Gaussian Xai-ahp Framework for Esge Decisions

Cesar Yoshio Machado Pedroza; Arthur Damasceno Vicente

DOI: 10.22167/2675-6528-202602779

Article derived from a Final Course Work (TCC), with content based on the student’s original work and adapted to the editorial format of the E&S Magazine with the support of the ResumeAI tool, an artificial intelligence solution developed by Instituto Pecege for textual synthesis and organization.

Summary

The measurement of ESG performance in the mining sector is a complex challenge, given the tension between the demand for metals essential for the energy transition and the often underestimated environmental costs. To address this issue, the XAI-AHP-Gaussian framework was developed, integrating machine learning, artificial intelligence explainability (XAI), and the probabilistic Analytic Hierarchy Process (AHP). The objective was not only to predict ESG scores but also to understand their main drivers and the degree of confidence in the explanations. The framework was applied to 24 years of historical data from Teck Resources Limited, a Canadian mining company. A significant structural break was identified in Teck’s ESG disclosure pattern in 2018, the year of adherence to the TCFD, with an 844.53% increase in the ESG Disclosure Index, suggesting that voluntary regulatory frameworks can alter corporate behavior. The XGBoost model identified Market Liquidity as the dominant predictor (87.5% SHAP importance), a result interpreted with caution due to its possible correlation with company size. SHAP-LIME triangulation demonstrated strong descriptive convergence (r=0.8580). The Gaussian-AHP assigned dominant weight to the Environmental dimension (56.38%). An exploratory event study around the Mount Polley disaster (2014) recorded an abnormal return of -60.03% for Teck, interpreted as a possible sector-wide reflection. The study demonstrated the framework’s utility for rigorous and interpretable ESG analysis, reinforcing the relevance of the ESGE concept and the integration of XAI in decision support systems.

Keywords: AHP-Gaussian; ESGE; Explainable AI; Machine Learning; Mining Sector.

1. Introduction

The mining sector faces increasing pressure related to environmental, social, and governance (ESG) factors. Although vital for the global energy transition, providing metals such as copper, lithium, and cobalt, this industry operates with significant environmental costs, often underestimated by traditional accounting methods (Hilson and Murck, 2000). The demand for these critical minerals is projected to increase substantially, between 400% and 500% by 2050, to meet the goals of the Paris Agreement (Azadi et al., 2020). This duality positions mining as part of the problem and the climate solution, making ESG governance not just relevant, but urgent for the sector.

The measurement of ESG performance in the mining sector is hindered by specific methodological issues. Firstly, there is inconsistency among ESG ratings provided by different agencies, such as MSCI, Sustainalytics, and Bloomberg, due to the adoption of distinct criteria and weights (Berg et al., 2022). Secondly, machine learning predictive models, like XGBoost and Random Forest, although achieving high accuracy, operate as “black boxes”, making it difficult to understand the reasons behind their predictions (Chen and Ma, 2021). Finally, there is a fragility in the weightings in multi-criteria decision-making methods, such as the deterministic Analytic Hierarchy Process (AHP), which treats expert judgments as fixed points, ignoring the inherent uncertainty in these assessments.

Tools to mitigate these challenges already exist, but they are rarely combined systematically. Methods such as SHAP (Lundberg and Lee, 2017) and LIME (Ribeiro et al., 2016) offer explainability for complex models, while DiCE (Mothilal et al., 2020) allows for the generation of counterfactual scenarios. In the domain of weightings, AHP-Gaussian (Saaty and Vargas, 1987; Hauser and Tadikamalla, 1996) addresses uncertainty by replacing point judgments with probability distributions propagated via Monte Carlo simulation. The gap lies in the absence of a systematic combination of these approaches for longitudinal ESG analysis in the mining sector.

The regulatory context, especially the recommendations of the Task Force on Climate-related Financial Disclosures (TCFD), created by the Financial Stability Board in 2015, makes this gap even more pressing. The TCFD disclosure model has quickly become a market expectation, with financial regulators in Canada, where the studied company operates, signaling adoption starting in 2017-2018. Evidence suggests that early adoption of the TCFD is associated with a lower cost of capital, possibly due to reduced information asymmetry with investors (Krueger et al., 2020). This study focuses on Teck Resources Limited (TSX: TECK-B), a diversified Canadian mining company, selected for its strategic relevance as one of the largest in the sector in Canada and for the availability of data that allow us to observe its transition and the evolution of the importance of ESG factors in its operations, including the reaction to events such as the Mount Polley disaster in 2014.

Given this problem and the need for a more rigorous and transparent approach to ESG evaluation, the present work is justified by proposing and empirically testing a framework that integrates explainable artificial intelligence (XAI) and the Gaussian Analytic Hierarchy Process (AHP-Gaussian) for the longitudinal analysis of ESG performance in the mining sector. The general objective of this study is to build a longitudinal ESGE dataset via text mining of corporate reports and market financial data; test structural breaks in ESG disclosure in 2018 via Chow test; develop predictive models with temporal validation; apply XAI triangulation validating methodological convergence; implement AHP-Gaussian with sensitivity analysis; perform exploratory event analysis around the Mount Polley disaster (2014, Imperial Metals), investigating possible sectoral reflection on Teck; and make the framework available in a decision support system (Streamlit) and Power BI dashboard for executive application.

2. Material and Methods

The research was characterized as descriptive-explanatory, seeking to describe the temporal evolution of the ESG performance of Teck Resources Limited through statistics and explain phenomena such as structural breaks and convergence of explainable artificial intelligence (XAI) methods. A mixed-methods approach was adopted, with a quantitative predominance, which included machine learning (ML) modeling, statistical tests, and Monte Carlo simulations, complemented by a qualitative component of text mining of corporate reports.

The study employed a longitudinal single-case strategy, allowing for a granular analysis of ESG dynamics over a 24-year period. The empirical subject was Teck Resources Limited (TSX: TECK-B; NYSE: TECK), a diversified Canadian mining company, founded in 1913 and headquartered in Vancouver, British Columbia. The company operates in four business lines: Copper, Zinc, Metallurgical Coal, and Energy.

The selection of Teck Resources Limited was justified by the availability of high-quality public data, including annual and sustainability reports since 2001, its exposure to multiple ESG challenges, and its sectoral relevance, being ranked among the top ten global companies in copper and the top five in metallurgical coal. This choice allowed for the observation of the transition and evolution of the importance of ESG factors in its operations.

The data collection covered the period from 2001 to 2024 and used three primary sources. Firstly, 24 annual and sustainability reports were downloaded from Teck’s corporate website. These documents, in PDF format, were converted to plain text using the pdfplumber library, with PyPDF2 as a contingency method, to preserve the spatial structure of text blocks.

From the full text of the reports, two variables were constructed: the Report Quality Score, defined as the character count as a proxy for report length, and the ESG Disclosure Index, which measured the frequency of mentions of sustainability anchor terms in the body of the document. This index was designed to measure disclosure volume, not actual ESG performance, and the count was absolute, without normalization by document length.

Secondly, historical series were extracted for Teck Resources (ticker TECK-B.TO) and for the S&P/TSX Composite index (^GSPTSE), used as a market benchmark. The extraction was performed via the Yahoo Finance API, obtaining closing prices, trading volume, and annual returns. Additional features were calculated such as Market Liquidity (average annual volume), Annual Return (year-over-year percentage change), and risk-return metrics, including Sharpe Ratio, Beta, 95% VaR, and Maximum Drawdown.

Third, for the exploratory event analysis around the Mount Polley disaster, which occurred on August 4, 2014, a market model was used. This model estimated Teck’s expected return via OLS regression in a 252-day pre-event estimation window, allowing for the calculation of abnormal return.

The ESGE (Environmental, Social, Governance, Economic) dataset was built by normalizing features via min-max scaling to the [0,1] range. The final dataset comprised 23 annual observations, referring to the period from 2001 to 2024, after removing one observation with a missing value. Four features were used in the modeling stage: ESG Disclosure Index, Report Quality Score, Market Liquidity, and Annual Return. Expanded ESG indicators were mapped but did not integrate into the final selection due to the unavailability of complete and comparable series over the 24 years.

The proposed XAI-AHP-Gaussian framework integrated five sequential modules. Module 1, named Feature Engineering ESGE, built normalized features from raw data. Module 2, ML Predictive Modeling, trained models using ensemble methods with temporal validation. Module 3, XAI Triangulation, applied global SHAP, local LIME, and counterfactual DiCE to explain model predictions.

Module 4, AHP-Gaussian Weighting, implemented Monte Carlo simulations for the probabilistic weighting of ESG dimensions. Finally, Module 5, Executive Dashboard, materialized the obtained insights into a decision support system developed in Streamlit and a dashboard in Power BI for executive application.

In descriptive and inferential statistical analysis, descriptive statistics such as mean, standard deviation, coefficient of variation, skewness, and kurtosis were applied to characterize the distributions of the variables. Normality was tested using the Shapiro-Wilk, Kolmogorov-Smirnov, and D’Agostino-Pearson tests. Pearson and Spearman correlations were calculated between the features, with a significance test via p-values (α=0.05).

To test the structural break in 2018, the Chow test was employed, specifying a linear regression model. The magnitude of the break was quantified by the percentage change between the means of the pre- and post-TCFD adoption periods. In Machine Learning modeling, the dataset was divided via TimeSeriesSplit (five folds), preserving the temporal order to avoid data leakage.

Three models were trained: Baseline (Mean Predictor), Random Forest, and XGBoost. Hyperparameters were optimized via GridSearchCV with cross-validation. Performance was evaluated by R² (coefficient of determination), RMSE (root mean squared error), and MAE (mean absolute error), with 95% confidence intervals for R² estimated via bootstrap (N=1,000 resamples). Feature importance was calculated via Gini importance (Random Forest), Gain importance (XGBoost), and Permutation importance (model-agnostic), and overfitting was verified through learning curves.

In the XAI analysis via SHAP-LIME-DICE triangulation, three complementary methods were applied to the XGBoost model. SHAP (TreeExplainer) calculated Shapley values for all features and instances, aggregating them via absolute mean to obtain global importance. LIME (LimeTabularExplainer) selected ten representative instances via k-medoids clustering, generating one thousand perturbations per instance for local explanations, with coefficients aggregated via weighted mean.

DiCE (DiceML framework) generated five counterfactuals per instance via gradient-based optimization, specifying a 10% increase in the ESGE score as the desired outcome, and analyzed the counterfactuals’ feasibility. SHAP-LIME convergence was validated through Pearson correlation between normalized importance vectors, with thresholds defined for strong, moderate, and weak convergence. The Mean Absolute Difference between SHAP and LIME importances was also calculated.

For the Gaussian Monte Carlo AHP, a 4×4 pairwise matrix was constructed for the ESGE criteria, using Saaty’s scale (1-9). The central judgments (µ_ij) were defined a priori by the author, based on ESG priorities reported in the literature for the mining sector (Kumari and Behera, 2021). Each element a_ij was modeled as a Gaussian random variable a_ij ~ N(μ_ij, σ²), where µ_ij represented the central judgment and σ controlled the uncertainty.

10,000 Monte Carlo simulations were performed. In each iteration, a stochastic matrix was sampled, the principal eigenvector (weights) and the Consistency Ratio (CR) were calculated, accepting the iteration if CR was less than 0.10. Four noise levels σe {0.05; 0.10; 0.15; 0.20} were tested, and σ=0.10 was selected as the baseline for balancing realism and robustness.

Financial benchmarking compared Teck Resources with the S&P/TSX Composite using risk-return metrics, such as Average Return, Volatility, Sharpe Ratio, Sortino Ratio, Beta, VaR 95%, CVaR 95%, and Maximum Drawdown. The difference in average returns was tested via independent t-test. For the event analysis of the Mount Polley disaster (2014), the abnormal return (AR) and cumulative abnormal return (CAR[0,+2]) for Teck were estimated using a market model via OLS in an estimation window of 252 pre-event trading days.

3. Results and Discussion

The analysis of longitudinal data from Teck Resources Limited revealed a complex dynamic in the evolution of ESG performance, with findings that underscore the importance of robust and interpretable methodological approaches. The results indicate substantial heterogeneity in ESGE variables over the 24-year period, reflecting the inherent volatility in the mining sector and corporate disclosure practices. The application of the XAI-AHP-Gaussian framework allowed not only for the prediction of ESG scores but also for the identification of key drivers and the assessment of confidence in the explanations, as proposed by the study’s general objective.

Descriptive statistics and structural break in 2018

The ESGE dataset, comprising 23 annual observations after the removal of a missing value, demonstrated considerable heterogeneity among the variables. The ESG Disclosure Index, for instance, showed a coefficient of variation of 131.10%, indicating sharp temporal volatility in Teck’s transparency practices. Market Liquidity also exhibited a high coefficient of variation of 95.27%, a result that can be attributed to the equity liquidity cycles that frequently accompany fluctuations in metal commodity prices. The Annual Return, in turn, recorded a standard deviation of 81.79%, with a coefficient of variation of -170.22% calculated on the absolute value of the mean, reflecting Teck’s significant exposure to price variations of metallurgical coal and copper in the global market.

The application of the Chow test to identify a structural break in 2018 revealed a statistically robust result for the ESG Disclosure Index, with an F-value of 53.94 and p<0.0001. This finding rejected the null hypothesis of parameter stability, indicating a significant change in the company’s reporting pattern, which coincided with the adoption of the Task Force on Climate-related Financial Disclosures (TCFD) recommendations that year. The magnitude of this break was notable, with an 844.53% increase in the average ESG Disclosure Index between the pre-TCFD period (average of 21.21) and the post-TCFD period (average of 200.43).

This result corroborates the hypothesis that voluntary regulatory frameworks can substantially influence corporate behavior in terms of transparency. However, it is crucial to interpret this finding with caution, as the study’s design does not allow for the establishment of a direct causal relationship. Competing factors, such as increasing stakeholder pressure and market expectations for greater ESG disclosure, may have operated simultaneously, contributing to the observed change. Market Liquidity also showed a statistically significant structural break in 2018, with F=28.37 (p<0.0001) and an increase of 115.73%, suggesting a possible association with increased investor interest in companies with robust disclosure practices, although other explanations remain plausible.

In contrast, the Annual Return did not show a statistically significant structural break (F=0.54, p=0.4705) in the same period. This data suggests that, despite the significant increase in Teck’s ESG transparency after adopting the TCFD, this change did not translate into superior financial performance in the short term. This observation is crucial for understanding the complexity of the relationship between ESG disclosure and shareholder value, indicating that financial benefits may be long-term or mediated by other factors not directly captured in this analysis.

Predictive Modeling via Machine Learning

In the predictive modeling stage, the XGBoost model demonstrated superior performance compared to Random Forest and the baseline model. The coefficient of determination (R²) for XGBoost was 0.9999, with a 95% confidence interval between 0.9996 and 1.0000. Although this value indicates an almost perfect fit to the in-sample data, it is imperative to interpret this result with caution. The small sample size (n=23) may lead to inflated fit metrics, which do not necessarily reflect a generalizable predictive capability for out-of-sample data. In small sample contexts, complex models tend to memorize the training signal, which can generate an apparent fit with no real predictive value.

Although no relevant divergence was identified between the training and test scores in the subset evaluated via TimeSeriesSplit, the limited sample prevents definitively ruling out the occurrence of overfitting. The mean squared error (RMSE) of XGBoost was significantly lower (3.275) compared to Random Forest (82.973), which reinforces the better relative fit of XGBoost within the analyzed sample. However, this superiority does not automatically extend to generalizable predictive capability. Market Liquidity emerged as the dominant feature, with 87.5% importance via Gain importance and 67.6% via Permutation importance.

The prominence of Market Liquidity as a primary predictor requires careful interpretation. It is possible that stock liquidity is capturing firm scale, which in turn may be correlated with ESG investments, rather than a direct ESG commitment by the firm. Both dimensions, liquidity and ESG investments, often correlate with firm size, which introduces an ambiguity that the available data do not conclusively resolve. Therefore, the analytical focus of this work rests more on the insights generated by XAI methods and the AHP-Gaussian framework, whose methodological robustness, demonstrated by Monte Carlo simulations, is independent of the degree of predictive fit of the underlying model. For future predictive applications, expanding the data panel and adopting more restrictive regularization strategies are recommended.

XAI Triangulation and SHAP-LIME Convergence

The triangulation of Explainable Artificial Intelligence (XAI) methods revealed a strong descriptive convergence between the explanations generated by SHAP and LIME, with a Pearson correlation coefficient of 0.8580. This value, although high, must be qualified. The correlation was calculated on only four importance pairs (one for each feature), resulting in few degrees of freedom (df=2) and preventing the assignment of inferential statistical significance (p≈0.14). Thus, the coefficient is treated as a descriptive indicator of ordinal agreement between the methods, rather than as inferential evidence of causality or generalization. This result, however, is compatible with the relative interpretative stability of the XGBoost model in the evaluated subset, suggesting that the model’s explanations are consistent.

Market Liquidity once again stood out as the most dominant feature, both in the global SHAP analysis (with a value of 192.357) and in the aggregated LIME (with 254.931). Annual Return occupied the second position in importance, with values of 101.641 for SHAP and 81.452 for LIME. A notable discrepancy was observed in the ESG Disclosure Index, where SHAP assigned 5.932 and LIME 19.949. This difference suggests that LIME may overestimate the local importance of this feature, possibly due to spurious correlations identified in specific subsamples, a behavior that can occur in local explainability methods.

Additionally, the DiCE tool was used to generate counterfactual scenarios. The results indicated that incremental improvements in the ESGE score would require simultaneous increases of impractical magnitudes (above 200%) in both Market Liquidity and Annual Return. This finding is suggestive of a model saturation at Teck’s current performance levels. The interpretation is that the ESGE framework, as applied, may be operating near the performance ceiling observed for the company in the analyzed period, indicating that large jumps in the score would require unrealistic magnitude changes in the input variables.

Probabilistic weighting via Gaussian AHP

The implementation of AHP-Gaussian with a standard deviation (σ) of 0.10 demonstrated robustness in weighting the ESGE dimensions. The Environmental dimension received the highest average weight, corresponding to 56.38%, with a 95% confidence interval between 52.93% and 59.54%. This result is highly compatible with the centrality attributed to environmental challenges in the mining sector, which include critical issues such as greenhouse gas emissions, water use, and tailings management, as widely discussed in the literature (Kumari and Behera, 2021).

The remaining dimensions were weighted as follows: Social obtained 26.25%, Governance 11.80%, and Economic only 5.57%. This prioritization pattern is consistent with the ESG priorities frequently reported for the mining sector, where environmental impacts are perceived as the most significant. The proportion of consistent simulations, with P(CR<0.10) of 98.78%, reinforces the stability of the obtained weights, even under moderate uncertainty in pairwise judgments. The absence of rank reversal for σ=0.10 and a minimum rate of 0.03% for σ=0.20 suggest an ordinal stability of the dimensions, even in the face of more substantial uncertainties in judgments.

The probabilistic approach of Gaussian AHP, by modeling judgments as Gaussian random variables, mitigated the limitation of not having a formal expert panel to define priorities. Instead of assuming point estimates and exact judgments, the uncertainty inherent in these assessments was modeled and propagated via Monte Carlo simulation. This lends greater methodological robustness to the framework, allowing weights to reflect a spectrum of possibilities and the confidence associated with each, aligning with the need for more rigorous and transparent ESG analysis.

Financial benchmarking and event study Mount Polley

Teck Resources Limited’s financial benchmarking against the S&P/TSX Composite index, for the period from 2001 to 2024, revealed a pattern of superior return, but accompanied by significantly higher volatility. Teck recorded an arithmetic average return of 48.23%, in contrast to the S&P/TSX Composite’s 16.44%, a statistically significant difference (p=0.0089). However, Teck’s volatility was approximately three times higher (135.16% versus 27.67% for the index). Teck’s Sharpe Ratio (0.3346) was lower than that of the S&P/TSX Composite (0.4857), indicating that the company’s additional return did not compensate for the incremental risk on a risk-adjusted basis.

The Teck Beta, at 0.7738, suggests moderate systematic exposure to the Canadian equity market. However, when combined with total volatility, which is about five times that of the index, this pattern points to a predominance of idiosyncratic, i.e., firm-specific, risk in Teck. This implies that much of the variability in Teck’s returns is not linked to general market movements but rather to factors intrinsic to the company or sector.

The event study around the 2014 tailings dam disaster at Mount Polley, an Imperial Metals operation in the same jurisdiction, recorded an abnormal return of -60.03% for Teck that year (p=0.0305). This result is interpreted as a possible reflection of a sector-wide regulatory shock in British Columbia, which indirectly affected Teck. It is important to note, however, that this interpretation has three crucial caveats. Firstly, the abnormal return was measured against a market benchmark, which already absorbs part of a purely sector-wide effect, making it difficult to isolate the specific impact of the event.

Secondly, the annual granularity of the analysis differs from conventional event studies, which use shorter time windows, approximating this analysis to an annual residual study (Campbell et al., 1997). Finally, the year 2014 was also marked by a sharp drop in the price of metallurgical coal, one of Teck’s main products, meaning that the effects are not separable in this study design. The Cumulative Abnormal Return (CAR[0,+2]) of +381.92% observed in subsequent years is compatible with a gradual recovery from a depressed base, but does not allow for causality to be attributed to specific company actions. These findings, taken together, indicate that the materiality of extreme ESG events may not be linear, being priced differently from gradual deteriorations, with implications for risk models used by institutional investors.

In summary, the results obtained demonstrate the utility of the XAI-AHP-Gaussian framework for rigorous and interpretable ESG analysis in the mining sector, as per the study’s objective. The identification of a significant structural break in Teck’s ESG disclosure in 2018, the dominance of Market Liquidity as a predictor (with due interpretive caution), the descriptive convergence of XAI explanations, the prioritization of the Environmental dimension in AHP-Gaussian, and the sectoral impact of the Mount Polley disaster provide valuable insights. Although the study presents limitations inherent to the sample size and the single-case scope, it reinforces the relevance of the ESGE concept and the integration of XAI and probabilistic multicriteria decision methods for decision support systems.

4. Conclusion

This study sought to develop and apply a framework that integrates explainable artificial intelligence (XAI) and the Gaussian Analytic Hierarchy Process (AHP-Gaussian) for the longitudinal analysis of ESG performance in the mining sector. A significant structural break was observed in Teck Resources Limited’s ESG disclosure pattern in 2018, the year of adherence to the TCFD, with an 844.53% increase in the ESG Disclosure Index, suggesting that voluntary regulatory frameworks can influence corporate behavior. The XGBoost model identified Market Liquidity as the dominant predictor, a result interpreted with caution due to possible correlation with business scale. Strong descriptive convergence was observed between the explanations generated by the SHAP and LIME methods, and the AHP-Gaussian assigned a dominant weight to the Environmental dimension, with 56.38%, aligning with the centrality of environmental challenges in mining. Additionally, the exploratory event study around the Mount Polley disaster recorded an abnormal return of -60.03% for Teck, interpreted as a possible reflection of sectoral regulatory shock. The main contribution of this work lies in the proposition and empirical testing of a methodologically rigorous and interpretable framework, which integrates XAI and probabilistic multicriteria decision-making methods, generating inputs for decision support systems for managers, regulators, and investors, and reinforcing the relevance of the ESGE concept.

The study’s limitations include the small sample size (n=23) and the focus on a single case, which restricts the generalization of findings. The disclosure index, constructed using anchor terms, served as a proxy for disclosure volume, rather than actual ESG performance, without normalization by document length. The Gaussian AHP judgment matrix was defined by the author, although probabilistic modeling mitigated inherent subjectivity. For future studies, it is suggested to expand the framework for a multi-firm analysis, which would allow testing the generalization of observed patterns. It is also recommended to improve the disclosure index, including normalization and the use of language models to capture contextual nuances. Another promising direction is the development of a dynamic version of the Gaussian AHP, with weights that update as new information becomes available, bringing the framework closer to real-world decision-making conditions.

Bibliographic References

Azadi M., Northey S.A., Ali S.H., Edraki M. 2020. Transparency on greenhouse gas emissions from mining to enable climate change mitigation. Nature Geoscience 13: 100-104.

Berg F., Koelbel J.F., Rigobon R. 2022. Aggregate confusion: the divergence of ESG ratings. Review of Finance 26: 1315-1344.

Campbell J.Y., Lo A.W., MacKinlay A.C. 1997. The econometrics of financial markets. 2nd ed. Princeton University Press, Princeton, NJ, USA.

Chen T., Ma X. 2021. Predicting ESG performance with machine learning methods. Sustainability 13: 8677.

Hauser D., Tadikamalla P. 1996. The analytic hierarchy process in an uncertain environment: a simulation approach. European Journal of Operational Research 91: 27-37.

Hilson G., Murck B. 2000. Sustainable development in the mining industry: clarifying the corporate perspective. Resources Policy 26: 227-238.

Krueger P., Sautner Z., Starks L.T. 2020. The importance of climate risks for institutional investors. Review of Financial Studies 33: 1067-1111.

Kumari S., Behera B. 2021. Sustainability assessment in the Indian mining sector using multi-criteria decision making. Resources Policy 72: 102119.

Lundberg S.M., Lee S.I. 2017. A unified approach to interpreting model predictions. Proceedings of the 31st International Conference on Neural Information Processing Systems 30: 4768-4777.

Mothilal R.K., Sharma A., Tan C. 2020. Explaining machine learning classifiers through diverse counterfactual explanations. Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency: 607-617.

Ribeiro M.T., Singh S., Guestrin C. 2016. Why should I trust you? Explaining the predictions of any classifier. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining: 1135-1144.

Saaty T.L., Vargas L.G. 1987. Uncertainty and rank order in the analytic hierarchy process. European Journal of Operational Research 32: 107-117.

Article originating from the Final Course Work of the Specialization in Data Science and Analytics of the MBA USP/Esalq

To learn more about the course, click here and access the MBX Academy platform