Implementation of bagging in time series forecasting

Abstract

Objectives. The purpose of the article is to build different models of bagging, to compare the accuracy of their forecasts for the test period against standard models, and to draw conclusions about the possibility of further use of the bagging technique in time series modeling.Methods. This study examines the application of bagging to the random component of a time series formed after removing the trend and seasonal part. A bootstrapped series combining into a new random component is constructed. Based on the component thus obtained, a new model of the series is built. According to many authors, this approach allows the accuracy of the time series model to be improved by better estimating the distribution.Results. The theoretical part summarizes the characteristics of the different bagging models. The difference between them comes down to the bias estimate obtained, since the measurements making up the bootstraps are not random. We present a computational experiment in which time series models are constructed using the index of monetary income of the population, the macroeconomic statistics of the Russian Federation, and the stock price of Sberbank. Forecasts for the test period obtained by standard, neural network and bagging-based models for some time series are compared in the computational experiment. In the simplest implementation, bagging showed results comparable to ARIMA and ETS standard models, while and slightly inferior to neural network models for seasonal series. In the case of non-seasonal series, the ARIMA and ETS standard models gave the best results, while bagging models gave close results. Both groups of models significantly surpassed the result of neural network models.Conclusions. When using bagging, the best results are obtained when modeling seasonal time series. The quality of forecasts of seigniorage models is somewhat inferior to the quality of forecasts of neural network models, but is at the same level as that of standard ARIMA and ETS models. Bagging-based models should be used for time series modeling. Different functions over the values of the series when constructing bootstraps should be studied in future work.