НАВЧАННЯ КОМБІНОВАНОЇ МОДЕЛІ ПРОГНОЗУВАННЯ ЧАСОВИХ РЯДІВ

Abstract

The method of construction of the combined model of forecast ing of time series based on basic models of forecasting is developed in the work. The set of basic models is dynamic, ie new prediction models can be included in this set. Models also can be deleted depending on the properties of the time series. For the synthesis of a combined model of forecasting time series with a given forecast step, the optimal step of prehistory is determined at the beginning. Next the functional is constructed. The optimal prehistory step is determined using the autoregression method for a fixed forecast step. It determines the period of time at which the accuracy of models from the base set is analyzed. For each basic model during the process of the construction of the combined model is determined by the weighting factor with which it will be included in the combined model. The weights of the basic models are determined based on their forecasting accuracy for the time period determined by the prehistory step. The weights reflect the degree of influence of the base models on the accuracy of the combined model forecasting. After construction of the combined model, its training is carried out and those basic models which will be included in the final combined model of forecasting are defined. The rule of inclusion of basic models in the combined model is established. While including basic models in the combined forecasting model, their weights are taken into account, which depends on the same parameter. The optimal value of the parameter is determined by minimizing the given functional, which sets the standard deviation between the actual and predicted values ​​of the time series. Weights with optimal parameters are ranked in decreasing order and are used to include basic models in the combined model. As a result of this approach, as predicted values for the real time series show, it was possible to significantly improve the forecasting accuracy of the combined model in many cases. The developed method of training provides the flexibility of the combined model and its application to a wide class of time series. The results obtained in this work contribute to solving the problem of choosing the most effective basic models by synthesizing them into one combined model.