Forecasting Copper Price with Multi-view Graph Transformer and Fractional Brownian Motion-Based Data Augmentation

Abstract

Copper price forecasting is crucial for both investors and governments due to its significant economic impact. Recently, machine learning techniques have been widely employed to construct copper price forecasting models, demonstrating high forecasting accuracy. However, there are two main limitations in these models: (1) the lack of ability to capture the non-Euclidean relationships among numerous features; and (2) using purely data-driven algorithms, which lack tractability and physical effectiveness. To address these challenges, this study proposes a multi-view graph transformer (MVGT) model for 1-month ahead copper price forecasting. MVGT integrates a parametric fractional Brownian motion module, which provides conditional expectations of future copper prices for data augmentation. Moreover, to comprehensively capture the non-Euclidean structure of copper features, MVGT introduces five graph generation methods. Furthermore, a multi-view graph transformers model is designed to provide structural copper feature embeddings, and an attention-based multi-view fusion mechanism is developed to enable the MVGT to comprehensively understand market trends while focusing on the most influential views. Experimental results on the COMEX and LME datasets demonstrate that MVGT outperforms baseline models in terms of training efficiency, forecasting accuracy, and generalization.