Abstract This paper introduces a novel technique for inflation forecasting that leverages density-based model averaging, utilizing Wasserstein barycenter and optimal transport metrics. By preserving the geometric properties of forecast distributions, this method provides an accurate and interpretable average of predictions from various machine learning models, from which point forecasts and confidence intervals can be derived. Extensive simulation studies demonstrate the method’s superior performance in capturing complex distributional features, especially in spatially distinct distributions. Additionally, the technique is applied to real-world inflation forecasting using historical data from the FRED-MD database. Empirical results show that this method significantly improves forecasting accuracy by providing nuanced density forecasts. These comprehensive forecasts offer valuable insights into future inflation trends and associated uncertainties, underscoring the method’s practical utility in economic forecasting. Comparative analysis with traditional methods highlights the potential advantages and broad applicability of this innovative approach.
Alessandro Spelta (Thu,) studied this question.