Non-ferrous metal price forecasting with Recurrent Neural Networks : how do they perform when forecasting multiple timesteps ahead?
Master thesis
View/ Open
Date
2021Metadata
Show full item recordCollections
- Master's theses (RealTek) [1826]
Abstract
This thesis aims to forecast the daily price of aluminum, copper and zinc from the London Metal Exchange five days ahead based on prices the previous five days using different recurrent neural networks. A “last-known observation” approach was used as a baseline for these models to beat which repeats the price at timestep five of the input data as the prediction for the next five days. Variables used for training and forecasting includes the prices of oil, gas, nickel, lead, tin, a US dollar index, aluminum, copper and zinc. Our results find that none of the single- or multi-layer LSTM or GRU models were able to out-perform the baseline model and in many cases the baseline significantly out-performed the recurrent neural network models. In general, the GRU models performed slightly better than the LSTM models, but not for all the metals. Further work could be done on multi-step commodity price forecasting by choosing a different time horizon or using intra-day data for a larger dataset. Other explanatory variables such as iron ore or coal could be included in the modeling and more complex networks such as the ResNet and LSTnet could be implemented.