Moving intervals for nonlinear time series forecasting

In this paper a new forecasting methodology to be used on time series prediction is introduced. The considered nonlinear method is based on support vector machines (SVM) using an interval kernel. An extended intersection kernel is introduced to discriminate between disjoint intervals in reference to the existing distance among them. The model presented is applied to forecast exchange ratios using six world's major currencies. The results obtained show that SVMs based on interval kernel have a similar behavior than other SVM classical forecasting approaches, allowing its performance to be seen as very promising when using high frequency data.