Does Earnings-Call Climate Exposure Improve Machine-Learning Forecasts of Stock Downside Risk? Out-of-Sample Evidence from U.S. Firms

Climate-related risks are now widely regarded as financial risks, and machine learning is often proposed as a means of turning climate information into sharper risk assessments.Whether firm-level climate exposure helps forecast a stock's future risk, rather than merely correlating with it in-sample, has received much less attention.We examine this question for 1,792 U.S.-listed non-financial firms (15,521 firm-years, with portfolio formation dates from 2012 to 2024), combining published earnings-call measures of firm-level climate change exposure with freely available accounting and market data.Our forecast target is each firm's downside semi-deviation over the following year.We estimate ordinary least squares and random forest models without climate exposure, with it, and with placebo climate variables randomly reassigned among firms in the same year and industry.We evaluate the forecasts in out-of-sample tests over 2017-2024 (10,836 firm-years) and use them to form sector-neutral portfolios that exclude the stocks with the highest predicted risk.Climate exposure does not improve forecast accuracy: adding it moves the random forest's out-of-sample R² from 10.25% to 10.30% (Diebold-Mariano p = 0.665), and the model with real climate data performs no better than 100 placebo versions (permutation p = 0.168).Random forests are more accurate than linear models (p = 0.022), but this advantage is only marginally significant after a small-sample correction (p = 0.069), and most of their gain over a naive past-risk forecast arises in the year after the COVID-19 shock; excluding that year, the pooled out-of-sample R² is -3.3% for the random forest and -8.4% for the linear model.Portfolios formed with and without climate information have almost identical risk.For U.S.-listed firms and a one-year horizon, earnings-call climate exposure therefore adds little to forecasts of downside risk once past volatility and basic firm characteristics are known, and the contributions of climate data and of model flexibility should be assessed separately.