Graphical models for the evaluation of multisite temperature forecasts: comparison of vines and independence graphs

摘要

Vine and independence graphs are employed to extract conditional independence relations from multivariate meteorological data so as to construct a simple graphical model which adequately represents the interrelationships between observations and corresponding model results at different sites. The independence graph approach identifies partial correlations of maximal order. Statistically negligible partial correlations are set to zero. Iterative proportional fitting is used to find a maximum likelihood distribution satisfying the stipulated zero partial correlations. The deviance between the fitted distribution and the original distribution measures the goodness of fit. The vine approach constructs a regular vine in which negligible partial correlations are set to zero. No proportional fitting is required. Again, deviance is used to measure goodness of fit. The connection between vines and continuous belief nets, where an arc from node i to j is associated with a (conditional) rank correlation between i and j is presented.