Ensemble methods for time series forecasting

Abstract

Improvement of time series forecasting accuracy is an active research area that has significant importance in many practical domains. Ensemble methods have gained considerable attention from machine learning and soft computing communities in recent years. There are several practical and theoretical reasons, mainly statistical reasons, why an ensemble may be preferred. Ensembles are recognized as one of the most successful approaches to prediction tasks. Previous theoretical studies of ensembles have shown that one of the key reasons for this performance is diversity among ensemble members. Several methods exist to generate diversity. Extensive works in literature suggest that substantial improvements in accuracy can be achieved by combining forecasts from different models. The focus of this chapter will be on ensemble for time series prediction. We describe the use of ensemble methods to compare different models for time series prediction and extensions to the classical ensemble methods for neural networks for classification and regression prediction by using different model architectures. Design, implementation and application will be the main topics of the chapter, and more specifically: conditions under which ensemble based systems may be more beneficial than their single machine; algorithms for generating individual components of ensemble systems; and various procedures through which they can be combined. Various ensemble based algorithms will be analyzed: Bagging, Adaboost and Negative Correlation; as well as combination rules and decision templates. Finally, future directions will be time series forecasting, machine fusion and others areas in which ensemble of machines have shown great promise.