EXPERIMENTAL STUDY OF THE ACCURACY OF COMPRESSION-BASED FORECASTING METHODS

Abstract

In information theory it is known that methods of data compression can be used for forecasting of stationary processes. In this paper an compression-based algorithm for time series forecasting was proposed and empirical study of its accuracy was carried out. The algorithm can operate with arbitrary methods of data compression. During the steps of the algorithm predicted values from different methods are combined, and the greatest impact on the end result is exerted by the method with the best compression ratio for the series. The algorithm can be used for forecasting of time series with discrete and continuous alphabets. To improve the accuracy of the forecast existing methods of time series preprocessing can be used. The empirical study of the efficiency of the proposed algorithm was conducted on time series from the M3 Competition and the T-index series. To generate forecasts well-known archivers were used. The results of the calculations showed that the obtained method has a relatively high accuracy and speed.