Monthly Empirical Quantile Mapping for Bias Correction of GEOGloWS Streamflow Simulations: Calibration and Temporal Validation Across 18 GRDC Stations in Romania

The GEOGloWS ECMWF Streamflow Service provides global river discharge forecasts but exhibits systematic biases in regions without local calibration. This study evaluates Monthly Empirical Quantile Mapping (MEQM) for bias correction of GEOGloWS historical streamflow simulations across 18 Global Runoff Data Centre (GRDC) gauging stations in Romania, covering diverse hydrological regimes from small tributaries to the Danube River. A split-sample validation approach was employed: calibration over 1991–2020 (30 years) and independent temporal validation over 1985–1990 (6 years). MEQM applies non-parametric quantile mapping separately for each calendar month using empirical cumulative distribution functions with approximately 11 quantiles per month (Sturges rule). During calibration, MEQM improved the Kling-Gupta Efficiency (KGE-2012) at all 18 stations, with mean KGE increasing from 0.26 to 0.68 and median absolute percent bias (|PBIAS|) decreasing from 22.34% to 1.52%. Seventeen of 18 stations achieved KGE ≥ 0.50, and 11 stations reached KGE ≥ 0.70. However, temporal validation revealed limited transferability: mean KGE decreased to 0.35, with only 7 of 18 stations maintaining KGE ≥ 0.50, and median |PBIAS| increased to 41.91%. Danube River stations (Baziaș, Orșova, Ceatal Izmail, Zimnicea) demonstrated excellent temporal stability (KGE 0.68–0.74 in validation), while small-to-medium tributaries showed severe performance degradation. Hydrological non-stationarity between the pre-1990 and post-1990 periods, likely driven by land-use changes following Romania’s 1989 socio-economic transition, was identified as the primary factor limiting temporal transferability. These findings highlight the importance of temporal validation for bias correction methods and suggest that MEQM is robust for large, regulated rivers but requires periodic recalibration for smaller basins subject to non-stationary hydrological regimes. To promote reproducibility and community reuse, all analysis code is released as an open-source Python package under the MIT license (https://github.com/flor187/gaf).