Multivariate LSTM-Based Intraday Gold Price Prediction with Rolling Time Series Validation

Abstrak

Projecting XAUUSD (gold vs. US dollar) prices on a one-hour interval is particularly challenging due to the market's dynamic and nuanced character. To address short-term financial forecasting, an advanced deep learning methodology utilizing Long Short-Term Memory (LSTM) models was employed. Historical XAUUSD data for 2024 was resampled to hourly intervals and supplemented with SMA, RSI, MACD, and Bollinger Bands to understand the market structure better. An LSTM model was developed using open, high, low, and close prices as inputs, with the close price designated as the output target. Data normalization was performed via MinMaxScaler. The model was validated using Time Series Cross-Validation (TSCV) with a rolling origin expanding window over five splits—a sophisticated method for evaluating performance. The results demonstrated the LSTM model's capability, showcasing a mean RMSE of 9.9574, a mean MAE of 7.4411, an R² score of 0.9535, and a remarkably low MAPE of 0.3009%. These findings indicate the advanced model effectively predicts intraday prices, even while grappling with complex and nonlinear patterns, offering a powerful instrument for trading professionals and researchers to cut through market noise.