Credit rating prediction using multi-modal temporal neural networks

Abstract

This research explores the integration of Recurrent Neural Networks (RNNs) and Graph Neural Networks (GNNs) to enhance corporate credit rating prediction. Corporate credit risk, characterized by the dynamic financial interactions and dependencies among firms, is effectively modeled using a multi-modal approach that merges graph-based and sequential data processing. This research employs Graph Convolutional Networks (GCNs) and various forms of RNNs, including Long Short-Term Memory (LSTM) units, Gated Recurrent Units (GRUs), and Transformers, to analyze corporate financial health both statically and dynamically. The models are trained using data from 318 US companies, synthesizing numerical data from quarterly financial statements and graph data constructed from stock market interactions to capture intricate patterns of corporate financial activities. The effectiveness of these multi-modal models is measured on both the 318 companies it was trained on, in a transductive way, as well as inductively 112 "unseen" companies that where excluded during model training. The superior performance in both seen and unseen data scenarios demonstrates robust generalization capabilities, essential for practical deployment in financial risk assessment.