Development of an aggregator for choosing the best forecasting method for groups of time series

Abstract

The paper presents a general algorithm for computing an aggregated time series forecast (TS), within which machine learning methods are used to adjust the parameters of a hybrid combined forecasting model. Also presented are the results of experiments on the application of the developed algorithm using the TS competition “Computational Intelligence in Forecasting” (CIF). The use of a neural network for choosing forecasting methods allowed, on average, for all experiments to reduce the error by 7.225%, as can be seen from the results of the experiments. The average error for the eight prediction methods chosen by the neural network turned out to be lower than the average error for all methods in 47 cases out of 50 (94%).