Nonlinear time series forecasting using MARS

Abstract

One of the most important uses of artificial neural networks is to forecast non-linear time series, although model-building issues, such as input selection, model complexity and parameters estimation, remain without a satisfactory solution. More of research efforts are devoted to solve these issues. However, other models emerged from statistics would be more appropriated than neural networks for forecasting, in the sense that the process of model specification is based entirely on statistical criteria. Multivariate adaptive regression splines (MARS) is a statistical model commonly used for solving nonlinear regression problems, and it is possible to use it for forecasting time series. Nonetheless, there is a lack of studies comparing the results obtained using MARS and neural network models, with the aim of determinate which model is better. In this paper, we forecast four nonlinear time series using MARS and we compare the obtained results against the reported results in the technical literature when artificial neural networks and the ARIMA approach are used. The main finding in this research, it is that for all considered cases, the forecasts obtained with MARS are lower in accuracy in relation to the other approaches.