Price indicators are often non-stationary stochastic processes. As a result, theoretically, the best forecast is simply the last observed value. In this study, we examine various approaches to constructing pseudo out-of-sample forecasts, using futures prices for oil and petroleum products as examples. We use simple econometric models (including different types of random walks, moving averages, linear trend models, and ARIMA) to compare the predictive performance of recursive and rolling (moving window) forecasting schemes. Our results show that the rolling forecasting scheme outperformed the recursive scheme more frequently. Linear trend models with a 6-year rolling window and ARIMA models with 8-year and 9-year rolling windows prove to be the most broadly applicable models for forecasting oil and petroleum product prices. Finally, we observe that forecast accuracy, on average, deteriorates (relative to a naive last-value forecast) as the forecast horizon increases from 24 months to 60 months.
