Multivariate postprocessing using Cholesky-based multivariate Gaussian regression

To obtain reliable joint probability forecasts, multivariate postprocessing of numerical weather predictions (NWPs) must take into account dependencies among the univariate forecast errors—across different forecast horizons, locations or atmospheric quantities. We develop a framework for multivariate Gaussian regression (MGR), a flexible multivariate postprocessing technique with advantages over state-of-the-art methods.

In MGR both mean forecasts and parameters describing their error covariance matrix may be modeled simultaneously on NWP-derived predictor variables. The bivariate case is straightforward and has been used to postprocess horizontal wind vector forecasts, but higher dimensions present two major difficulties: ensuring the estimated error covariance matrix is positive definite and regularizing the high model complexity.

We tackle these problems by parameterizing the covariance through the entries of its basic and modified Cholesky decompositions. This ensures its positive definiteness and is the crucial fact making it possible to link parameters with predictors in a regression.  When there is a natural order to the variables, we can also sensibly reduce complexity through a priori restrictions of the parameter space.

MGR forecasts take the form of full joint parametric distributions—in contrast to ensemble copula coupling (ECC) that obtains samples from the joint distribution. This has the advantage that joint probabilities or quantiles can be easily derived.

Our novel method is applied to postprocess NWPs of surface temperature at an Alpine valley station for ten distinct lead times more than one week in the future.  All the mean forecasts and their full error covariance matrix are modelled on NWP-derived variables in one step. MGR outperforms ECC in combination with nonhomogeneous Gaussian regression.