MACHINE LEARNING APPROACHES TO PREDICTIVE MODELING OF LIVESTOCK DEVELOPMENT

The subject of this research is the forecasting of the population dynamics of sheep and goats in Ukraine under wartime, economic, and climatic challenges that directly affect livestock farming and national food security. The present study is of particular pertinence, given the pivotal role that small ruminants play in ensuring the supply of meat, dairy products and wool. A decline in animal numbers may result in shortages of livestock products, deterioration of the socio-economic situation in rural areas, and a reduction in Ukraine's export capacity. The purpose of the paper is twofold: firstly, to develop short- and medium-term forecasting models for sheep and goat populations by combining classical statistical techniques with modern machine learning approaches; and secondly, to identify the strengths and weaknesses of these models across different forecasting horizons. The methodological framework utilised is founded upon statistical data of an official nature, as recorded by the Main Department of Statistics in the Odesa region for the period 2007–2025. Four approaches were employed: the SARIMAX statistical model; the additive Prophet model; and LSTM neural networks, which were implemented using PyTorch and Keras. Forecast performance was evaluated using the RMSE, MAE, MAPE and MASE metrics, enabling a comprehensive comparison of the models. The results confirm that the sheep and goat population in Ukraine has been in persistent decline, with sharper falls observed since the start of the full-scale war. The Keras-based LSTM model proved to be the most accurate for short-term forecasts (12 months), while the PyTorch-based LSTM model demonstrated the greatest stability for medium-term predictions (24 months). The SARIMAX and Prophet models achieved moderate accuracy but struggled to reproduce peaks and troughs in the time series. The study's scientific novelty lies in its integration of statistical and neural network forecasting approaches, while explicitly accounting for wartime disruptions and crisis-related factors. This makes the research one of the first attempts to adapt hybrid forecasting models to the Ukrainian livestock sector in wartime. The practical value lies in the ability to use the developed models at both the micro and macro levels. Individual farms can use them to optimise production and resource planning, while policymakers can use the forecasts to design effective state support programmes, strengthen food security strategies and ensure the sustainable recovery of the agricultural sector.