Sensitivity analysis for time lag selection to forecast seasonal time series using Neural Networks and Support Vector Machines

Multi-step ahead forecasting is an important issue for organizations, often used to assist in tactical decisions. Such forecasting can be achieved by adopting time series forecasting methods, such as the classical Holt-Winters (HW) that is quite popular for seasonal series. An alternative forecasting approach comes from the use of more flexible learning algorithms, such as Neural Networks (NN) and Support Vector Machines (SVM). This paper presents a simultaneous variable (i.e. time lag) and model selection algorithm for multi-step ahead forecasting using NN and SVM. Variable selection is based on a backward algorithm that is guided by a sensitivity analysis procedure, while model selection is achieved using a grid-search. Several experiments were devised by considering eight seasonal series and the forecasts were analyzed using two error criteria (i.e. SMAPE and MSE). Overall, competitive results were achieved when comparing the SVM and NN algorithms with HW.