Copper price movement prediction using recurrent neural networks and ensemble averaging

The motivation for this paper is to investigate the use of three promising types of recurrent neural networks (RNNs), i.e., the long short-term memory (LSTM) network, the bidirectional LSTM (BiLSTM), and gated recurrent unit (GRU), when applied to the task of copper price prediction. This is done by deploying these RNNs with adequate data slicing and augmentation procedures to extract useful information from both domestic and international copper-related market indices for forecasting out-of-sample copper price movements. These RNN models are then empirically tested under various input window lengths, with the memory-free ANN model serving as a benchmark. The results show that the RNN models with memory units are superior to the memory-free ANN model, but the choice of a longer input window length may not produce a better prediction performance. To optimize the prediction results, the RNN models with relatively low forecasting errors are further combined using various ensemble averaging approaches (i.e., AVG or OLS, with the former being simpler) to integrate different model forecasts. Empirical findings show that the best ensemble prediction model is formed by combining just two RNNs (LSTM and BiLSTM) with shorter input window length, through the simpler AVG approach. Our results therefore suggest that in a sense, the simpler is better philosophy should apply for the deployment of RNN models for the task of copper price prediction.

Automated summary

The motivation of this study is to investigate the use of three promising types of recurrent neural networks (RNNs) for copper price prediction. The results show that RNN models with memory units outperform memory-free ANN models, and longer input window lengths may not necessarily result in better prediction performance. By combining two RNNs (LSTM and BiLSTM) with shorter input window lengths using a simpler averaging approach, the best ensemble prediction model can be obtained.