A Novel Gated Fusion CNN-LSTM Model for Multi-Horizon Intraday Gold Price Forecasting

Abstract

Gold serves as a key inflation hedge and portfolio stabilizer, making accurate price forecasting essential for investors. Hourly gold prices exhibit pronounced non-linearity and microstructure noise that limit traditional econometric models, motivating a shift toward deep learning. Existing CNN-LSTM architectures cascade convolutional and recurrent layers sequentially, without a mechanism to reconcile their complementary representations. We propose a Dual-Branch CNN-LSTM architecture with Gated Fusion, combining a convolutional-recurrent deep branch with a parallel raw-input skip branch, adaptively merged by a learned gate inspired by the Gated Multimodal Unit. Input sequences are restructured via sliding windows across three forecast horizons (24→1, 48→2, and 72→3 hours). The model is validated on 37,140 hourly XAUUSDm observations (Exness, 2020-2026) against 1D-CNN, LSTM, and Sequential CNN-LSTM baselines, achieving the best or tied-best accuracy across all configurations. For the 24-hour horizon, MAE = 9.00 USD, R² = 0.9994, and DA = 52.6%, on the original USD scale. Diebold-Mariano tests confirm significant gains over the 1D-CNN and Sequential CNN-LSTM baselines (p < 0.001), while McNemar tests confirm directional-accuracy gains in 8 of 9 comparisons (p < 0.05). These results show a modest but robust improvement over prior CNN-LSTM designs, offering a reliable foundation for risk management pending economic backtesting.