Ex-Post (Pseudo) Out-of-Sample Forecast Accuracy of Proposed Oil Price Models

Abstract

This thesis aims to test and compare some of the most frequently applied models in the forecasting literature, for their ability to produce accurate ex-post (pseudo) out of-sample forecasts of the crude oil price. These models range vastly in complexity, ranging from the most parsimonious idea of price-today-is-price-tomorrow approach to more sophisticated and stochastic models. All models will be assessed with the commonly used proxy for the oil price, namely the West Texas Intermediate (WTI) benchmark price, sampled in both daily and monthly frequencies. A model’s forecast accuracy will be evaluated employing a set of various loss functions that differ in their way of penalizing the forecast errors. Additionally, the models’ forecasts will be tested for being directionally accurate in predicting the actual price changes. Finally, model selection and estimation will be analysed across different lengths of historical price data, to examine what effect the choice of sample period has on the forecast results. The empirical results of this analysis show that neither the deterministic or stochastic models evaluated are able to forecast the price of crude oil with an adequately desired accuracy. It was also found that forecast results are highly sensitive to the choice of sample period for historical prices used as input for model estimation, and that certain models perform better when only recent market data is used as input.