International Journal of Computer Theory and Engineering

Abstract

Time series forecasting is a challenging task in many fields. Due to the complex non-linear relationship between the multidimensional features of the time series data, improved time series forecasting requires a forecasting model that combines multiple prediction models. Ensemble learning performs better than single learning model and discovers regularities in dynamic and non-stationary data. In literature, single level neural network ensembles are used for the prediction problems [1- 5]. This paper introduces a novel two level ensemble learning approach based on Radial Basis Function networks (RBF), K - Nearest Neighbor (KNN) and Self Organizing Map (SOM) for time series prediction with the aim of increasing the prediction accuracy. The evaluation of the proposed Pattern Prediction Ensemble Model (PAPEM) using three input datasets such as, Mackey dataset, Sunspots dataset and Stock Price dataset shows that the proposed PAPEM model performs better than the individual classifiers.