Abstract:
Data analytics especially predictive analytics is very important research domain which includes time series forecasting.
Nonlinear nonstationary time series are challenging to predict. This paper presents the outcome of the research study in finding
better forecasting methods for nonlinear nonstationary time series. Rolling forecast approach and locally adaptive empirical mode
decomposition (EMD)-based hybridization were employed with autoregressive integrated moving average (ARIMA) and exponentially
weighted moving average (EWMA). Thus, two methods were EMD-ARIMArolling and EMD-EWMArolling of which the later was
found better in this study. Also, EMD-EWMArolling was combined with ARIMArolling and EWMArolling using affine combinations
to develop affEEArolling and affEEErolling methods. Proposed affEEArolling and affEEErolling along with six other compared
methods were employed on nine closing price stock datasets from NASDAQ Financial-100 companies and compared using accuracy
measurements. From the results, it was found that proposed methods significantly improved forecast accuracy and outperformed the
compared methods (e.g., in ACGL dataset, affEEArolling reduced RMSFE by 55.98% where rolling forecast, EMD-hybridization
and affine combination improved 43.7%, 4.24% and 18.28% respectively and affEEErolling improved 56%). Hence, EMD-based
hybridizations and forecast combinations can be useful tools for time series forecasting. In addition, such EMD-based advanced methods
can be considered for inclusion in financial technologies.
