Hybridization of nonlinear and linear models for time series forecasting

Abstract

Nowadays, time series forecasting is very challenging due to the uncertainties of real world events that are influenced by many indefinite factors and rapid changes. This scenario requires forecasting methods that work efficiently with incomplete and multivariate data. Otherwise, the solutions will tend to trap into a local minimum or will encounter over fitting problems that would lead to inaccurate predictions. Due to the complexity of time series data, there is no single method that best fit all situations. Therefore, a hybrid model is proposed to improve individual model performance. However, most of the existing hybrid models utilize univariate data which requires huge historical data to obtain accurate forecasting result. Thus, in this research, a new hybrid model is developed by combining nonlinear and linear models to overcome the current discrepancies of the existing hybrid models. The proposed hybrid model integrates the nonlinear Grey Backpropagation Particle Swarm Optimization and linear Autoregressive Integrated Moving Average model (GBPSO_ARIMA) by combining new features. These features include multivariate time series data, cooperative feature selection (CFS), optimization algorithm (PSO) and hybrid sequence process. The proposed model is tested on four different sets of time series data that exhibit different time series behaviours and data scales. The performance of the proposed model is analysed and compared with the individual and hybrid models. The experiments show that the proposed hybrid model is more robust for all datasets regardless of the sample sizes and data behaviours. The validation of these results is further tested using hypothesis test and analysed in terms of stability and adaptability. The results have proven that the proposed hybrid model provides a better alternative tool for time series forecasting. This is due to its promising performance and robust capability in handling multivariate small and large scale time series data in incomplete data situations, which is also capable in reducing over fitting issues and local minimum problems.