Qiguo Sun, Xibei Yang & Meiyu Zhong of School of Computer Science and Engineering, Jiangsu University of Science and Technology, Zhenjiang 212003, Jiangsu Province, China, have recently published a paper on copper price forecasting with multi-view graph transformer and fractional Brownian motion-based data augmentation in National Resources Research: https://link.springer.com/article/10.1007/s11053-024-10442-1
Copper is one of the most important materials in our global economy, widely used in construction, electronics, renewable energy, and transportation. Its price fluctuations can impact entire industries and even national economies. Accurately predicting copper prices is crucial for businesses, investors, and governments. This paper introduces a cutting-edge AI model that significantly improves the ability to forecast copper prices one month in advance.
The Challenge
Traditional methods, such as basic machine learning models, struggle to analyze the complex factors influencing copper prices. These include global economic indicators, supply-demand imbalances, and long-term price patterns. Two main issues arise:
- Complex Relationships: Most models fail to capture how multiple factors interconnect and influence prices.
- Physical Realism: Many purely data-driven models are effective but cannot explain why certain predictions are made.
The Solution: MVGT Model
The researchers created a new model called the Multi-View Graph Transformer (MVGT). This model integrates two key innovations:
- Graph-Based Analysis: Instead of treating features (like demand, inventory, or economic data) separately, the model builds a “graph” to represent relationships between these features. For example:
- Causal Graph: Shows which factors (like supply) influence others (like price).
- Similarity Graph: Connects factors with similar trends.
- Cointegration Graph: Identifies features that move together over time, like stock prices.
- Fractional Brownian Motion (fBm): This mathematical model accounts for long-term trends and price movements in copper. It provides a reliable “baseline” forecast that enhances the overall accuracy.
How It Works
The MVGT model combines the graph relationships and fBm insights to identify meaningful patterns in copper pricing data. An “attention mechanism” helps focus on the most important features while ignoring irrelevant noise. The model processes information from multiple perspectives (graphs) to form a well-rounded understanding of market trends.
The Results
The MVGT model was tested on two major datasets:
- COMEX (New York) and LME (London) copper prices.
The results showed that MVGT outperformed traditional methods like Artificial Neural Networks (ANN) and Long Short-Term Memory (LSTM) models. It was more accurate, trained faster, and could better predict market behavior.
For example, MVGT achieved an accuracy score of over 90% on both datasets, far surpassing older models. The ability to capture subtle relationships between features gave it a clear edge.
Why It Matters
- Investors: MVGT provides more reliable price forecasts, helping investors make better decisions.
- Businesses: Companies using copper can anticipate price changes and manage costs effectively.
- Policymakers: Governments can better plan for economic stability and resource management.
Looking Ahead
While MVGT is a significant advancement, it doesn’t yet account for unpredictable factors like political events or sudden policy changes. Future improvements might integrate large language models to analyze news and geopolitical developments, further enhancing predictions.
Photo by Bruna Fiscuk on Unsplash