Comparative Assessment of Machine Learning and Statistical Models for Precipitation Forecasting in Semi-Arid Regions

Accurate precipitation forecasting is critical for climate resilience, water resource management, and agricultural planning, particularly in semi-arid regions with high rainfall variability. This study provides a comprehensive comparison of statistical, machine learning, deep learning, and hybrid approaches for monthly precipitation forecasting in Konya Province, Türkiye, using long-term climatic data from 1958 to 2025. The evaluated models include SARIMAX, LASSO, CatBoost, LightGBM, XGBoost, Multilayer Perceptron (MLP), Long Short-Term Memory (LSTM), and a Hybrid LSTM–XGBoost framework. Model inputs consist of key meteorological variables—temperature, humidity, evaporation, sunshine duration, and wind speed—together with seasonal indicators and lagged or rolling features, depending on model structure. A chronological 70/10/20 train–validation–test split was applied to all machine learning models, while SARIMAX incorporated exogenous variables for time series modeling. Model performance was evaluated using R2, RMSE, and MAE. The results show that tree-based ensemble models (CatBoost, LightGBM, and XGBoost) outperform traditional statistical approaches by effectively capturing nonlinear relationships and seasonal variability, although they tend to underestimate rare extreme precipitation events. The standalone LSTM model exhibits moderate predictive performance, indicating that temporal dependency modeling alone is insufficient. In contrast, the Hybrid LSTM–XGBoost model achieves the best overall results, demonstrating that integrating temporal learning with nonlinear regression substantially improves forecasting accuracy. The findings highlight hybrid modeling as a reliable framework for precipitation forecasting in semi-arid regions, supporting sustainable water management and climate adaptation efforts.