Abstract: the purpose of the study is to justify the choice of artificial neural network architectures for inclusion in the neural network module of the hybrid short-term forecasting system “SGM Horizon” developed by the authors. The hybrid model system and the SGM Horizon software package are designed to forecast macroeconomic indicators at the country and regional level. The hybrid approach ensures that the required quality of forecasts is achieved for the entire indicator system. Neural network models play a key role here. The article provides an overview of the work on the use of neural networks in forecasting, with special attention paid to research in the field of forecasting economic time series. The methodological basis of the research is a hybrid approach that includes machine learning models, such as neural networks and decision trees, along with econometric models. Two basic architectures are used for predicting economic indicators based on neural networks: direct distribution networks and recurrent networks. In direct propagation networks, a multi – layer perceptron is used to predict time series; in recurrent networks, a network architecture with long-term and short-term memory (LSTM) is used. Based on the analysis of tools for solving modeling problems based on neural networks, the functions of the C++ and C# FANN (Fast Artificial Neural Network) library were used for their implementation in SGM Horizon. Within the framework of SGM Horizon, a module of Artificial neural networks has been developed that allows time series forecasting based on multi-layer perceptron architectures and recurrent networks with long-and short-term memory (LSTM). The forecast of indicators of macroeconomics, state budgets, the social sphere and foreign economic activity of the Russian Federation is carried out in the “SGM Horizon” system. At the first stage, the system predicts all indicators based on a regression model. As a result, out of a total of 175 indicators, high accuracy and quality values were obtained for 125 indicators. At the second stage, neural network predictive models were built for the remaining 50 indicators using the neural network module. Results of satisfactory accuracy were obtained for 45 indicators. The hybrid approach to modeling and forecasting of socio-economic indicators of the Russian Federation developed by the authors makes it possible to achieve high accuracy of the forecast for the entire set of indicators under study. The forecasting of indicators of macroeconomics, state budgets, social sphere and foreign economic activity of the Russian Federation carried out in the SGM Horizon system demonstrates the effectiveness of this approach. Calculations based on the multi-layer perceptron architecture for predicting 45 out of 50 indicators gave more accurate forecasts compared to the basic regression model.
Keywords: artificial neural networks, socio-economic indicators of the Russian Federation, forecasting, time series, hybrid information and analytical system