Carbon Credit Evaluation for Green Finance Based on Multimodal Data and Dynamic Ratings

Abstract

This study aims to construct a carbon emission credit evaluation system that meets the development requirements of green finance and ensures high scientific validity and accuracy. By integrating both structured and unstructured multimodal data—including corporate Environmental, Social, and Governance (ESG) reports, financial information, policy response data and news sentiment—this study adopts a joint modeling approach based on Graph Convolutional Networks (GCN) and Light Gradient Boosting Machine (LightGBM) to assess the carbon emission credit levels of enterprises. The dataset is obtained from the China Carbon Exchange, the Green Bond Public Information Platform, and the Wind ESG Rating Database, covering data from 1,824 enterprises between 2017 and 2022. The experimental results show that the proposed method achieves an accuracy of 86.1% in predicting the probability of corporate carbon violations, representing an improvement of 12.7% over the traditional logistic regression model. Moreover, this system can dynamically generate highly interpretable rating reports at the industry level, showing broad application prospects in key areas of green finance, such as green credit approval and the design of carbon asset securitization products.