Forecasting in the Fog: Real-Time versus Revised-Data Evidence on Machine Learning's Edge over the Phillips Curve
ML inflation forecasts are almost universally trained on fully revised data, even though real-time forecasters never have such data, and reported feature importances are typically computed in-sample, conflating predictive relevance with retrospective fit. This paper asks whether the ML advantage over the Phillips curve documented in Agyekum (2026) survives when models are trained and evaluated on real-time (ALFRED) vintages rather than revised series, and whether SHAP feature-importance rankings are an artifact of in-sample estimation. Using 2000-2026 U.S. data on unemployment, CPI and PCE inflation, payrolls, real GDP, and the 10-year-2-year Treasury spread, vintage-consistent panels are built for four traditional models (random walk, AR(1), Phillips curve, ADL-OLS) and four ML models (Random Forest, Gradient Boosting, Elastic Net, SVR), re-estimated recursively at 3-, 6-, and 12-month horizons (208, 206, 204 forecasts). Real-time/revised accuracy differences are small and, apart from one exception at 6 months (Gradient Boosting vs. Phillips curve, DM = -1.671, p = 0.097), indistinguishable under Diebold-Mariano tests; Gradient Boosting alone shows consistent positive skill at longer horizons. The random walk remains a strong short-horizon benchmark, consistent with the puzzle in Agyekum et al. (2026) for exchange rates. Using walk-forward, out-of-sample SHAP, a Random Forest on revised data assigns dominant importance to PCE inflation (mean |SHAP| = 0.778, rank 1 of 9), while on real-time data it assigns PCE negligible importance (0.039, rank 6), relying instead on current CPI (0.834 vs. 0.223). This twenty-fold swing, larger than the in-sample estimate, is invisible to point-forecast metrics and shows the model's PCE reliance is substantially a hindsight artifact. An RSI summarizes the accuracy gap by model and horizon, with implications for auditing ML inflation forecasts.