Streamflow prediction with machine learning: evaluating predictability across hydroclimatic regimes

Accurate streamflow prediction is crucial for water resources management under increasing climate variability. This study compares the performance of Random Forest (RF) and Extreme Gradient Boosting (XGBoost) across 24 river basins with varying climates and flow regimes. Using 43 years (1980–2022) of daily rainfall and streamflow data, models were trained via Recursive Feature Elimination and validated using rolling cross-validation. Metrics used include NSE, RMSE, MAE, and Pbias. RF outperformed XG-Boost in 18 basins, especially under intermittent and low-flow conditions. Both models performed best (NSE > 0.90) in humid, stable basins, but accuracy declined (NSE < 0.5) in arid and variable catchments. Streamflow variability was a stronger predictor of model performance than rainfall variability. RF showed greater robustness in noisy datasets, while XGBoost excelled in high-flow settings. The study presents a diagnostic framework for selecting ML models based on hydroclimatic and flow regime characteristics.