Global and decomposition evolutionary support vector machine approaches for time series forecasting

Multi-step ahead time series forecasting (TSF) is a key tool for supporting tactical decisions (e.g., planning resources). Recently, the support vector machine (SVM) emerged as a natural solution for TSF due to its nonlinear learning capabilities. This paper presents two novel evolutionary SVM (ESVM) methods for multi-step TSF. Both methods are based on an estimation distribution algorithm search engine that automatically performs a simultaneous variable (number of inputs) and model (hyperparameters) selection. The global ESVM (GESVM) uses all past patterns to fit the SVM, while the decomposition ESVM (DESVM) separates the series into trended and stationary effects, using a distinct ESVM to forecast each effect and then summing both predictions into a single response. Several experiments were held, using six time series. The proposed approaches were analyzed under two criteria and compared against a recent evolutionary artificial neural network (EANN) and two classical forecasting methods, Holt–Winters and autoregressive integrated moving average. Overall, the DESVM and GESVM obtained competitive and high-quality results. Furthermore, both ESVM approaches consume much less computational effort when compared with EANN.