Forecasting Gold–Brent–U.S. Dollar Return Connectedness: Persistence Versus Macro-Financial Information

This study investigates whether U.S. macro-financial indicators improve one-step-ahead forecasts of return connectedness among gold, Brent crude oil, and the U.S. Dollar Index beyond autoregressive persistence. We estimate a time-varying Total Connectedness Index (TCI) using a TVP-VAR model with a 10-step generalized forecast-error variance decomposition and historically calibrated initialization. The U.S. Consumer Price Inflation is in line with the scheduled release dates, while predictors without historical data are not considered for the main analysis. Six forecasting models are evaluated using chronological validation and expanding-window estimation over 374 out-of-sample observations (8 May 2024–10 March 2026). The AR(2) benchmark achieves the lowest root mean squared error (RMSE=0.4898, out-of-sample R2=0.9768), compared with 0.4908 for macro-financially augmented linear regression and 0.5313 for XGBoost combining autoregressive and external predictors. The augmented XGBoost model produces 17.65% higher mean squared forecast loss than AR(2) (Holm-adjusted Diebold–Mariano p=2.10×10−5). Rolling-origin SHAP analysis attributes 94.91% of aggregate mean absolute feature attribution to the two lagged TCI predictors, indicating that historical connectedness dominates the attribution structure of the fitted nonlinear forecasting model. Exploratory inflation-regime comparisons do not provide robust evidence of a discrete 3% threshold. Within the evaluated three-market connectedness target, selected macro-financial predictors, one-step forecasting horizon, and 374-observation test period, none of the alternative models improved upon the AR(2) benchmark. These findings highlight the importance of historically consistent target construction, information availability, and persistence-aware forecasting benchmarks. They do not establish that macro-financial information generally lacks predictive value or that nonlinear forecasting models are inherently inferior.