Abstract Hydrological models are calibrated on specific periods based on how well simulations match streamflow observations by selecting the parameter vector(s) that provide the highest accuracy. This accuracy can decrease significantly during extrapolation to periods not seen during calibration, especially when they are characterized by different climate conditions. Here, we develop a novel calibration objective that relies on the joint calibration of model accuracy and time‐invariance of residuals, hypothesizing that such invariance can improve the temporal generalizability of hydrological models. We test this approach using a hydrological model and 208 catchments in Western Germany. We evaluate it based on (a) the accuracy loss between calibration and evaluation, (b) the compromise between accuracy and residual invariance, (c) the dependence of residuals on variations in forcings, and (d) the instability of parameters between two calibration periods. We find that our approach reduces the loss in accuracy between calibration and extrapolation as compared to calibration approaches solely based on accuracy. Furthermore, our method combined with the Boxcox streamflow transformation weakens the dependence of residuals on variations in the forcing. However, we find that invariant residuals do not necessarily imply improved parameter stability between different calibration periods. Furthermore, the gain in temporal generalizability comes at the cost of a decrease in accuracy, which should be considered based on the application. Our results highlight that the new RESIdual STability (RESIST) calibration objective has the potential to improve the temporal generalizability of hydrological models for climate‐impact studies.
Astagneau et al. (Sun,) studied this question.