Statistical downscaling

Statistical downscaling estimates local or regional climate from large-scale atmospheric predictors using relationships fitted to observations. A transfer function — regression, analogue method, weather typing, or increasingly a neural network — is trained to map coarse fields onto local conditions, then applied to model output or reanalysis.

Against dynamical downscaling the practical difference is cost. A dynamical run takes substantial HPC time for one realisation; a fitted transfer function applies to a large ensemble across many scenarios almost for free. Where the question needs many realisations rather than one detailed one, that is decisive.

The assumption that does the damage

Every statistical method assumes the fitted relationship remains valid outside the range it was fitted on. This is the stationarity assumption, and under a changing climate it is exactly the assumption that cannot be relied upon.

The relationship between a large-scale predictor and local temperature holds in the observed record because of physical processes operating in the climate that produced that record. Change the climate and those processes may change too — a circulation pattern that reliably produced local warmth may stop doing so, or the local response to it may shift. Nothing in the fitting detects this, and the method produces confident output regardless.

The problem is worst precisely where the answer matters most. Extremes are rare in the training record by definition, so the transfer function is least constrained there, and the future contains conditions the record does not. A method validated well on the observed period can be systematically wrong about the period it was built to say something about.

This is the concrete case for process understanding over statistical description. A dynamical model simulates physics, so it can produce conditions unlike anything in the observed record; a fitted relationship can only extrapolate one.

Where it is the right tool

• Bias correction of model output against observations, which is nearly universal in impact studies and is statistical downscaling under another name.
• Large scenario ensembles where dynamical simulation is infeasible and the alternative is no local information at all.
• Regions and variables where the physical relationship is strong, well understood, and unlikely to change — orographic temperature lapse being the friendliest case.

The defensible use is with the assumption stated rather than assumed, and ideally checked against a dynamical run over some part of the range. Presenting statistically downscaled projections without that caveat is where the method gets misused.

See also: dynamical downscaling, regional climate modeling, physical climatology, and climate on the stationarity problem generally.