Projections are rooted in either glocal climate model (GCM) or regional climate model (RCM) data. The differences between GCMs and RCMs are not as clear as their names would lead you to believe - even the RCM relies on information from the GCM. RCMs typically include greater detail (i.e., better refined topography features and more complex physics calculations) in the model for a specific region to presumably improve the simulation. Computation limitations currently make it uneconomical to run a global model at such high spatial resolution, so RCMs are used to dynamically downscale (increase information) over a subset of the globe. However, the climate for a region is not independent of the climate processes outside of that region, so a GCM is used to calculate the conditions at the boundary of the region simulated by the RCM. The ability of the RCM to correctly simulate the regional climate depends on the ability of the GCM to correctly simulate the boundary conditions for that region and important localized climate-related dynamics. The quality of the information coming from the RCM depends on the quality of the information provided by the GCM, so ultimately it is important to know the strengths and weaknesses of the GCM. read more about GCMs and RCMs
There is another kind of downscaled projection that is not based on physics like the RCM. Several statistical methods have been developed to downscale climate projections from GCMs. There are several important differences between dynamical and statistical downscaling that should be considered as part of the evaluation of the quality of information that downscaled projections provide. One important distinction is that most of the advantages of statistical over dynamical downscaling are in the stage of the data processing. It is much less computationally expensive to produce statistically downscaled projections than run the RCM for several decades to produce the same amount of information. However, if you have the choice between using statistically or dynamically downscaled projections that have an identical spatial and temporal coverage, there are many more advantages to using the dynamically downscaled projection. Mainly, the climate information in the dynamically downscaled (RCM) projection is based on high resolution simulations of climate processes whereas the information in the statistically downscaled projection is based on the statistical relationship between observed large- and small-scale variables. A major limitation of statistical downscaling is that it assumes the statistical relationships between large- and small-scale climate variables are stationary (do not change over time), which is not necessarily true under changing climate conditions. For example, lake-effect snowfall in the Great Lakes region is related to how much ice is covering the lakes in winter, and if lake-ice continues to decrease in the future we can expect a new climate regime for lake-effect snowfall. Under the assumptions of statistically downscaled projections the relationship between lake-ice and lake-effect snowfall is not allowed to change. read more on the differences between statistical and dynamical downscaling