Neural network forecasting of interbudget transfers based on regional socioeconomic indicators

The modern digitalization of public administration increases the need for tools for accurate budget forecasting. Despite the accumulated research in the field of interbudgetary relations, the problem of identifying the quantitative relationship between the socio-economic indicators of regions and the distribution of federal transfers, especially with the use of artificial intelligence methods, remains insufficiently studied. The lack of studies using neural networks to forecast financial assistance to the constituent entities of the Federation and the high demand for improving the efficiency of budget planning were the key reasons for choosing this topic. The objective of the study is to identify the relationship between the indicators of regional socio-economic development and the volume of interbudget transfers, as well as to build a model for forecasting future transfers using the backpropagation method. The objectives include forming a data sample, constructing the structure of the neural network, its training, and testing. The methodological basis is a fully connected multilayer perceptron with two hidden layers of 34 neurons, trained on normalized data using the backpropagation method. The sample includes 16 annual socio-economic indicators and quarterly data on transfers for 85 regions of the Russian Federation for 2015-2021, ensuring high model representativeness. The sample timeframe is limited to 2022, as data for the subsequent period are characterized by changing external conditions, which affects their comparability with the previous time series. The results demonstrate a steady reduction in error in the test sample and confirm the existence of a statistically significant relationship between socio-economic parameters and transfer volumes. The model successfully predicts the amount of transfers for the next quarter. The practical significance of the model lies in its applicability to decision support in planning federal financial assistance and developing a «model budget.» Limitations include the absence of qualitative factors and political and administrative parameters in the sample. Future research could aim to include a wider range of indicators and compare neural networks with alternative machine learning methods.