Distributional Forests for Short-Term Density Predictions of U.S. Inflation

Recent work by Medeiros et al. (2019, Journal of Business & Economic Statistics) shows that point forecasts of the random forest machine learning algorithm systematically outperform well-established benchmarks at predicting U.S. inflation. This article extends their work from point to density forecasts derived from large-scale datasets by means of the novel distributional forest algorithm of Schlosser et al. (2019, The Annals of Applied Statistics), which allows us to simultaneously account for locally evolving means, time-varying volatilities, nonlinearities as well as heavy-tailed densities – all of which constitute stylized characteristics highlighted in the recent literature. We show that distributional forests typically achieve superior point and density predictions over random walk, auto-regressive and stochastic volatility benchmarks