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

Area of Science:

• Quantitative Finance
• Machine Learning
• Econometrics

Background:

• Accurate commodity price forecasting is crucial for financial markets.
• Deep learning models offer potential for improved financial time-series prediction.
• The impact of architectural complexity on deep learning forecasting accuracy remains an active research area.

Purpose of the Study:

• To investigate whether increased architectural complexity in deep learning models enhances financial forecasting accuracy.
• To compare the performance of various deep learning architectures against traditional econometric and machine learning models.
• To evaluate model generalizability across different commodities (copper, aluminum, zinc).

Main Methods:

• Utilized daily spot price data from the Shanghai Metals Market (Jan 2015 - Sep 2025).
• Applied a standardized preprocessing pipeline including z-score normalization and sliding window techniques.
• Systematically evaluated eighteen models: GRU, LSTM, CNN-BiLSTM-Attention hybrids, ARIMA, GARCH, Random Forest, XGBoost, and Transformer.

Main Results:

• The standard Gated Recurrent Unit (GRU) model demonstrated the lowest error rates (MAE=1032.85) and highest R-squared (0.907) on copper data.
• GRU models showed strong generalizability on aluminum (MAE=167.51, R-squared=0.918) and zinc (MAE=254.23, R-squared=0.952) datasets.
• Ablation studies revealed that adding complexity (attention, bidirectional layers, convolutional modules) decreased predictive accuracy, with significant differences confirmed by the Diebold-Mariano test (p < 0.05).

Conclusions:

• Unnecessary architectural complexity in deep learning models can hinder financial forecasting performance.
• Simpler, robust models like the standard GRU are effective for commodity price prediction.
• Findings advocate for parsimonious model selection in financial time-series analysis.