Daily Crude Oil Prices Forecasting Using a Novel Hybrid Time Series Technique

This paper introduces a new hybrid time series forecasting technique to obtain an efficient and accurate daily crude oil prices forecast. The proposed hybrid technique combines the features of various regression, time series, and machine learning models to improve forecast accuracy. First, it involved processing the original crude oil prices time series to address missing values, variance stabilization, and normalization. Second, it divides the purified crude oil price time series into two major parts: the nonlinear long-term trend (long-run fluctuations) and the residual (short-term fluctuations). The secular non-linear trend series is modeled using various regression models, including linear, spline, lowess, and smoothing spline regression models. On the other hand, the short-run random part is modeled and forecasted using four benchmark time series (linear and non-linear AR, ARMA, ESM) and four machine learning (Neural network autoregressive, Random forest, support vector regression with polynomial and radial basal functions) models. The forecast from both parts is summed to obtain the final forecasting results. The proposed hybrid time series method builds a strong emphasis on effectively measuring the nonlinear long-term trend, which has been disregarded in prior works. A variety of evaluation metrics are employed to evaluate the proposed effectiveness. The testing findings demonstrate the hybrid forecasting technique's accuracy and efficacy. Specifically, an MAE of 1.28106, an RMSE of 1.59259, a PCC of 0.94158, and a DS of 0.82149 are obtained for the Brent and WTI oil markets when the Lowess regression and NPAR models are combined. With regard to WTI oil prices in particular, the model's resilience in producing precise forecasts is further demonstrated by the lowest MAE of 1.25730, RMSE of 1.55390, and PCC of 0.93890.