Financial Resilience Under a Major Gold-Price Regime Shift: A Benchmark-Based Assessment of Machine Learning, Deep Learning, and Econometric Models

Machine-learning and deep-learning models are increasingly applied to financial-market forecasting, yet their reported accuracy is often measured in price levels and rarely tested against a stringent benchmark under changing market conditions. It therefore remains unclear whether added model complexity delivers genuine, resilient predictive value or merely reflects the persistence of the underlying price series. This study addresses that gap by evaluating one-day-ahead gold-price forecasts within a transparent, leakage-resistant, benchmark-centred framework during a major structural regime shift. Using daily COMEX gold futures (2014–2026), we compare Ridge Regression, SVR-RBF, Random Forest, GARCH(1,1), and a bidirectional LSTM against a naïve random walk, with ARIMA and exponential-smoothing benchmarks, using strictly chronological splitting, training-only preprocessing, a return-space target, and two-sided and one-sided Diebold–Mariano tests. Structural-break analysis confirms a pronounced regime shift, with the gold price rising from roughly USD 1051 to USD 5318/oz and volatility breaks falling inside the out-of-sample period. All models attain near-unity price-level R2 but non-positive return-space R2, and none of the flexible models significantly outperforms the naïve benchmark; only a heavily regularised linear model shows a marginal one-sided edge attributable to gold’s upward drift rather than to learned structure, while GARCH’s contribution lies in calibrated uncertainty rather than point accuracy. Increasing model complexity did not yield resilient incremental predictive value, and the study offers a reproducible template for benchmark-based evaluation of forecasting systems.