ESG Performance Prediction and Driver Factor Mining for Listed Companies Based on Machine Learning: A Multi-Source Heterogeneous Data Fusion Analysis

Abstract

With the acceleration of global economic integration and the growing focus on sustainable development, Environmental, Social, and Governance (ESG) factors have become key standards for evaluating a company's long-term value and risk. However, accurately measuring the ESG performance of listed companies and identifying the underlying driving factors remains a significant challenge. This paper proposes a Transformer-based multi-source heterogeneous data fusion model, MSformer, which analyzes diverse data, including financial reports, news, social media comments, and government announcements. It categorizes the data into three types: time-series structured data, time-series structured mapped data, and textual data. The model enhances feature extraction using the Spatial Frequency-coordinated Attention Mechanism (SFHA) and employs Support Vector Regression (SVR) for prediction. Experimental results show that MSformer outperforms other advanced models, achieving an outstanding 87.4% multi-class accuracy and 0.517 average prediction error, proving its effectiveness and advantage in ESG prediction.