Authors - Mark Fedorchenko, Olena Kopishynska, Yurii Utkin, Igor Sliusar, Leonid Flehantov, Olha Barabolia, Nadiia Protas, Tetiana Dugar Abstract - Crop yield forecasting based on small official statistics is different from forecasting with dense satellite, field, or weather datasets: the sample is short, temporal leakage is easy to introduce, and machine learning (ML) should not be accepted unless it beats transparent baselines. This paper presents a baseline-first and reliability-aware workflow for farm management and regional advisory systems. Wheat, maize, and sunflower are evaluated for Poltava, Vinnytsia, Cherkasy, and national-level Ukraine data for 2010-2024. ElasticNet, XGBoost, and LightGBM are compared with naive lag-1, linear-trend, LINEST, and Autoregressive Integrated Moving Average (ARIMA) baselines under a forward temporal design. The contribution is a decision layer that recommends ML only after it clears a practical mean absolute error (MAE) margin and then reports empirical validation-residual bands, test coverage, feature-group diagnostics, and compact farm management systems (FMS)-compatible forecast cards. The Poltava workflow recommends FORECAST.LINEAR for wheat (MAE 0.49 t/ha), LightGBM for maize (MAE 0.69 t/ha), and LightGBM for sunflower (MAE 0.04 t/ha). Across the external check, ML is recommended in 7 of 12 region-crop cases. The results show that ML can help in small official-statistics settings only when checked against simple baselines and reported with reliability diagnostics.
Open Zoom