Predicting Corporate Carbon Footprints for Climate Finance Risk Analyses: A Machine Learning Approach

Abstract

This paper uses machine learning to improve the prediction of corporate emissions so that financial regulators and investors can make better decisions about climate transition risk. The need for predictions arises because only a subset of global companies report emissions. The novelty is to use machine learning rather than the conventional regression approaches and naïve models implemented by data providers. Our best-performing model is a two-step framework that applies a Meta-Elastic Net learner to combine predictions from multiple base-learners. It results in an accuracy gain based on mean absolute error of up to 30% as compared with the existing models. We find that prediction accuracy can be further improved by incorporating additional predictors (energy data) and additional firm disclosures in particular sectors (utilities and real estate) and regions (Asia and Latin America). This provides an indication of where policymakers should concentrate their efforts for greater disclosure.

Keywords: climate change; corporate carbon footprints; machine learning

JEL Classification: G17; Q51; Q52; Q54