Within the last 40 years or so there have been a large number of contributions to the scientific literature (journal articles, conference presentations and books) on various aspects of catchment hydrology (rainfall-runoff) models. The focus of these contributions has been on the structure of the models (both mathematical and hydrological), the software used to implement them, methods of calibration (both manual and automatic), estimation of parameters in ungauged situations, their practical application for solving problems and, with increasing frequency in recent years, the uncertainties associated with their outputs. These models are generally defined as mathematical representations of the hydrological cycle at the catchment scale and it is interesting to note how many of the contributions focus on the 'mathematical' part of that definition and how many on the 'hydrological' part. More than 20 years ago, KlemeΕ‘ (1986) contested that for '. . . a good hydrological model it is not enough to work well. It must work well for the right reasons'. The main focus of this commentary is whether or not that message is true, whether active researchers and practitioners recognize it as an issue and whether or not it is possible to decide if a model is working for the right reasons in ungauged (i.e. no observed streamflow data) catchments.
Does the level of complexity influence whether models can work well for the right reasons? The argument for so-called 'physics-based' models arose partly from the desire to include hydrological processes more explicitly (Abbott et al., 1986). However, there seems to be a point at which added complexity in the model structure is not matched by our ability to quantify the model parameters realistically, given typically available data resources. This is an important consideration in data scarce areas which are often the very areas where hydrological models have the potential to provide valuable information for the purposes of water resources planning and management. Beven (1989) questioned the use of algorithms based on small-scale process observations in models that are used at much larger scales. At the other end of the complexity scale, it is perhaps difficult to imagine how it is possible to assess whether highly simplified models are doing the right thing for the right reason. Simple models may be easier to calibrate, particularly when automatic calibration approaches (Vrugt et al., 2003) are being used, but this then becomes a largely mathematical fitting exercise. The inevitable process lumping that occurs in simple models implies that the parameters have little physical meaning, are just mathematical constants and are difficult to extrapolate to ungauged catchments. The desire for parsimonious models (Jakeman and Hornberger, 1993; Perrin et al., 2001), expressed by some contributions to the literature, must surely emanate from the school of hydrological modellers who are looking for elegant mathematical solutions to a complex input (i.e. climate signals)-output (streamflow) problem. It has been claimed in the past (Loague and Freeze, 1985) that more complex models have too many parameter interactions and are fraught with problems of equifinality and lack of parameter identifiability (Beven, 2006), whereby different parameter sets can result in very similar streamflow