FORECASTING KEY INDICATORS OF CROP PRODUCTION USING ECONOMETRIC MODELS AND MACHINE LEARNING METHODS (A CASE STUDY OF THE REPUBLIC OF SAKHA (YAKUTIA)

Abstract

The article presents a comparative testing of econometric and machine learning models for forecasting key indicators of crop production in the Republic of Sakha (Yakutia) under conditions of short time series (2010–2024). Significant differences in the effectiveness of alternative approaches were identified: the inclusion of agroclimatic predictors (growing degree days, precipitation, and average growing season temperature) reduced the yield forecasting error to 5,4% (ARIMAX model) compared to 13,2% for purely trend-based methods. It was established that for forecasting gross harvest, the most balanced results are provided by the ARIMA+ARIMAX ensemble (average MAPE 8,1%) and Ridge Regression (9,0%), whereas Random Forest generates constant predictions under limited sample conditions, making it unsuitable as a primary forecasting tool. Robustness assessment with varying training sample lengths demonstrated the highest stability for Ridge Regression (average MAPE variation range of 6,7 percentage points), while ARIMAX exhibited high sensitivity to outliers at the end of the training period. Comparison of the calculations with official regional forecast documents confirmed a systematic overestimation of planned indicators, caused by independent modeling of sown areas and yield without accounting for their functional relationship and climate risks. The low accuracy of official forecasts is attributed to the high interannual volatility of agroclimatic factors, which is ignored in traditional methodologies. The feasibility of transitioning to combined forecasting schemes is substantiated, in which sown areas are calculated using trend or administrative methods, while yield is modeled with the mandatory inclusion of exogenous weather indices. The proposed approach enables minimizing strategic planning errors, optimizing budget expenditures, and enhancing the reliability of food security risk assessments in the region's zone of risky agriculture.