Copper Price Forecasting of COMEX Copper Based on Machine Learning: A Comparative Study of LSTM, ARIMA, and SVR Models
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Ten years of COMEX front-month closes (2 506 days) are rolled forward one observation at a time to produce 1 501 out-of-sample one-step forecasts. A single-layer LSTM with 64 hidden units and dropout 0.15 beats the in-house ARIMA(0,1,0) benchmark by 1.5 c/lb RMSE and 19 percentage points in directional accuracy; ε-SVR with an RBF kernel finishes between the two. The neural network keeps prediction error below 10 c/lb even during the 2020-Q2 volatility spike, a period when the linear model exceeds 15 c/lb. Gate-activation traces show that the forget unit discards history once rolling kurtosis exceeds 5.5, letting the cell state adapt to fat-tail shocks without external regressors. The specification satisfies the treasury desk’s runtime limit (38 s per retraining) and 65 % hit-ratio veto, cutting monthly copper VaR by 12 %.
Fields of Study & Subjects
Stock Market Forecasting Methods
Forecasting Techniques and Applications
Financial Risk and Volatility Modeling
Publication details
- Journal / source
- IEEE International Conference on Power, Electronics and Green Energy (ICPEGE)
- Pages
- 1237–1241
- Publication date
- 28 January 2026
- Language
- EN