Predicting commodity returns with climate variables: Statistical loss functions vs. economic value

We test whether large-scale climate indicators contribute to the predictability of commodity futures returns and generate economic value for investors. Using monthly data on fourteen commodities, we conduct a pseudo out-of-sample forecasting exercise based on a range of models that include automatic variable selection and nonlinear regime-switching specifications. Climate-related predictors rarely outperform standard benchmarks in terms of statistical forecast accuracy. Yet, when embedded in a portfolio choice problem, they deliver economically meaningful gains. These results illustrate a disconnect between statistical predictability and economic relevance. • Climate principal components add little to out-of-sample predictability. • Autoregressive benchmark outperforms stepwise and hidden Markov models. • Climate signals improve certainty equivalent returns in portfolio choice. • Statistical loss and economic value yield sharply different rankings. • Regime-switching models enhance gains from climate information.