Time Series Prediction Based on Recursive Update Gaussian Kernel Function Networks

Abstract

This paper presents a new method of predicting the values of time series using recursive update Gaussian Kernel Function Networks. First, the input structure of time series prediction model is determined by the phase space analysis of time series. Then, the one step time series prediction model is trained using the Gaussian kernel function network. In the case of multiple step time series prediction, the estimated value is used along with previous input data to make a prediction model for the right next prediction step and the same process is recursively updated until it reaches the desired prediction step. In this model, the prediction model is trained in such a way that the accumulated error due to the recursive prediction method is reduced as much as possible. For the demonstration of the proposed method, the time series prediction of Kosdaq (one of the Korean composite index) data was performed. As a result, the proposed model outperforms other prediction models such as a simple recursive prediction model, direct prediction model and also other widely used regression methods, such as support vector machines and k-nearest neighbors.