Intelligent Analysis and Forecasting of Fruit Crop Yields in Tajikistan: A Comparative Study of Time Series and Machine Learning Models

Abstract

This paper presents the findings of the first comprehensive study on data mining and forecasting of fruit crop yields in the Republic of Tajikistan based on time series covering 1991-2030. Missing data for 1992-1998 were reconstructed using monotonic piecewise cubic Hermite interpolating polynomials (PCHIP). Preliminary statistical analysis employing the Jarque-Bera, ADF, and KPSS tests revealed the series to be nonstationary, necessitating the use of models robust to this property. Six forecasting models were compared: ARIMA, Prophet, exponential smoothing (ETS), a linear trend, a hybrid Prophet-ARIMA model, and XGBoost gradient boosting. Forecast quality was assessed over the 2024-2030 test period using MAE, RMSE, and MAPE. The Prophet model delivered the best performance (MAPE $=\mathbf{6. 9 1 \%})$, demonstrating an ability to capture a non-linear trend under non-stationary conditions. Residual diagnostics (Ljung-Box test, $\mathbf{p}=\mathbf{0. 2 6 3})$ confirmed the adequacy of the model. A yield forecast for 2031-2035 was obtained, with 95% confidence intervals indicating a steady increase to 179.3 centners per hectare. Additionally, statistically significant differences in productivity across categories of farms were identified. The results may serve as a basis for strategic planning in the agricultural sector and illustrate the effectiveness of machine learning techniques when analysing non-stationary agricultural time series in Central Asia.