GOLD PRICE FORECASTING: A FORECAST COMBINATION APPROACH.

Abstract

We forecast monthly gold prices using forecast combinations with equal weights and Bates and Granger (1969) weights. Component models include non-seasonal exponential smoothing models, univariate autoregressive integrated moving average (ARIMA) model, distributed lags regressions with ARMA errors, vector autoregressions, as well as feed-forward neural network autoregressions with a single hidden layer. In terms of root mean squared error (RMSE), the equally weighted forecast combinations of both univariate and multivariate component models managed to beat the benchmark ARIMA model at 7 to 12 steps ahead forecast horizons. The equally weighted forecast combination of univariate models beat the benchmark ARIMA in 11 of the 12 different h-steps ahead forecast horizons considered. Other observations include the better performance of equally weighted forecast combinations over those using Bates and Granger (1969) weights, and the superior forecasting performance of simpler models over more complex models with numerous predictors.