Improving tools for predicting agricultural development: a combination of recurrent neural networks and multiple regression for regional analysis

the article considers the urgent task of improving the accuracy and objectivity of forecasting the economic dynamics of the agro-industrial complex at the regional level. An algorithmized methodological approach developed by the author for combined medium-term forecasting of agricultural development indicators based on the combined use of recurrent neural networks (RNNW) with long-term short-term memory architecture (LSTM) and methods of correlation and regression analysis is presented and tested. The study was conducted on the example of the agro-industrial complex of the Kaliningrad region using statistical data for the period from 2000 to 2023. The advantage of the proposed approach is substantiated in terms of the limited statistical base and the relative simplicity of predictive calculations compared with traditional methods. The work details the stages of building a neural network model and multiple regression models for key indicators of agriculture and gross regional product. The analysis of the results revealed the polarization of the forecast estimates obtained by various tools, which allows for a more comprehensive understanding of possible development trajectories. The effectiveness of using a neural network apparatus for predicting the regional and sectoral dynamics of the agro-industrial complex is proved even with a limited amount of historical data. Conclusions about the expediency of using a combined approach to complement forecasts and increase their validity are formulated, as well as directions for further research are outlined.

agro-industrial complex, regional economy, forecasting, neural networks, correlation and regression analysis, econometric modeling, Kaliningrad region, combined approach, economic dynamics