<p>The present study highlighted the capability of machine learning (ML) models, combined with the empirical Natural Resources Conservation Service (NRCS) method, to simulate the complex rainfall-runoff relationship. The novel hybrid approach presented a robust foundation for enhancing runoff prediction by integrating ML’s pattern recognition ability with the NRCS method’s empirical reliability. Four hybrid ML-NRCS models were evaluated: Random Forest (RF) -NRCS, K-nearest neighbors (KNN)-NRCS, Extreme Gradient Boosting (XGBoost)-NRCS, and Decision Trees (DT)-NRCS. The evaluation utilized 31&#xa0;years of daily rainfall data from the Erbil meteorological station, along with corresponding runoff data calculated using the NRCS method. The XGBoost-NRCS model outperformed other models by maximizing the coefficient of determination (R²) and Nash–Sutcliffe Efficiency (NSE) and minimizing the Mean Squared Error (MSE) and Mean Absolute Error (MAE). The XGBoost-NRCS model improved model generalization, providing a significant edge over other ML models in identifying critical patterns and handling sparse hydrological data. Furthermore, the findings revealed that all monthly-based models deliver better results than daily based models, suggesting that data aggregation enhances prediction accuracy. As a validation step, the models were tested using the dataset of the Sulaymaniyah meteorological station. The XGBoost-NRCS model yielded excellent predictions of monthly runoff but less effective predictions of daily runoff due to the high variability in daily rainfall data across diverse regions. A key finding of this study is the hybrid XGBoost-NRCS model’s strong predictive capability and broader applicability across different datasets. This study introduces a novel and scalable methodology for improving runoff estimation, particularly in data-scarce environments.</p>