A Machine Learning Based Anatomy of Firm Level Climate Risk Exposure
Abstract
We construct firm-level climate risk exposures by utilizing two natural language processingtechniques (LDA and word2vec) on quarterly earnings conference call transcripts.
This unsupervised learning method automatically generate five topics, all aligned with
popular concerns about climate change. We then conduct empirical analysis on one of
the topics that put high weight on words about natural disaster. This topic has a significant
negative association with firm's sales growth and profitability, indicating that our
measure exactly capture firm level disaster exposure. Moreover, firm with higher disaster
measure tend to earn higher stock returns in the future years. Appropriate long-short
portfolios based on this topic generates positive return, which cannot be explained by
common risk factors and other firm characteristics.