Forecasting Oil and Gold Prices with Regime and Structural Demand Features

Abstract

This study investigates one-step-ahead forecasting of West Texas Intermediate (WTI) crude oil and gold prices over 2018–2025, with out-of-sample evaluation from 2022 to mid-2025. Linear Regression, XGBoost, and Random Forest are compared using feature sets that include cross-market signals, regime indicators, and global electric vehicle (EV) adoption as a structural demand proxy. Both log-return and price-level targets are examined. While return-based modeling is commonly used to address non-stationarity, it yields limited practical gains. In contrast, price-level evaluation reveals more stable and interpretable performance differences across models. XGBoost achieves the lowest errors, with a price-level MAPE of 1.82 percent for oil and 0.74 percent for gold. EV adoption does not improve short-horizon prediction. Errors increase under higher market stress, and residual analysis indicates that remaining volatility dynamics are better captured by GARCH filtering.