¹Iğdır University, Faculty of Agriculture, Agricultural Economics Department, TR-76000 Iğdır - TÜRKİYE
²Iğdır University, Iğdır Vocational School, Transportation Services Department, TR-76000 Iğdır - TÜRKİYE
³Iğdır University, Iğdır Vocational School, Finance-Banking and Insurance Department, TR-76000 Iğdır - TÜRKİYE
⁴Akdeniz University, Faculty of Economics and Administrative Sciences, Econometrics Department, TR-07058 Antalya - TÜRKİYE DOI : 10.9775/kvfd.2026.37347 Goat milk is a valuable food source because of its digestibility, nutritional value and suitability for value-added dairy products. This study forecast Türkiye's annual raw goat milk production for 2025-2029 and compared ARIMA with Decision Tree, MARS, XGBoost and SVM models using FAOSTAT data for 1961-2024. Nested expanding-window cross-validation, in which hyperparameters were tuned within each training window without access to the test year, produced 36 one-year-ahead forecasts for 1989-2024. Although MARS had the lowest errors (RMSE 33.07 versus 34.73 thousand tonnes for ARIMA), no machine-learning model differed significantly from ARIMA in Holm-adjusted Diebold–Mariano tests. The parsimonious ARIMA (2,1,0) model was therefore retained for probabilistic projections. Its point forecast declined from 468.22 thousand tonnes in 2025 to 420.39 thousand tonnes in 2029, whereas the 95% residual-bootstrap interval widened to 204.46-608.23 thousand tonnes, spanning nearly the entire historical range of the series. Sensitivity checks using post-2009 data and a fixed rolling window reproduced these conclusions. The forecasts therefore indicate substantial uncertainty rather than a firm decline and do not by themselves support specific policy measures; productivity, value-added processing and agricultural data systems warrant further investigation. Keywords : Agricultural planning, ARIMA, Cross-validation, Forecasting, Goat milk, Machine learning, Model comparison, Time series analysis, Türkiye