A Novel Weighted Ensemble Technique for Time Series Forecasting Ratnadip Adhikari and R.K. Agrawal School of Computer and Systems Sciences, Jawaharlal Nehru University, New Delhi-110067, India {adhikari.ratan,rkajnu}@gmail.com Abstract. Improvement of time series forecasting accuracy is an active research area having significant importance in many practical domains. Extensive works in literature suggest that substantial enhancement in accuracies can be achieved by combining forecasts from different mod- els. However, forecasts combination is a difficult as well as a challenging task due to various reasons and often simple linear methods are used for this purpose. In this paper, we propose a nonlinear weighted ensem- ble mechanism for combining forecasts from multiple time series models. The proposed method considers the individual forecasts as well as the correlations in pairs of forecasts for creating the ensemble. A successive validation approach is formulated to determine the appropriate combi- nation weights. Three popular models are used to build up the ensemble which is then empirically tested on three real-world time series. Obtained forecasting results, measured through three well-known error statistics demonstrate that the proposed ensemble method provides significantly better accuracies than each individual model. Keywords: Time Series Forecasting, Ensemble Technique, Box-Jenkins Models, Artificial Neural Networks, Elman Networks. 1 Introduction Time series forecasting has indispensable importance in many practical data mining applications. It is an ongoing dynamic area of research and over the years various forecasting models have been developed in literature [1,2]. A major concern in this regard is to improve the prediction accuracy of a model without sacrificing its flexibility, robustness, simplicity and efficiency. However, this is not at all an easy task and so far no single model alone can provide best forecasting results for all kinds of time series data [3,4]. Combining forecasts from conceptually different methods is a very effective way to improve the overall forecasting precisions. The earliest use of this prac- tice started in 1969 with the monumental work of Bates and Granger [5]. Till then, numerous forecasts combination methods have been developed in litera- ture [6,7,8]. The precious role of model combination in time series forecasting can be credited to the following facts: (a) by an adequate ensemble technique, the P.-N. Tan et al. (Eds.): PAKDD 2012, Part I, LNAI 7301, pp. 38–49, 2012. c  Springer-Verlag Berlin Heidelberg 2012