Machine Learning for Inflation Forecasting: Can High-Dimensional Macro Indicators Beat the Phillips Curve?

Abstract

Accurate inflation forecasting is critical for monetary policy and investment decisions. The Phillips Curve has been the dominant framework, yet its predictive accuracy has deteriorated since the 2008 Global Financial Crisis. This study compares five forecasting models—Autoregressive (AR), Phillips Curve, Elastic Net, Random Forest, and LSTM—using 45 engineered features from 46 FRED series spanning 705 months (1967–2025). Using a chronological 70/30 split and 48-window expanding evaluation, Elastic Net achieves an RMSE of 0.4117, a 46.79% improvement over the Phillips Curve (RMSE $=0.7737)$, confirmed by the Diebold–Mariano test $(\mathbf{D M}=3.404, p=0.0007)$. Random Forest and LSTM perform worse than the Phillips Curve baseline, suggesting that non-linear complexity does not automatically improve inflation forecasting. Granger causality tests reveal that the Unemployment Rate does not Granger-cause inflation $(p=0.533)$, while the Federal Funds Rate $(p<0.001)$ and Industrial Production $(p=0.004)$ do. Elastic Net is the only model with Theil's $U<1(\mathbf{0. 9 2 6})$, meaning it alone beats a naive random-walk forecast.