What makes a robust calibration period? Insights into the effects of data properties

Hydrological modelling requires robust parameter sets to ensure stable performance across different time periods. This study evaluates 20 split-sample strategies, including calibration on the entire available period, using two monthly lumped models (GR2M and abcd) across 770 global catchments. We analyse robustness and assess how it is influenced by calibration data properties (length, temporal proximity to the application phase, and hydroclimatic features). Results show that the full-period calibration is the most robust strategy, a finding validated across four objective functions. Robustness increases with longer calibration periods and proximity to the application phase. Higher robustness is also found to be linked to (i) higher values of humidity indices (e.g. P/PET, Q/P), (ii) higher variability in the annual runoff regime, and (iii) higher sum of local peaks above certain runoff-dependent thresholds.