Steel price index forecasting through neural networks: the composite index, long products, flat products, and rolled products

Forecasting commodity prices is a vital issue to a wide spectrum of market participants and policy makers in the metal sector. In this work, the forecast problem is investigated by focusing on the Chinese market, with the daily steel price indices of the composite index, long products, flat products, and rolled products spanning a 10-year period from June 15, 2011, to April 15, 2021. The non-linear auto-regressive neural network is employed as the forecasting model and model performance corresponding to a variety of settings is explored over algorithms for model estimations, numbers of hidden neurons and delays, and ratios for splitting the data for each of the four price indices. Models that are relatively simple and generate forecasts of high accuracy and stabilities are arrived at. Particularly, relative root mean square errors (RRMSEs) of 0.49%/0.49%/0.52%/0.55%, 0.51%/0.44%/0.51%/0.49%, and 0.52%/0.53%/0.56%/0.53%, respectively, for training, validation, and testing, and the overall RRMSE of 0.50%/0.49%/0.53%/0.54% are reached for the composite index/long products/rolled products/flat products. The results could, on the one hand, serve as standalone technical price forecasts. They could, on the other hand, be combined with other (fundamental) forecast results for forming perspectives of price trends and carrying out policy analysis.