Maximal overlap discrete wavelet transform gaussian process regression for monthly crude oil price forecasting.

Abstract

Crude oil is one of the most traded commodities in the world and its prices have a significant impact on the global economy. Although many forecasting models have been developed for predicting oil prices, it remains one of the most challenging forecasting tasks due to the high volatility of oil prices. This study proposes a combination of Maximal Overlap Discrete Wavelet Transform (MODWT) with Gaussian Process Regression (GPR) for monthly crude oil price forecasting. For this purpose, monthly West Texas Intermediate (WTI) and Brent crude oil prices are used to evaluate the performance of proposed model. Feature selection in time series forecasting is important to improve learning accuracy and reduce computational time by eliminating redundant and irrelevant input. In order to identify significant input time lag for the models, Neighborhood Component Feature Selection (NCFS) is first used to rank the input lag based on their importance. MODWT is used to decompose the time series data to details (high frequency) and approximation (low frequency) components. The proposed model, MODWT-GPR model is developed by using the decomposed time series components obtained from MODWT. Two types of MODWT-GPR models with three different wavelet families (daubechies, symlets and coiflets) are studied and compared to get the best forecasting performance. Standard MODWT-GPR model only uses the decomposed wavelet components, while Hybrid MODWT-GPR includes the original input lag series in additional to the decomposed wavelet component. The proposed models are compared with other benchmark models including Autoregressive Integrated Moving Average (ARIMA), Artificial Neural Network (ANN) and the conventional GPR model. The forecasting performance of the tested method are measured by using root mean square error (RMSE), mean absolute error (MAE) and mean absolute percentage error (MAPE). The result shows that the integration of MODWT improves the performance of the conventional GPR model and Hybrid MODWT-GPR outperforms all other forecasting models used in this study. Future study suggests the application of different kernels to study their impact on crude oil price forecasting.