Stochastic model of mineral prices incorporating neural network and regression analysis

Abstract In mine valuation commodity price forecasts are required to assess the economic viability of a project. The forecast must cover a relevant period to capture the trend and volatility in prices within a mining business cycle. Mineral commodity prices are volatile, which means that the results of evaluation tools that do not treat the stochasticity of metal prices rigorously may be misleading. A model has been set up in which the gold price is modelled by a three-step procedure: a multivariate, normally distributed random variable generator (MNDRVG), the method of multilayer feed-forward neural networks (MFNN) and multiple regression analysis. The first step uses MNDRVG to generate the input data. In the second step the data are fed into an orthogonal projection model to estimate the parameters of a linear regression. In the third step the estimated regression is used to forecast the gold price. Because not all the input data are known at the time of the forecast, MFNN is used to generate the unknown input data of the exogenous factors of the forecasting equation. Data on forecasting variables are normally available for a few years; to obtain realistic estimates from the MFNN and multiple regression models more data are needed to train the network. The MNDRVG model provides a convenient method for extending the data set since it takes into account the uncertainty in the parameters and the spread of the data around the mean. A data set of an identified socio-politico-economic cycle was used to validate the model. Analysis of the results shows that the estimated mineral price model predicts the annual average gold price with very high precision. The main novelty of the methodology is the simulation and rigorous analysis of the randomness property associated with mineral price to reduce the estimation and forecasting errors.