Predicting the Future of Sustainable Finance: A Statistical and Machine Learning Analysis of ESG-Driven Investment Portfolios

Abstract: Global investments in Environmental, Social, and Governance (ESG) assets have surpassed USD 30 trillion, thus making sustainable finance one of the prevailing themes of the decade. A key question is whether ESG portfolios either outperform traditional investments or if so, which ESG dimensions play a significant role. This article builds a unique dataset of fund-level return data with ESG-rated portfolios, and applying both statistical and machine learning approaches to examine performance. We use regression models (linear, ridge and XGBoost) to predict returns and classification models (logistic regression and random forest) to predict benchmark outperformance. A primary contribution of this paper is the application of Bayesian regression, which considers uncertainty and produces probabilistic forecasts rather than point estimates. We also conduct relative feature importance analysis to reveal which pillar – environmental, social or governance – provided the most explanatory power for return forecasts. Early-stage findings indicate that ESG portfolios have often yielded competitive returns while exhibiting additional resilience during market volatility. Among the ESG factors, the Environmental pillar is the most predictive for creating long-term value. Overall, these findings offer new evidence that sustainability practices have the potential to provide additional ethical and financial benefits. This paper signifies a shift from descriptive reporting of ESG analytics to focusing on predictive, uncertainty-aware modelling, which articulates the combination of machinelearning with statistically sound practices. The implications for corporate strategy, public policy, and investment management are noteworthy, and it contributes to advancing SDGs 8 (Decent Work and Economic Growth), 12 (Responsible Consumption and Production), and 13 (Climate Action).

Keywords: ESG portfolios, supervised machine learning, regression models