Deep Multi-layer Perceptron Regression for ESG-Driven Stock Return Prediction in China's Energy Industry

In recent years, as China accelerates its “dual‐carbon” goals and global investors increasingly integrate non‐financial metrics, the Environmental, Social and Governance (ESG) framework has become pivotal for evaluating corporate sustainability. This study aims to quantify the impact of ESG performance on annual stock returns for A‐share listed companies in the energy sector. We first conduct exploratory data analysis—including time‐series plots and correlation heatmaps—using Wind and Huazheng ESG scores alongside key financial indicators (net profit, total assets, return on assets, leverage, and asset turnover). To capture complex, nonlinear relationships, we apply a Multi-layer Perceptron (MLP) regressor—a feed-forward neural network with two hidden layers (10 and 5 units), ReLU activation, and the Adam optimizer—on standardized feature inputs, splitting the dataset into training and testing subsets. Our results indicate a statistically significant positive association: each one-point ESG increase corresponds to a 4.5% uplift in stock return (p < 0.05), and the MLP model achieves robust out-of-sample prediction accuracy (low MSE). Mediation analysis further shows that improved ESG scores boost returns partly via enhanced profitability (18% indirect effect through ROA). These findings underscore the value of deep learning techniques for forecasting market performance based on ESG metrics and offer investors and policymakers a data-driven tool to integrate sustainability into financial decision-making.