Gold Price Direction Prediction Using Machine Learning with Walk-Forward Validation

This study presents a robust machine learning framework for gold price direction prediction using non-overlapping time-series data and walk-forward validation. Unlike prior studies that report overly optimistic results due to temporal leakage and overlapping samples, this work ensures a realistic evaluation by constructing independent observations and employing a rolling validation strategy. A set of technical indicators, including RSI, MACD, and Bollinger Bands, is used as predictive features. Multiple models, including Logistic Regression, Random Forest, and Support Vector Machines, are evaluated against a persistence-based baseline. Experimental results show that all machine learning models outperform the baseline, achieving an average accuracy of approximately 69% and balanced accuracy above 70%. Additionally, the Random Forest model achieves an AUC score of 0.786, indicating strong discriminative capability. Feature importance analysis highlights the significance of momentum and volatility indicators in predicting market direction. The findings demonstrate that, while financial markets remain inherently noisy, carefully designed validation strategies can yield reliable and interpretable predictive performance. This study contributes to the field by providing a realistic and reproducible framework that enhances the reliability of financial forecasting models and supports more informed decision-making in practical investment scenarios.