Optimized Feature Selection and Self-Attention- Enhanced Deep Learning Model for Gold Price Prediction

Gold plays a pivotal role in financial markets, but economic volatility makes forecasting its price challenging. This paper presents a hybrid deep learning model integrating convolutional neural networks (CNN), gated recurrent units (GRU), and self-attention (SA) to capture complex patterns and improve forecasting accuracy. The model is trained on 15 years of historical data (January 11, 2010 – January 10, 2025) and includes opening, high, low, and closing prices with 15 technical indicators (TI). Pearson correlation helps identify essential feature selection, minimize noise, and improve efficiency. The model predicts closing prices for 1-day, 2-day, and 3-day periods, providing valuable short-term trading insights. Performance is evaluated using root mean square error (RMSE), mean absolute percentage error (MAPE), and R-squared (R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>). Results show that the hybrid approach outperforms traditional models like CNN, LSTM, GRU, and CNN-GRU. The results emphasize its utility as a reliable tool for forecasting gold prices in fluctuating markets. Within the 1-day time frame, the model achieves its optimal performance, accordingly exhibiting RMSE, MAPE, and R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> values of 0.985, 25.683, and 0.990.