Advanced neural network forecasting system with continuous learning for grouped time series
Keywords:
Abstract
Introduction: Forecasting of grouped time series is a complex applied problem when it is necessary to take into account both intra-series and inter-series relationships, and promptly respond to changes in the laws of series formation. Purpose: To improve forecasting systems for grouped time series, allowing to increase the accuracy of the received forecasts. Results: We develop an improved system of neural network forecasting of grouped time series with continuous learning, which includes the correction unit that allows to automatically select hyperparameters of forecasting and make the most correct associative call of information from the neural network memory. We propose new rules for the implementation of the system in software execution with improved associative recall of information from neural network memory, which increases the stability of the functioning of neural networks themselves. We develop and programmatically implement the algorithm of the correction unit operation, which provides the selection of the attenuation coefficient, the retrospective depth of the components of the grouped time series, as well as the threshold of neuronal excitation. The dependence of prediction accuracy on the size of neural network channels is investigated. The example of forecasting market indicators demonstrates the advantage of the developed system in comparison with known analogues. Practical relevance: The improved software system makes it possible to obtain more accurate time series forecasts for solving applied problems. Thus, the average absolute error is reduced by 2–35%, the average absolute percentage error by 4–37% and the standard error by 3–29%. Discussion: In the future, to increase the efficiency of the system, it is necessary to develop rules for the automatic selection of other hyperparameters, as well as to optimize the algorithm for their selection to reduce computational costs.
