the LSTM variant to measure the impulse response of PD on a shock of one standard deviation of each type of shock, as to assess the speed of the physical vs. transition shocks’ effects on financial risk. Figure 8 shows how the cumulative percentage change in PD over the following eight quarters (2 years) after a shock.

The percentage increase in PD is highest in quarter 2 (6 months) for a flood shock and will gradually diminish but will stay above baseline values for eight quarters (5%). This persistence is due to the long-lived nature of asset damage and costs of reorganisation of the supply chain. The immediate answer to a carbon policy shock is a 9% increase in PD in the same quarter, and a subsequent gradual fall of 3% over 8 quarters. The initial response is in line with the expectations channel–markets immediately adjust prices for risk. The water scarcity for drought is the slowest, with only 12% increase in drought in quarter 3 (9 months) and 6% drought persistence after 8 quarters, this signifies the slow emergence of drought and the cumulative drought effects on agricultural and industrial production.

We further investigate the temporal heterogeneity by looking at the evolution of the SHAP importances for the climate features over our sample time period (2014–2023). The absolute value of the rolling window (12-month) averages of the SHAP are shown in Figure 9 for flood depth, drought index, and carbon policy shock. Flood depth importance increased from 7.7 bps in 2014 to 15.6 bps in 2023 (+101.7%), especially since 2018 when there have been several high-profile flood events and a rising awareness among investors. The pattern of drought importance is more volatile, reaching a high around 2016 of 12.2 bps, followed by peaks around El Niño years. The importance of carbon policy shock increased from 5.8 bps in 2014 to 15.3 bps in 2023, with a significant jump in 2021 of 10.4 bps after the EU’s “Fit for 55” package and the increase in carbon prices. This relative importance over time indicates that the physical and transition risks have gained a new significance in the financial markets in recent years.

6.5 Geographic heterogeneity

The transmission mechanism should be region specific as physical climate shocks are inherently local. We divide the sample into three general geographic regions: North America, Europe and Asia–Pacific (APAC). Figure 10 presents a regional average SHAP of the mean absolute SHAP values of flood depth and drought index, normalized by the respective regional average PD value.

North American companies have the largest impact (normalized SHAP = 1.8) with respect to flood depth, due to the large number of companies located in coastal regions and along river valleys, where flooding is more likely to occur, and also because the land use restrictions are less stringent than in Europe. The impact on Europe is also significant, though lower (1.2), reflecting improved flood protection, but also the high value of assets in flood prone areas. Asia Pacific has the lowest normalized SHAP (0.9) as a higher percentage of technology and manufacturing companies are in elevated or well protected zones, and perhaps because our APAC sample is skewed towards larger, more resilient companies. The percent vulnerability to flooding in North America is 1.5 times higher than that of Europe. Apart from drought, Europe has the highest normalized SHAP with 1.5, which is indicative of the sensitivity of agriculture and water-dependent industries in southern Europe. North America falls in the middle of the pack (1.3) and APAC is on the bottom again (0.8). These geographic variations highlight the need for geographic adaptation in climate risk evaluation.

6.6 Extended model validation and performance visualizations

In addition to the quantitative robustification in Section 6.3, we give some further graphical validations of the predictability of the ANN. Four diagnostic plots for the test set (15% of the data, held-out) are shown in Figure 11.

The predicted vs. actual PD (in bps) for the MLP model is shown in Panel (a). The points are closely grouped around the 45° line with a good overall fit as evidenced by an R ² of 0.914 for the test set. The distribution of the residuals (actual–predicted) is shown in Panel (b). Residuals are close to being normally distributed, with a mean of −1.51 bps, a standard deviation of 7.36 bps, a Kolmogorov-Smirnov test does not reject the null hypothesis of normality (p = 0.14). Panel (c) plots the c. d.-functions for the predicted PD of the top quartile of firms (high-leverage firms) and bottom quartile of firms (low-leverage firms). The median PD of 30.2 bps of the high-leverage group is right-shifted relative to the low-leverage group, which has a median PD of 14.7 bps, further reflecting the amplifying effect of leverage on the PD that ANN captures. In (d), the learning curves (training loss and validation loss) are plotted with respect to the number of epochs. The training loss smoothed out and the validation loss levels off after around 50 epochs, and no over fitting is visible, and early stopping was invoked at 78 epochs. These diagnostics indicate good overall fit and generalization, formal calibration assessment, confidence intervals, seed sensitivity, and split-sample stability are reported in Section 6.7.

Finally, to assess the stability of our estimation of the channel percentages across different random seeds and data splits, Table 11 shows the mean, standard deviation, and 95% bootstrap confidence interval of the channel percentages (for flood shocks) for 1000 bootstrap repetitions of the whole training/validation/testing process. The confidence intervals are narrow compared to the mean estimates in all three channels: importantly, the 95% CI for the direct asset damage (53.8%–61.1%) does not overlap with the one based on supply chain disruption (28.7%–36.4%), suggesting that the ranking and approximate size of the channel decompositions are not sensitive to sampling variation.

TABLE 11. Bootstrap stability of channel decomposition (flood shocks).

Table 12 provides a comprehensive summary of feature importance (mean absolute SHAP) across all major input variables, sorted descending. Flood depth remains the single most important feature (mean |SHAP| = 21.3 bps), followed by leverage (14.7 bps), drought index (11.2 bps), and carbon policy shock (9.8 bps). The ESG environmental score and disclosure risk flag rank lower but are still non-negligible, confirming that both physical and transition-related variables matter.

TABLE 12. Global feature importance (mean |SHAP|) for PD prediction.

The confusion matrix of binary classification (default vs. non-default) with a PD threshold of 100 bps is shown in Table 13. On the test set, the ANN results in a true positive rate (sensitivity) of 0.79, a true negative rate (specificity) of 0.87, and an overall accuracy of 0.85. The false positive rate is 0.13, and the false negative rate is 0.21. These values are more favorable than for the linear logit model (accuracy 0.738, sensitivity 0.650).

TABLE 13. Confusion matrix for binary default classification (threshold = 100 bps), based on the subset of test observations (N = 1,300) with an available realized default label.

Table 14 shows computational performance of our ANN framework. The final MLP (2 hidden layers, 64/32 neurons) was trained on the complete training set (34,480 observations) in 126 s using a standard GPU card (NVIDIA A10). The inference (prediction and SHAP computation for the test set, 1,848 firm-quarters) took 4.2 s. These times are suitable for real-time risk monitoring.

TABLE 14. Computational performance of ANN framework.

6.7 Extended validation: Class balance, calibration, and split-sample stability

No class-rebalancing technique was used in training the ANN or any of the classification baselines (Table 14), but instead the ANN was trained directly on the observed class distribution (21.6%–35.1% prevalence across the classification subsample). We believe that a moderate imbalance and not a severe one, so standard training is appropriate, but we report Brier score and calibration slope (Table 16) as complements to AUC, because rank-based metrics such as AUC are less sensitive to imbalance than calibration metrics.

In addition to overall model reliability measures, we provide class balance, calibration, estimation uncertainty, seed sensitivity, and regional and sectoral stability measures. The overall test set prevalence of defaults (PD ≥ 100 bps) is in the range of 7.4% (Technology) to 14.6% (Energy) across the sectors and 10.9% (Europe) to 12.5% (Asia-Pacific) across the regions of the held-out test set (Table 15). This moderate non-extreme class imbalance cannot be seen as an artifact of a degenerate class distribution, since the reported AUC and confusion-matrix values (Table 14) are not.

TABLE 15. Regional and sectoral stability of ANN performance on the held-out test set.

Table 15 gives AUC and RMSE values for each region/sector. The performance is fairly consistent, from the lowest AUC of 0.79 for the Technology section to the highest of 0.87 for the Energy section, and from the lowest of 26.8 bps to the highest of 33.2 bps for RMSE. Gendered results were also found, as described in the earlier section, that the Technology sector had the lowest climate shock exposure of all sectors analyzed, where, in all three metrics (AUC, RMSE and Brier score), the lowest performance is observed. The situation is similar in all other Asia-Pacific countries regionally, likely due to the smaller sample of countries that are more likely to be resilient, discussed in Section 6.5. None of the regions or sectors is as low as it would be considered unreliable (AUC values are all above 0.70, this is the threshold of the linear logit model). The stability was evaluated at the regional and sectoral level measuring the same classification subset N = 1300 as in Table 15.

We trained the main MLP five times with different random initializations of the parameters, and the same architecture and hyperparameters as in Table 16. This performance was found across different seeds, with AUC – 0.841 ± 0.003, RMSE – 28.9 ± 0.4 bps, and R ² – 0.912 ± 0.002, which shows negligible variation performance across different seeds and highlights that this is not a strong effect of a favorable initialization.

TABLE 16. Random seed sensitivity analysis.

We also measure the uncertainty of the estimation by 1,000 bootstrap resampling of the test set (Table 17). The 95% confidence interval of the AUC for the ANN (0.828–0.854) does not overlap with that of the logit baseline (0.704–0.736) and does not overlap with the random forest baseline (0.786–0.816), meaning that the performance advantage of the ANN is stable and not just due to sampling variation.

TABLE 17. Bootstrap confidence intervals for model performance (1000 bootstrap samples).

Lastly, calibration is evaluated formally, instead of by visual inspection only (Table 18). The ANN model also performs best with the lowest Brier score (0.089), and the lowest Expected Calibration Error (ECE = 0.031), the calibration slope is close to 1, which means the PD values predicted by the ANN model are close to the default rates over the range of probabilities. Beyond discrimination (AUC), both baseline models are relatively under-calibrated (higher Brier score and ECE, and calibration slopes further away from 1), indicating that the advantage of the ANN goes beyond discrimination to suitability of probability estimates, which is especially important for applications of stress-testing where absolute PD values rather than only ranking are relevant.

TABLE 18. Calibration performance.

7 Discussion

Artificial neural networks (ANNs) could serve as a substantial tool for capturing the complexity and nonlinearity of ways climate hazards could influence financial risk. While the absolute performance of the model stands out, it goes beyond that and could offer more detail on how indirect climatic stressors interact with internal firm characteristics to pose identifiable financial risks. The results provide interpretive value in terms of understanding the channels of propagation, and they will be of relevance to both policymakers and corporate strategists, while also noting the limitations of the model.

The analysis indicates that financial climate hazards rarely come from a single, discrete, and sequential occurrence such as a natural disaster. Rather, they result from repeated variability across time, derived from a synergistic, dynamic set of interactions between endogenous firm characteristics (capital structure, sectoral context, ESG preparedness) and exogenous factors (unstable regulatory context, frequency of hazard). Their interactions are often non-linear, with significant lag effects that traditional econometrics do not capture. In this way, the ANN captures latent interactions and identifies the risk an organization faces as a multidimensional process - the expected risk of different combinations of variables. For example, firms in heavily regulated energy-intensive industries showed disproportionately large SHAP attributions not only to physical regulatory policy changes. Such as carbon taxes, but also to anticipatory signals like announcements of policy changes, which resulted in volatility changes despite lacking any physical change. Conversely, within the technology industry, the model attributed smaller default-risk contributions to negative physical conditions for organizations with high ESG capability, a pattern consistent with the role of reputational capital and stakeholder perceptions. The patterns found confirm the view that climate risk is not one-dimensional, but rather influenced by firm-specific groupings that filter and amplify external shocks in different ways.

Overall, from the regulatory perspective, the key takeaway is the implication for regulators to incorporate new machine learning models into climate stress testing exercises. Model-based approaches improve the depth and predictive validity of the supervisory tools used by financial institutions (i.e., central banks and financial regulators) and any other actors reliant on predictive modelling. The outcomes from the study that included measures of firm-level ESG alignment bring the modelling closer to international regulatory standards (e.g., IFRS Foundation, European Financial Reporting Advisory Group) and can facilitate the alignment of risk analytic and disclosure-based regulations. The model’s ability to distinguish vulnerability profiles by industry and region provides the basis for differentiated policy responses, allowing regulatory authorities to move from prescribing uniform requirements to calibrated responses to risk (e.g., capital buffers, disclosure thresholds, or transition plan reviews).

At the same time, there are also similar implications at the corporate position. Senior finance and risk representatives can utilize ANN-based climate-financial risk models not just as compliance mechanisms, but as key parts of their internal risk management systems. Additionally, by exploring exposure under various climate and policy scenarios, firms can be able to identify weaknesses or vulnerabilities in their operational models, supply chains, or financial strategies. For instance, a manufacturing company that is facing water restrictions and emissions fines could use the model’s results to determine the financial implications of re-localizing their supply chain, decarbonizing their processes, or seeking better ESG integration. The focus on firm-specific data instead of sector-wide aggregates makes the information much more usable, and relevant to specific organizations, and ultimately supports better capital allocation, firm-specific valuation adjustments, and decoding external auditor plans for the firm. For an approach this individualized, we can move climate risk analysis from a macroprudential approach to a firm-centric decision-making tool.

7.1 Limitations and boundary conditions

Our approach has 4 important boundaries even though it has a lot of advantages. First, the sample of 1,232 publicly listed companies of North America, Europe and Asia-Pacific is not representative of the universe of all companies, but rather of large-cap developed market companies with relatively good disclosure practices, especially in Asia-Pacific. The results should be interpreted with some caution for small and private companies, and emerging markets that have different exposure and resilience characteristics. Second, the physical climate variables are derived from ERA5 data, which are available at 0.25 ° × 0.25 ° resolution (around 25–30 km), and location uncertainty is also present in the facility geocoding. These limitations could underestimate or overestimate local exposures and cause attenuation bias in the effects reported. Third, PD is built following the Merton (1974) distance-to-default methodology, which is based on equity prices and volatility. Some of the predictive power of the model might be driven by financial market expectations of financial risks, rather than a unique climate-to-default channel. We can take steps to reduce, but not eliminate, this possibility, through placebo tests and lagged controls. Lastly, the ANN is only trained and validated using the observations of 2014–2023. Because of its nonlinear threshold and leverage-amplification effects, it therefore represents observed relationships in the past and may not necessarily extrapolate to future, unprecedented, climate shocks or extreme regulatory stress-test scenarios, without further out-of-sample validation.

Design of the validation and potential leakage at the firm level. We split the data into train/validation/test units at the firm-quarter level, rather than the firm level, so that quarters from the same firm can be seen in each of the subsets. There exist some firm-specific characteristics that remain stable over time (such as a firm’s sector, region, or leverage at the beginning of the training period), which can be learned partly during the training period, allowing for better performance on the held-out firm’s observations. We see this as a boundary condition on the reported performance metrics, not a restriction that threatens the qualitative results on non-linear channel transmission and channel decomposition based on almost exclusively within-model SHAP attribution. Future research should confirm the performance using a strictly firm-level (grouped) train-test split as was not done here.

Moreover, our identification strategy does not completely overcome other threats to causal inference, whereas our placebo test does rule out spatially correlated model artifacts. The estimated transmission is not unbiased due to the fact that firm location is not randomly assigned and the unobserved characteristics of firms that are correlated with location and financial robustness (selection) may bias the estimate. Variation in regional confounding factors also might create false correlations. The reverse causality is also a possibility: poorer companies may invest less in their resilience infrastructure, which could make them seem more exposed than they actually are. The contribution by SHAP has a special role, in particular in explaining predictive contributions within the fitted model, and should not be interpreted as an estimate of the causal effect. We therefore consider our results as solid rather than as definitive causal estimates and as non-linear, consistent with causal interpretation. A quasi-experimental design approach, including an event study analysis centered on a discrete, unexpected event in the flood (such as a flood) or an instrumental variables design using precipitation shocks upstream of the event that are not correlated with firm-level characteristics, would be valuable for separately estimating the causal component of this transmission.

8 Conclusion

This research aimed to understand the various and convoluted processes through which climate-related shocks, whether the shock is physical or due to transition, translate to observable financial risks at the firm level. Through the use of an artificial neural network (ANN) architecture designed to accommodate nonlinear dependencies, latent interactions, and high dimensionality, we constructed a flexible framework that captures the transmission processes of climate risk with detail that few studies in the literature can achieve. The model was empirically validated across a wide range of firms, sectors and geographies demonstrating the merit of data learning approaches and the diversity of climate risk exposure of firms.

Our key findings can be summarized along three dimensions. First, the ANN significantly outperformed traditional econometric baselines (e.g., linear regression, GARCH) and even more advanced machine learning models (e.g., random forest), in terms of predictive accuracy, generalization, and responsiveness to complex scenario testing. Second, interpretability tools, such as SHAP, allowed us to open the “black box” and understand the most important risk drivers, which spanned firm-specific financial measures, climate policy uncertainty, and physical hazard frequency. Third, we saw in our sectoral and regional analyses that climate risk transmission is not uniform, carbon-intensive firms and firms in areas exposed to acute physical risks exhibit disproportionately high sensitivity, especially in transition scenarios with carbon taxation or regulatory tightening. Our results illustrate the possibilities of ANN-based frameworks not merely as predictive modelling but also as an assessment mechanism for climate stress testing, ESG risk benchmarking, and supervision. In terms of regulatory function, this kind of modelling may provide enough granularity for central banks or financial regulators, at the level of supervisory mandates for stress testing and scenario analysis in the fashion recommended by the NGFS (Network for Greening the Financial System). For practitioners including asset-level risk managers, CFOs, and ESG strategists, the ability to conditionally model climate-financial dynamics has real usefulness for capital allocation, repricing assets, and scenario planning in strategy. Despite the demonstrated advantages, several limitations merit discussion. First, the availability and quality of climate-financial datasets remain a constraint, particularly in emerging markets where disclosure is sparse or non-standardized. Second, while explainability techniques mitigate some of the opaqueness associated with deep learning, there remains a trade-off between model complexity and transparency especially in regulatory contexts. Third, our model is currently calibrated for firm-level analysis, extending the methodology to sovereign debt portfolios or system-level risk requires additional structural and temporal modeling. Building on these insights, several avenues for future research are both feasible and valuable. One promising direction involves integrating real-time climate indicators into the input architecture of the model. This would allow for adaptive, near-real-time risk forecasting, which is particularly relevant for financial institutions managing dynamic portfolios. Another extension involves embedding standardized climate scenarios, such as those provided by NGFS or the Shared Socioeconomic Pathways (SSPs), directly into the training and stress testing loops of the ANN. This integration would align model outputs with macroprudential planning needs and allow for harmonization with international scenario-based initiatives. Lastly, expanding the analysis from firm-level granularity to a portfolio, sectoral, or sovereign level could open new possibilities for climate risk assessment in investment management, banking supervision, and fiscal policy design. In sum, this work demonstrates that the fusion of artificial intelligence and climate-financial data has the potential to significantly advance both academic understanding and practical management of climate-induced financial risk. As climate change continues to reshape the contours of economic and financial stability, such interdisciplinary and data-intensive approaches will be indispensable.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. During the preparation of this manuscript, the author(s) used AI-based language tools to assist with grammatical editing and occasional paraphrasing of previously drafted text to improve clarity and readability. These tools were not used to generate original scientific content, data, analysis, or conclusions. All substantive content, data analysis, interpretation, and scientific claims are the sole responsibility of the author, who reviewed and take full responsibility for the final text.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fenvs.2026.1909195/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Alex Oriel Godoy, University for Development, Chile

Reviewed by: Shubham Metha, Northwest Bank, United States

Carluys Suescum Coelho, Centro de Estudios Gerenciales Avanzados, Venezuela