Abstract:
The research objective of the present study is the development of a model for increased accuracy of steel-price forecasts, which is of paramount importance for firms who use steel as an input and thus need to make informed decisions with regard to an optimal amount and type of hedge against unfavourable steel-price movement. To achieve its aim, the study forms weighted average combinations of steel price forecasts generated separately by a transfer function ARIMA model (ARIMA-TF) and an artificial neural network model (ANN), as both models are shown to contribute independent information with regard to target variable (steel price) movement. A generalized reduced gradient algorithm (GRG) method is employed to estimate the component model forecast weights, which is a novel approach introduced by this study. The data set employed includes a time series of monthly steel prices (cold rolled flat steel) from February, 2012 to November, 2020. Explanatory variables include iron ore price, coking coal price, capacity utilization, GDP and industrial production. With regard to the out of sample forecasts of all models (component and combining), mean absolute percentage forecast errors (MAPE) are calculated and model comparisons are made. The study finds that the combining model formed with the gradient algorithm approach in which the weights are constrained to be nonnegative and sum to one has the lowest MAPE of all models tested, and overall is found to be very competitive with other models tested in the study. The policy implication for firms that use steel as a major input is to base their hedging decisions on a combination of forecasts generated by ARIMA-TF and ANN models, with the forecast weights generated by a constrained generalized reduced gradient algorithm (GRG) method.
