Model-Inspired Predictors for Model Output Statistics (MOS)

Abstract

This article addresses the problem of the choice of the predictors for the multiple linear regression in model output statistics. Rather than devising a selection procedure directly aimed at the minimization of the final scores, it is examined whether taking the model equations as a guidance may render the process more rational. To this end a notion of constant fractional errors is introduced. Experimental evidence is provided that they are approximately present in the model and that their impact is sufficiently linear to be corrected by a linear regression. Of particular interest are the forcing terms in the coupling of the physics parameterization to the dynamics of the model. Because such parameterizations are estimates of subgrid processes, they are expected to represent degrees of freedom that are independent of the resolved-scale model variables. To illustrate the value of this approach, it is shown that the temporal accumulation of sensible and latent heat fluxes and net solar and thermal radiation utilized as predictors add a statistically significant improvement to the 2-m temperature scores.