Neural Networks in the Digital Economy

Abstract

An analysis of the mathematical models and methods, which are used to build the algorithms of the digital sector of the economy, made it possible to draw the following conclusion: the modern development of mathematics and information technology provides a wide range of possibilities for setting and solving various problems in the field of economics. One of them is the analysis of time series in order to predict the behavior of a dynamic system. The ability to use artificial neural networks for this is justified by the theorems of mathematics by Takens, Gorban, and Kolmogorov, which made it possible to reduce the problem of forecasting time series to the problem of approximating a continuous function of several variables. Based on the analysis of neural networks and modern methods of their training, a phased construction of the ANN, which determines the main characteristics of the ANN, is proposed, the correct choice of which will allow to qualitatively solve the task of forecasting the time series. The main stage of the scheme under consideration is the selection and training of a neural network, which requires an assessment of all possible architectures and methods for training neural networks, since there are still no general criteria to confidently choose the optimal «methodology» for training a neural network, despite their active using. The choice of the most important properties of a neural network: its architecture, block size, learning speed, remains largely empirical. A high-quality solution to any economic problem requires a comprehensive analysis of existing training algorithms for neural networks in order to choose the best one. This inhibits using of ANNs to solve the practical problems of the economy. It’s proposed to use artificial neural networks (ANNs) only to solve such problems of economics and management, where the formalization of the decision stages is either quite complicated or impossible at all.KeywordsDigital economyNeural networksDeep learningInterpolationParallel computingJEL CodesC45C88