Technical note: Benchmarking large-domain model performance under sampling uncertainty

Abstract. Large-domain hydrologic modeling studies are becoming increasingly common. The evaluation of the resulting models is however often limited to the use of aggregated performance scores that show where model accuracy is higher and lower. Moreover, the inherent uncertainty in such scores (i.e., the sampling uncertainty), stemming from the choice of time periods used for their calculation, often remains unaccounted for. Here we use a collection of simple benchmarks whilst accounting for this sampling uncertainty to provide context for the performance scores of a large-domain hydrologic model. These benchmarks are simple ways of predicting the variable of interest and include, for example, the long-term daily mean flow, daily precipitation scaled by the average rainfall-runoff ratio, and a basic 2-parameter model that represents a catchment's diffusive response to precipitation inputs. Our test case consists of simulations from the National Water Model v3.0 for approximately 4900 basins across the United States. The benchmarks suggest that there are considerable constraints on the model's performance in approximately one-third of the basins used for model calibration and in approximately half of the basins where model parameters are regionalized. Sampling uncertainty has limited impact: in most basins the model is either clearly better or worse than the benchmarks, though numerous cases remain where sampling uncertainty makes it difficult to clearly distinguish between model and benchmark performance. The areas where the benchmarks outperform the model only partially overlap with areas where the model achieves lower performance scores, and this suggests that improvements may be possible in more regions than a first glance at model performance values may indicate. A key advantage of using these benchmarks is that they are easy and fast to compute, particularly compared to the cost of configuring and running the model. This makes benchmarking a valuable tool that can complement more detailed model evaluation techniques by quickly identifying areas that should be investigated more thoroughly.