Does Earnings-Call Climate Exposure Improve Machine-Learning Forecasts of Stock Downside Risk? Out-of-Sample Evidence from U.S. Firms (Cureus Journal of Computer Science.)

Summary

• The study evaluates out-of-sample forecasts of downside risk for 1,792 U.S.-listed non-financial firms across 15,521 firm-years from 2012 to 2024.p.1
• Adding earnings-call climate exposure moves the random forest's out-of-sample R² from 10.25% to 10.30% with a Diebold-Mariano p = 0.665.p.1
• Evaluating forecasts against 100 placebo versions demonstrates that real climate data performs no better than randomly reassigned data, yielding a permutation p = 0.168.p.1
• Excluding the single year following the COVID-19 shock reduces the pooled out-of-sample R² to −3.3% for the random forest and −8.4% for the linear model.p.1
• The random forest model lowers the mean squared error by −0.01053 compared to the linear model, yielding a Diebold-Mariano p-value of 0.022.p.15

Key insights

1. Adding firm-level climate change exposure to conventional risk predictors does not materially improve out-of-sample forecasts of stock downside risk.p.1
2. The random forest model yields an out-of-sample R² of 10.25% without climate variables and 10.30% with climate variables.p.1
3. Models using real climate data perform no better than 100 placebo versions with shuffled climate variables, producing a permutation p-value of 0.168.p.1
4. Most of the random forest's predictive advantage over a naive past-risk forecast is concentrated in the year following the COVID-19 shock.p.1
5. Earnings-call climate exposure measures are highly sparse, with regulatory exposure being zero in 81.0% of firm-years and physical exposure zero in 92.7%.p.13
6. Past volatility is the dominant predictor in the climate-informed random forest, registering a mean absolute SHAP value of 0.1289.p.18
7. Portfolios built with and without climate forecasts show nearly identical realized risk, with monthly expected shortfall reaching 8.96% and 8.97%.p.19