Cross-Industry Heterogeneity in ESG Event Pricing: Evidence from Short-Horizon Abnormal Returns

Replication package for "Cross-Industry Heterogeneity in ESG Event Pricing: Evidence from Short-Horizon Abnormal Returns."

This package provides code and data to reproduce all empirical results in the paper. The study examines how industry membership shapes short-horizon cumulative abnormal returns (CARs) following ESG-related news events for S&P 500 firms over 2009–2020. The central finding is that industry fixed effects improve out-of-sample predictive R² by 0.09–0.17 relative to specifications without industry controls, operating through a return dispersion mechanism rather than a shift in median direction. Kruskal-Wallis tests fail to reject equality of industry medians at all horizons, while Levene tests reject equality of variances at p < 0.001. Structured materiality features derived from SASB topic overlap and FinBERT pillar probabilities add little incremental explanatory power beyond industry controls.

The included dataset (esg_events_sasb.csv) contains 675 ESG events identified from approximately 1.4 million Benzinga news headlines, scored using FinBERT-ESG for pillar classification, DistilBERT for sentiment, and BART-Large-MNLI for severity, disclosure mismatch, and SASB topic indicators. After cleaning and market model estimation, the analytical sample comprises 598 firm-event observations across 153 S&P 500 firms and 48 SASB industries.

The replication script (replicate.py) reproduces the full pipeline from event cleaning through Random Forest cross-validation and produces all figures and intermediate summary statistics used in the paper. All reference datasets (S&P 500 universe, SASB materiality map, GICS-to-SICS industry mapping) are included. The raw Benzinga headline data must be downloaded separately from Kaggle (link provided in README.md) and is only required to re-run the NLP ingestion stage from scratch.

* Version 2 fixes a date-alignment bug for events that occur on non-trading days. Version 3 updates the repository description to reflect the final analytical sample.