A Multimodal Explainable AI Architecture for Dynamic ESG Performance Prediction: Integrating Sustainability Intelligence, Investor Relevance, and Risk Dynamics in Saudi Listed Companies

This study develops an explainable machine-learning framework for predicting Environmental, Social and Governance Index Performance (ESGIP) of Saudi listed companies by integrating sustainability indicators, investor-relevance measures, and financial-risk dynamics. The framework combines descriptive statistics, Augmented Dickey–Fuller and Phillips–Perron stationarity tests, correlation analysis, ANOVA, principal component analysis, XGBoost, long short-term memory, Random Forest optimized by genetic algorithm and particle swarm optimization, and SHapley Additive exPlanations. All variables were stationary at level I(0). Sustainability Reporting Quality showed the strongest positive correlation with ESGIP (r = 0.8663), while Tobin’s Q was the strongest positive investor-relevance predictor (r = 0.6531) and Financial Distress Score the strongest negative predictor (r = −0.5921). PCA revealed a dominant sustainability–governance structure, with PC1 explaining 82.71% of variance in C1, compared with 45.01% in C2. In testing, RF-PSO improved NSE by 1.93% and reduced MAE by 4.84% for C1, while also outperforming XGBoost and LSTM for C2. SHAP identified Sustainability Reporting Quality and Tobin’s Q as leading ESGIP parameters. The framework supports transparent ESG assessment, sustainable investment screening, and risk-aware decision-making aligned with Saudi Vision 2030. These findings demonstrate the value of integrating corporate sustainability information with market and risk indicators to improve interpretable ESG intelligence in emerging markets.