Multi-Relational Graph Convolution Network Prediction of Climate Change Impact on Firms

The paper investigates whether one can learn about the degree of climate risk exposure of firms by extracting information about their fundamentals and network connections and applying innovative graph-based artificial intelligence (AI) methodologies. We tackle this problem by creating a multi-relational financial network (MRFN) dataset and devise novel variations of graph convolutional networks (GCN), tailored to predict the risk of firms in a semi-supervised fashion. To the best of our knowledge, the MRFN dataset built reflective of prowess of financial analysts and market sentiment of investors is a first of its kind, defined as two-layered network with node level financial features and risk labels. A multi-layered nature of MRFN motivates two formulations of multi-relational GCN (MR-GCN), namely a GCN via network aggregation and a GCN via supra-graph. We show that the proposed MR-GCN models outperform the conventional uni-relational GCN models in terms of better classification accuracy for climate risk prediction by approximately $35\%$. We also find that network sparsity can be leveraged to improve the risk exposure prediction accuracy. While the main emphasis is on climate risk prediction of firms, we demonstrate the efficacy and replicability of the proposed network learning framework on another pivotal financial risk assessment problem, the credit risk prediction of firms. The MRFN benchmark datasets and graph-based AI methodologies, developed in this paper, can serve as an essential supplement for the application of network learning for financial risk assessment problems