The Impact of Market Volatility Regimes on Gold Price Prediction Accuracy: A VIX-Based Machine Learning Approach

Abstract

This study analyzes the impact of market volatility regimes on gold price prediction accuracy using the VIX indicator and compares machine learning model performance across different market conditions. Daily data from September 2014 to November 2025 (2,773 observations) includes gold prices, VIX, DXY, and S&P 500. Volatility regimes are classified into Calm (VIX<15), Normal (15≤VIX<25), and Crisis (VIX≥25). Granger Causality tests validate predictive relationships, followed by a comparison of three models —ARIMA, LSTM, and GRU—at 1-day and 7-day horizons using walk-forward validation. Results show VIX change has the strongest predictive power (F-stat=9.676, p<0.001), followed by DXY and S&P 500. The GRU model performs better, with an RMSE of 0.98% and directional accuracy of 51.2%. Critical finding: accuracy varies substantially across regimes—Calm periods achieve RMSE of 0.61% (Dir.Acc=54.2%), while Crisis periods increase to 1.34% (Dir.Acc=47.3%). Short-term predictions (1-day, RMSE=0.67%) significantly outperform 7-day forecasts (RMSE=0.92%). Volatility regimes significantly influence the accuracy of gold predictions. GRU models excel during low-to-normal volatility but degrade during crises. Investors are advised to employ adaptive strategies with wider confidence intervals when the VIX is≥25. This research contributes a regime-aware forecasting framework for gold portfolio risk management.