Measuring Corporate Risk Using Large Language Model Embeddings: Evidence on Corporate Climate Risk and Supply Chain Restructuring

Abstract

We propose a novel text-based approach to measuring corporate climate risk that combines large language model (LLM) sentence embeddings with supervised machine learning. We find that this combined approach improves the precision of climate-related sentence classification by 23% relative to predefined keyword methods and by 26% compared to a recent LLM agent-based classifier. Using the resulting corporate risk measure, we further explore how firms react to supplier climate risk. We find that after controlling for firm- and supplier-level non-climate risk, supplier climate risk exposure increases the likelihood of terminating existing suppliers by 5.3% and adding new suppliers by 6.3%. Our analysis further suggests that such restructuring is strategic. First, the restructured supply networks exhibit greater geographic diversity, indicating firms’ efforts to mitigate future climate-related disruptions by diversifying supplier locations. Second, firms prioritize adding new suppliers that produce more unique products or occupy more central positions in the supply chain network. We also document that such restructuring is most pronounced for supplier backward-looking climate risk and extreme physical climate risk. Firms with higher operational efficiency and greater product differentiation respond more aggressively to supplier climate risk.