Leveraging GCM-based forecasts for enhanced seasonal streamflow prediction in diverse hydrological regimes

• The quality of streamflow predictions is attributed to local hydrological regimes. • Initial hydrological conditions enhance skill across seasons for up to 4 months. • GCM forcing improves streamflow accuracy, particularly in the first 4–8 lead weeks. • In the snow-driven areas of Sweden, ESP outperforms GCM-based forecasts in winter. • GCM-based streamflow forecasts excel ESP in the southern rainfall-driven areas. Seasonal hydrological forecasts are vital for managing water resources and adapting to climate change, aiding in diverse planning and decision-making processes. Currently it is unknown how different forecasting methods, considering initial hydrological conditions and dynamic meteorological forcing, perform across the Swedish river systems, despite the significant socio-economic implications. Here we explore the drivers that mostly impact streamflow predictions and attribute the added quality of these predictions to local hydrological regimes. We compare the accuracy of seasonal streamflow forecasts driven by dynamic GCM-based meteorological forecasts with those generated by the Ensemble Streamflow Prediction (ESP) method. The analysis spans across about 39,500 sub-catchments in Sweden encompassing various climatic, geographical and human-influenced factors. Results show that the streamflow predictability varies in space due to the country’s diverse hydrological regimes. Regardless of the regime, updating the models to achieve the best possible initial conditions is crucial for enhancing forecast skill across all seasons for up to 4 months. GCM-based meteorological forcing notably improves short-term streamflow accuracy, showing significant impact particularly up to 4–8 weeks lead time depending on the local hydrological regime. In the snow-driven northern regions, ESP demonstrates superior performance over GCM-based streamflow forecasts in winter. Conversely, in the southern regions, where conditions are predominantly influenced by rainfall, GCM-based forecasts show higher performance up to 4–6 weeks ahead, regardless of the season. In river systems with high human influences, streamflow climatology outperforms ESP and GCM-based forecasts underscoring the challenges of accurately modelling artificial reservoir management and the need for better access to management data. These insights guide the development of an advanced national seasonal hydrological forecasting service, and highlight the need for region-specific forecasting strategies indicating areas where predictability is enhanced by improved monitoring, hence initial conditions, and/or meteorological forcings. Finally, we discuss the applicability of these forecasting methods to other regions worldwide, thereby placing our new insights within a global context. Our research focuses on improving streamflow predictability in Sweden, especially looking ahead to the coming weeks to seasons. We compare the Ensemble Streamflow Predictions (ESP) method that uses historical weather data to predict future streamflows, and a dynamic forecasting method based on bias-adjusted ECMWF SEAS5 system that considers current weather conditions and predicts how climate evolves. We cover all river systems across Sweden, each with unique climatic and physiographic characteristics. Streamflow predictions based on SEAS5 are generally more accurate than ESP in the short term, regardless of the season; however, for longer time horizons, ESP often becomes just as good, if not better, especially in winter. We show that one single forecasting method may not work best for all areas depending on the regional hydrological regime and the season. ESP is more effective in handling the variability caused by snow and rain, which is crucial in determining streamflow in the northern mountainous areas. In southern Sweden, where streamflow is more influenced by rainfall than snow, SEAS5 generates more accurate streamflow in the short term for up to 6 weeks ahead. In regions with large rivers controlled by human interventions, neither method consistently outperforms the other, even in the short term.