Comparative Evaluation of Deep Learning Model Complexity for Forecasting Non-Ferrous Metal Prices

This study examines whether increasing architectural complexity improves forecasting accuracy in deep learning-based financial models. Using daily spot price data from the Shanghai Metals Market for copper (Cu), aluminum (Al), and zinc (Zn) spanning January 2015 to September 2025, a standardized preprocessing pipeline was applied, including z-score normalization and sliding window sequence construction (window length = 30, forecast horizon = 1). A total of eighteen models were systematically evaluated, including gated recurrent units (GRUs), long short-term memory (LSTM) networks, convolutional neural network-bidirectional LSTM-attention hybrids (CNN-BiLSTM-Attention), as well as traditional econometric models (autoregressive integrated moving average and generalized autoregressive conditional heteroskedasticity), machine learning models (random forest and extreme gradient boosting), and a Transformer-based model. All deep learning models were trained exclusively on Cu data and evaluated on independent Al and Zn datasets to assess generalizability. Results show that the standard GRU model achieves the lowest error rates (mean absolute error [MAE] = 1032.85; root mean square error = 1344.30) and highest explanatory power (coefficient of determination [R²] = 0.907) on the Cu test set, while also performing strongly on Al (MAE = 167.51, R² = 0.918) and Zn (MAE = 254.23, R² = 0.952). Ablation analysis demonstrates that adding architectural components such as attention mechanisms, bidirectional layers, and convolutional modules reduces predictive accuracy. Statistical testing using the Diebold-Mariano test indicates that most performance differences are significant (p < 0.05). These findings highlight the limitations of unnecessary model complexity and support the use of simpler, robust approaches for commodity price forecasting.