Coking coal futures price index forecasting with the neural network

Coking coal price forecasting is a significant issue for investors and policy makers. This study explores usefulness of the nonlinear autoregressive neural network for this forecasting problem in a dataset of daily closing prices of the coking coal futures traded in China Dalian Commodity Exchange during January 4, 2016-December 31, 2020. Through examining various model settings across the algorithm, delay, hidden neuron, and data splitting ratio, the model leading to generally accurate and stable performance is reached. Particularly, the model's inputs are the lagged coking coal futures prices and output is the 1-day ahead price forecast. The model is based on the two-layer feedforward network with six delays and two hidden neurons, which is trained through the Levenberg-Marquardt algorithm, and leads to relative root mean square errors of 1.84%, 1.85%, and 1.84% for the training, validation, and testing phases, respectively. Usefulness of the machine learning technique for the price forecasting problem of the coking coal price is illustrated. Results here might be used on a standalone basis as technical forecasts or combined with fundamental forecasts to form perspectives of price trends and perform policy analysis.