Learning for Inflation Forecasting with Dynamic Feature Spaces

Inflation forecasting becomes especially challenging during structural breaks, such as the COVID-19 pandemic, when historical patterns are disrupted. We evaluate lightweight ML strategies for adapting to dynamic feature spaces, including rolling-window training, interpolation, and reversible normalization. Using U.S. CPI data and pandemic-specific features (cases, deaths, vaccines, hospitalizations), we compare traditional econometric models, tree ensembles, and deep sequential architectures. We find that ensemble methods deliver the most accurate forecasts (RMSE 0.36 with interpolated training data, 0.61 without), while normalization layers improve neural networks' robustness under distribution shifts. These results show that simple, low-cost adaptation methods sustain forecast reliability during shocks, aiding policymakers.