Predicting Copper Prices with Cutting-Edge Neural Networks

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Aug 4 2024

In a paper published in the journal Mathematics, researchers developed and tested a one-dimensional convolutional neural network (1D-CNN) to accurately predict global copper prices. They used extensive data from November 1991 to May 2023 to train and validate the model, outperforming traditional forecasting methods.

The study showed that CNN delivered dependable forecasts for copper prices extending to 2027. The research underscored the model's potential as a valuable tool for policymakers, offering insights into managing price volatilities and informing strategic decisions in the energy sector.

Background

Past work has highlighted the critical role of copper in energy production and its significance for sustainable development. Copper's high thermal and electrical conductivity make it essential for energy infrastructure, such as power transmission lines and electric motors.

The global demand for copper, driven by its use in various industries, has led to complex price fluctuations, impacting economic sectors and challenging traditional forecasting methods. These factors underscore the importance of accurate copper price prediction for effective resource management and policy planning.

Copper Forecasting

Forecasting copper prices is crucial due to copper's pivotal role in various industries, particularly energy production and telecommunications. Conventional models like autoregressive integrated moving averages (ARIMA) and generalized autoregressive conditional heteroskedasticity (GARCH) often find it challenging to capture the complex and non-linear fluctuations in copper prices.

A 1D-CNN was developed to address these limitations, offering superior accuracy by recognizing intricate patterns in time-series data. This innovative approach, tailored specifically to copper price forecasting and enhanced with diverse predictor variables, demonstrates significant advancements over traditional methods, providing valuable insights for mining companies, policymakers, and investors. The study represents a pioneering contribution to copper price forecasting using deep learning (DL), setting a new benchmark for future research.

CNN Architecture

A convolutional neural network (CNN) generally comprises multiple layers, including an input layer, convolutional layers, pooling layers, a fully connected layer, an activation function layer, a flattening layer, and an output layer. While the convolution process is linear, these feature maps are transformed through an activation function layer to introduce nonlinearity into the network. Sigmoid, rectified linear units (ReLUs), and exponential linear units (ELUs) are examples of common activation functions.

The pooling layer simplifies the feature maps generated by the convolution layers to minimize overfitting and improve computing efficiency. The network can collect and analyze the features for the final output because of the fully connected layer, a multi-layer perceptron (MLP), which connects every neuron to every other neuron in the layer before it.

This architecture allows CNNs to capture intricate patterns in data effectively, making them highly suitable for tasks like predicting copper prices. This design enables CNNs to capture intricate patterns in data, making them highly effective for predicting copper prices. This structure allows CNNs to effectively identify complex patterns in data, making them particularly adept at tasks such as forecasting copper prices.

CNN Layer Functions

A CNN's fully connected layer (FC layer) linearly maps input vectors to another, with neurons between adjacent layers paired. In CNNs, a convolutional layer applies filters across input spectra to generate feature maps, with padding used to maintain the input-output size. Classical preprocessing techniques, like detrend and Savitzky–Golay derivatives, can be replaced by convolutional layers that optimize filter selection through training, particularly in spectroscopic calibration.

Activation functions such as sigmoid, tanh, ReLUs, and ELUs infuse the network with non-linear properties. ReLUs and ELUs address vanishing gradients and "dead neuron" issues, although ELUs may be slower in large networks.

CNN Accuracy

This study utilized a CNN to estimate global copper prices based on various input variables, including silver, aluminum, nickel, gold, iron, coal, and crude oil prices. The model was developed and evaluated using monthly data, allocating 80% for training and 20% for testing, with forecasts aggregated into annual predictions for comparison. The CNN demonstrated remarkable accuracy, with an R-value of 0.98.

The team observed significant error reduction over 1000 epochs, with the model attaining a mean squared error (MSE) of 0.0337, a mean absolute error (MAE) of 0.119, and a root mean squared error (RMSE) of 0.1835. Performance comparisons with forecasts from the International Monetary Fund (IMF) and the International Society of Automation (ISA) highlighted CNN's effectiveness in predicting future copper prices. The research also evaluated the CNN technique against several alternative methods, showing that the CNN provided superior predictive accuracy.

Conclusion

To sum up, this study addressed the gap in copper price forecasting by using a 1D-CNN, incorporating factors such as silver, aluminum, and crude oil prices. The CNN model was assessed through MAE, MSE, and RMSE metrics, showcasing its high dependability and performance.

The findings have significant implications for policymakers and energy system developers, aiding in creating robust energy plans and enabling informed economic interventions. This pioneering method set a new precedent for analyzing energy policies and their influence on the energy industry.