A Transformer-Based Framework for Precious Metal Price Prediction With Explicit Cross-Metal Dependency Modeling

Abstract

Precious metals serve as critical investment assets and industrial commodities in global financial markets. Accurate price forecasting is challenged by their strong nonlinearity, long-range temporal dependencies, and complex inter-metal linkages. Traditional econometric models and existing deep learning approaches often overlook explicit cross-metal interactions, leading to suboptimal multi-metal prediction performance. This paper proposes an improved Transformer architecture for multi-precious metal price forecasting, which jointly models individual temporal dynamics and cross-metal correlations in a unified framework. The method integrates metal-specific embedding, hybrid multi-head attention for temporal and cross-metal modeling, a cross-attention mechanism for leader–follower relationship capture, and an Informer-based sparse attention for long-sequence efficiency. A hybrid loss function combining Huber regression and inter-metal correlation constraints is designed to preserve realistic market co-movements. Experiments on daily futures data of eight precious metals from 2005 to 2025 show that the proposed method outperforms baseline models including linear regression, gradient boosting, SVR, LSTM, GRU, and CNN across multiple prediction horizons. The results validate that explicit cross-metal dependency modeling significantly enhances forecasting accuracy in interconnected precious metal markets.