A New Non-negative Variable Weight Combination Forecasting Method Based On Sliding Window

Combination Forecasting(CF)combines several different kinds of single forecasting methods by feasible weighting method,so as to realize the comprehensive utilization of the effective information provided by various single forecasting methods.According to whether the weights coefficients of CF change with time,it can be divided into variable weights CF and invariable weights CF.There are many methods of combination of invariable weights,and they are easy to operate,so it is wildly used in practice.However,compared with the variable weights CF method,the forecasting accuracy of the invariable weights is higher at different time points by different individual forecasting methods.In view of the existing problems,in this paper,the sliding window weighting method is introduced.By setting reasonable window length,the non-time-varying weights are changed into time-varying weights by sliding weights,thus improving the accuracy of CF.The main contents of this paper are as followsFirst,the object of study is not only a single point-value sequence,but also an interval-type sequence.The new model proposed in this paper is not only applied to single-point sequence values but also combined with interval sequence values to verify the applicability of the method.Secondly,on the basis of the CF method of entropy weights and the CF method of traditional variation coefficient weights,an improved CF method of variation coefficient weights is put forward,which can transfer the traditional variation coefficient based on the prediction data level to the residual data level,and effectively eliminate the situation that the variation degree of traditional variation coefficient is weakened due to the data order of magnitude.Thirdly,considering that the CF model with invariant weights can't predict the weights of each time well in some cases,the sliding window mechanism is introduced.Combining with the idea of entropy method,that is,the larger the entropy value of the system,the smaller the information it contains and the smaller the var-iation degree of some index of the system.Vice versa,the smaller the entropy value of the system,the larger the information it contains,and the greater the variation degree of a certain index of the system is.On this basis,starting from the basic definition of coefficient of variation in statistics,this paper sets the appropriate window length,defines the variation degree of prediction errors of each single prediction model based on sliding window,and then obtains the weights coefficient,so as to realize the transformation from invariable weights to variable weights.This method has wide applicability and can improve the accuracy of combination prediction effectively.