Hard to Beat, Not Impossible: Improving Seasonal Ensemble Streamflow Predictions (ESP) in Mountainous Catchments

Abstract. The Ensemble Streamflow Prediction (ESP) method has been extensively applied in operational forecasting and is widely considered a robust benchmark for predicting spring–summer snowmelt runoff in mountain environments. Nevertheless, it remains unclear how, when and where the implementation of data assimilation and/or post-processing techniques offer potential for improving the quality of seasonal and monthly streamflow forecasts during the snowmelt season. Here, we compare alternative ESP hindcast configurations – implemented with the conceptual rainfall-runoff model HBV.IANIGLA – across 21 Andean catchments spanning a pronounced hydroclimatic gradient (28–37° S), focusing on spring-summer (i.e., September–March) streamflow hindcasts produced for five initialization times over a 38-year period (September/1982–March/2020). Data assimilation experiments include the ensemble Kalman Filter (EnKF) and the Sequential Importance Resampling Particle Filter (SIR-PF), and the suite of post-processing techniques includes Quantile Mapping, Linear Regression, Trace Weighting and Random Forest. For completeness, the alternative ESP configurations are also compared against a simple statistical model using initial hydrologic conditions (IHCs) – computed as the sum of simulated soil and snow water storages – as the only predictand. Results show that, at the seasonal scale, the advantage of SIR-PF over EnKF becomes evident only for hindcasts initialized on August 1 or later, while post-processing can substantially degrade skill at lead times of two months or longer. However, combining data assimilation and post-processing for September 1 initializations improves skill and reliability of raw ESP monthly hindcasts, particularly at the beginning and end of the snowmelt season. The simple IHC-based statistical method also improves reliability at both seasonal and monthly scales, despite mixed results for other metrics. Finally, catchment attributes help explain where improvements are most likely: seasonal gains are favored in semi-arid basins with lower precipitation and runoff, higher aridity, and poorer hydrologic model performance, whereas monthly gains show distinct geographic and hydroclimatic patterns.