Balanced-Sampling-Based Heterogeneous SVR Ensemble for Business Demand Forecasting

An accurate demand forecasting model has academic and practical significance to supply chain management. However, multi-source data and error data have great effect on the demand prediction accuracy. Therefore, a balanced-sampling-based ensemble of heterogeneous support vector regression forecasting method named BS-EnHSVR (Balanced-Sampling-based Ensemble of Heterogeneous SVR) is proposed in this paper to improve the prediction accuracy by employing balanced sampling and heterogeneous ensemble learning techniques. Training dataset is firstly classified to different clusters by using clustering algorithm, and then sample data from each cluster equally to generate training subset for training different individual SVR models with different training parameters for ensemble. Experimental results on beer sales show that the proposed method has good usability and generalization ability.