Explainable machine learning and multi-objective optimization for forecasting corporate SDG performance

The current practice of sustainability assessments in the corporate world is based on backward-looking ESG scores, which offer limited insight into dynamic, non-linear and multi-objective sustainability transitions. This study proposes an Explainable Artificial Intelligence (XAI) Decision Intelligence Framework for predicting and optimizing corporate Sustainable Development Goal (SDG) performance. The framework leverages 4,000 firm-year observations from 12 industrial sectors across the globe from 2016 to 2023, incorporating hierarchical imputation, winsorisation, lead alignment, machine learning, TreeSHAP, Partial Dependence Profiles and Pareto optimisation. Out-of-sample evaluation across four sequential expanding-window forecast periods identifies Ridge Regression as the champion forecasting benchmark (mean OOS R² = 0.8949; RMSE = 2.6884; MAE = 2.1374), while LightGBM is retained as the nonlinear analytical surrogate (mean OOS R² = 0.8787; RMSE = 2.8871; MAE = 2.3124). TreeSHAP identifies concentrated model-specific predictive attributions, while the partial dependence analysis indicates heterogeneous nonlinear response patterns within the analytical surrogate. The sectoral Pareto frontiers clearly show a range of trade-offs, with frontier scores ranging from 81.71 to 89.36, depending on the sector. Small negative or near-neutral changes are simulated in selected sectors under the + 25% Green R&D scenario. Overall, the framework facilitates sustainable decision-making on an explainable, forward-looking and sector-sensitive basis. It offers corporate managers, investors and policy-makers interpretable evidence to allocate resources, plan for transition and implement sector-specific sustainability interventions.