| Abstract: |
Companies are subject to stringent expectations in terms of social responsibility, particularly in managing risks associated with their environmental, social, and governance (ESG) practices. These practices are evaluated using ESG risk scores. Traditionally, ESG risk scores are generated by firms like Sustainalytics and MSCI, which primarily focus on larger corporations. Consequently, entities investing in smaller companies, such as venture capital firms, private equity firms, and individual investors, face a challenging and resource-intensive process for initial risk assessment. However, our research has uncovered a novel approach through the application of machine learning techniques and the use of multimodal large language models based on publicly released company reports. This approach enables the prediction of ESG risk scores with an accuracy of 68.09%, offering a viable tool for preliminary analysis. Significantly, this research introduces a pioneering framework that utilizes a new architecture for analyzing ESG practices, transforming the traditional assessment process for both large and small companies alike. Our research shows high accuracy in predicting risk assessments and simplifies the evaluation process. Nonetheless, there is potential for enhancing this accuracy through further refinement of the models, improvements in data extraction, and continued exploration of additional modeling techniques. |