Crude oil price forecasting based on the reconstruction of imfs of decomposition ensemble model with arima and ffnn models

Abstract

The development of economic and industry depend upon how well the accuracy of crude oil price forecasting is managed. The study aims to reduce computation complexity and enhance forecasting accuracy of decomposition ensemble model. The propose model comprises four steps which are (i) decomposing the complex data into several IMFs using ensemble empirical mode decomposition (EEMD) method, (ii) reconstructing the decomposed IMFs through autocorrelation into stochastic and deterministic components, (iii) forecasting every reconstructed component, and (iv) ensemble all forecasted components for the final output. IMFs in the stochastic component are analysed separately. The findings confirm that the stochastic component contributed more variation as compared to deterministic component. For verification and illustration, Brent, West Texas Intermediate (WTI) daily, weekly, monthly and yearly, and Pakistan monthly spot crude oil prices were used as sample study. The empirical results indicated that the proposed model statistically outperformed all the considered benchmark models including the most popular auto-regressive integrated moving average (ARIMA) model, feed forward neural network (FFNN) model, decomposition ensemble model (EEMD-ARIMA and EEMD-FFNN), reconstruction decomposition ensemble model with stochastic and deterministic components (EEMD-(S+D)-ARIMA and EEMD- (S+D)-FFNN) and Rios and De Mello (RD) reconstruction decomposition ensemble model with stochastic and deterministic components (EEMD-RD-ARIMA and EEMD-RD-FFNN). To determine the performance, two descriptive statistical measures were applied, including the root mean square error (RMSE) and mean absolute percentage error (MAPE). The MAPE of the proposed EEMD-individual stochastic and deterministic (ISD)-FFNN model for daily and weekly data of Brent and WTI are <1%, however, for monthly Brent, WTI and Pakistan data are <5% shows a good fit produce by EEMD-ISD-FFNN. The MAPE of the model EEMDISD- FFNN for yearly Brent data is <30% indicate a reasonable fit and for WTI <20% implies a good fit. Whereas the MAPE of the EEMD-(S+D)-FFNN model for Brent yearly data <20% display a good fit and for WTI data <10% indicate excellent fit. In nutshell, the recommended model for yearly data is EEMD-(S+D)-FFNN. In conclusion, the proposed method of reconstruction of IMFs based on autocorrelation enhanced the forecasting accuracy of the EEMD model.