Justification of agricultural production forecasts and problems of their relevant implementation (on the example of the Orel region)

In the paper, the authors tried to summarize the existing developments on issues of agribusiness-foresight methodology, and specifically on the justification of agricultural production forecasts using machine learning methods. The main goal was to form a forecast for the next three years on the volume of agricultural production in the Orel region at actual and comparable prices. Besides, the production data for the Russian Federation as a whole and price indices of agricultural producers were used. The work used "classical" methods of time sequences modeling: OLS, ETS, ARIMA, their derivatives and combinations. More complex algorithms, based on bagging, boosting or deep learning were not considered, as the original data would not give a significant increase in prediction accuracy. Also, the univariate data was the main analysis, with exclusive inclusion of additional measurement in individual models. The operation algorithm used in the machine learning is shown in details. The optimal model was selected on the learning sample, and the models were validated using the RMSE loss function on the test sample. The first step on the training sample was to select parameters for the main series. For the second and third steps auxiliary models for the two-dimensional methods of the first step were selected. As a result, a short-term three-year forecast was calculated in actual and comparable prices, and the limits of confidence intervals were determined. Considering the complex of crisis phenomena of 2020, the problems of choosing a scenario of possible movement of production dynamics were discussed. In the face of increasing uncertainty, decision-making in management should be based on an appropriate methodological basis.