Forecasting Food Inflation in Real Time with Tabular Foundation Models

Food-price inflation is volatile and difficult to forecast, especially around structural breaks where historical relationships cease to hold. We build a monthly forecasting pipeline using public data from FRED to predict U.S. Food CPI dynamics from 1967--2025, evaluating models with strict time-based splits and publication-lag rules to reduce look-ahead bias. On a challenging post-shock test window (Dec. 2023--Dec. 2025), TabPFN is the only method with positive out-of-sample skill (MAE 1.16; $R^2$ 0.69), while all other models yield negative $R^2$ values. Feature-importance diagnostics reveal that failed models rely on temporal trend extrapolation rather than stable economic covariates, consistent with a regime shift. These results provide evidence that tabular foundation models can be unusually robust on small, low-frequency macro-financial datasets where standard ML pipelines break down.