Nonlinear Combinational Forecasting Based on Support Vector Machine

Abstract

Traditional forecast techniques apply a single forecaster to carry out the task. However, this forecaster might not be the best for all situations or databases. In this paper, we propose a nonlinear combinational forecasting model mainly based on SVM and several single forecasting methods such as Grey theory, Grey Verhulst model, exponentiation model, exponent model and linear model to carry out data forecasting. During the process of the forecast, firstly the model group fits several single forecast approaches are used to form a model group, and a set of data in time sequence. Secondly, the fitted results by different traditional predictive models in time sequence act as the input of the support vector machine regression (SVMR) model, and the changeable weights of the input quantities, which are the important part of the combinational SVMR prediction model, can be obtained by relative SVMR approach based on known input and output samples. Several single prediction approaches are used to build first model group. Then the prediction results of the first model group constitute a multi-dimension vector and act as the input sample of Support Vector Machine (SVM) model. In the paper, the procedure of the combinational prediction based on SVMR is discussed in details. Two examples in time sequence have proven that the proposed model can balance fitting and extrapolation, and make a break through the traditional prediction rules. At the same time, the issue that the good fitting with bad extrapolation resulted by excessive learning is solved effectively in this paper. Moreover, the proposed combinational prediction model has higher prediction accuracy by comparing with the other approaches. The validity and feasibility of the proposed model has also been proven through the experiences.