Future changes and uncertainties due to climate change are most meaningful when placed in context. Therefore, once the variables of interest have been identified, it is important to understand how well they are currently known and can be simulated both now and in an altered climate. This includes model simulation uncertainties (e.g., how well does the hydrology model used capture peak flows, low flows, seasonality), but also other uncertainties that could impact decisions (e.g., operation changes because of aging infrastructure or increases in water demand) (NRC 2009; Brekke et al. 2009).

Ideally, models should represent all relevant processes well. If certain processes are poorly captured, the model’s ability to simulate the climate sensitivities of dominant processes could be in question. Yet models will always be limited by being simplifications of the real world (Clark et al. 2008; Carslaw et al. 2018). Therefore, for practical purposes, models are most often evaluated on how well they do at simulating key, measurable processes, especially those relevant to the impact of interest. For example, if the decisions relate to flooding, then hydrology model performance on short timescales matters. If, however, the decisions relate to water needs for drought, performance on shorter timescales may be less relevant. Evaluations should include how well model outputs are simulated historically (what is the current ability to simulate the variable of interest) and how sensitive they are to an altered climate. The latter can be done through evaluating whether modeled values respond accurately to a range of different climate conditions or through simple perturbations of the most relevant climate variables (e.g., Vano et al. 2012). This does not provide a comprehensive evaluation of how well future changes can be simulated, as this may not be knowable, but it can provide confidence that model sensitivities are physically reasonable and that further exploration using a model or approach is warranted. Additionally, techniques exist that can be used to evaluate how well a model performs under climatic conditions significantly different from those it was developed to simulate (Refsgraad et al. 2013).

Evaluations of how sensitive models are to change can reveal the extent to which climate impacts can be adequately simulated for decision making purposes. In many cases, this provides helpful context for a more detailed analysis.

In some cases, however, this can change the trajectory of the study. There are several possible outcomes:

(1) Evaluations reveal variables of interest are insensitive to climate. This could be because they really are, e.g., rain-dominant basins do not experience a seasonal shift in their hydrograph because there is no snow to melt (Elsner et al. 2010). Or, it might be an artifact of the hydrologic model design (e.g., temperature sensitive parameters, such as evaporative demand, have been fixed) which does not allow the model to account for climate change (Willows and Connell 2003). In these situations, it makes little sense to do full climate simulations unless more climate-sensitive impacts are also of interest (e.g., streamflow temperature in rain-dominant basins) or the hydrologic model is reconfigured to be sensitive to changes in climate (e.g., evaporative demand, the key pathway by which temperature might influence water balance is no longer fixed).

(2) Evaluations reveal other non-climate uncertainties like population changes on exposure to extreme heat (Jones et al. 2015) or land use and resource availability (Olsen et al. 2015) have an equal or greater impact than climate. It is then up to the decision maker to decide which future impacts should be explored first.

(3) Evaluations reveal large variability in results and the climate change signal is less noticeable in the midst of the noise or has not yet emerged from the range in natural variation. This could be either from natural variability of climate systems (e.g., Hamlet 2011) or model uncertainties (Reclamation 2014a). Notably, however, relative uncertainties change depending on the region (local, regional, global), time horizon, and amount of models included to represent the variable of interest. For example, uncertainties from emission levels dominate other types of uncertainties as planning horizons increase (Hawkins and Sutton 2009, 2011) and other sources of uncertainties are introduced as more models are added (Eisner et al. 2017). Overall, even when relative uncertainties are large, there is still a need for understanding the range of plausible futures to use in stress tests, and practices designed for future flexibility and appropriate safety factors or freeboards should be used (Olsen et al. 2015).