Performance of different models in iron ore price prediction during the time of commodity price spike

Abstract

Future prediction of commodity price based on available data is very important for mining investors and operators. Commodity prices cointegrate and show Granger causality to and from one another. This research reviewed five different estimation techniques which are Bivariate Non-Linear Regression (BNLR), Multiple Linear Regression (MLR), Multiple Non-Linear Regression (MNLR) as well as logsig and tansig model of Levenberg-Marquardt Artificial Neural Network modelling to simulate the future iron ore price based on 12 other monthly commodity prices and indices including LNG, aluminium, nickel, silver, Australian coal, zinc, gold, oil, tin, copper, lead, and Commodity Price Index (Metals).Six different models were tested in the paper to forecast the iron ore prices from 1 to 6 months over 10 months period. Linear model (purelin) using Levenberg-Marquardt technique was able to exhibit the best forecast result with average accuracy of 5.92% for 1 month ahead, 9.48% for 2 months, 11.21% for 3 months, etc. It is important to highlight that high accuracy is achieved (accuracy under 5% between forecasts and actuals in 40–50% cases) by purelin model for up to 2 months forecast for the period between July 2020 and April 2021. This indicates that prediction of iron ore price for the coming month is possible for up to 2 months period using the purelin model. It can be noted that the period tested was unstable for iron ore prices where rapid surge in iron ore price was observed. Same principle can be applied in the time of next commodity price cycle.