Text mining in climate finance

Abstract

This thesis is to explore the relationship between climate risk and stock returns.

In the first chapter, I develop a quantitative dynamic general equilibrium model to explore how climate regulatory risk is reflected in cross-sectional asset pricing. The representative household has preferences for low carbon. Greener environmental preference makes the market price of climate policy risk more negative conditional on a low elasticity of substitution between green and brown capital. The quantitative implications of the model can rationalize the recent empirical evidence on positive price of carbon risk. Finally, using textual analysis to measure transition risk from 10K filings, the paper shows that lower transition risk-exposed firms carry a 8.4\% annualized risk premium. Consistent with the model, the price of transition risk tends to be negative when climate awareness is high.

In the second chapter, we propose a measure of firm-level climate regulatory exposure based on 10-K filings. Using the 2016 Trump election as an exogenous shock to perceived climate regulatory risks, we identify a positive effect on stock returns for firms with higher climate regulatory exposures; they experience economically and statistically significant higher cumulative returns post-election. In the year following the election, firms with higher climate regulatory exposure experience higher carbon emissions and lower investor attention. Both findings indicate that, post-election, investors become less concerned with climate regulatory risks. Results are robust to physical climate, trade, tax and oil price exposures.

In the third chapter, We develop a machine learning framework to extract the text-implied climate risk in the SEC 10K filings for the text-implied price of risk estimation. We exploit both supervised and unsupervised learning algorithms in NLP to accomplish this task. The proposed risk exposure is testable within the Fama-MacBeth framework and has statistical sufficiency. Empirically, this technique could generate a monthly abnormal return on climate risk from 1.2\% to 2.1\%, depending on the specification of the benchmark model. The pricing finding is robust at the firm level for the Fama-MacBeth regression. Additionally, we find that the sufficient text-implied risk exposure is consistent with recent empirical evidence with sensible external validation.