Location errors in QPFs over the Calabria region: Does a multi-model poor man's ensemble over-perform each member?

The aim of this work is to evaluate the quality, in terms of location errors, of multi-model poor man's ensemble (PME) forecasts against the single model ones over the Calabria region. Several strategies were adopted to combine precipitation forecasts by three limited area models (LAMs), namely the mean, the median, and a probabilistic matching approach. The Contiguous Rain Area (CRA) analysis was the method selected to detect and quantify the location errors of the forecast precipitation patterns with respect to the corresponding rain gauge-based analyses. Two best-fit criteria, the minimization of mean squared error and the maximization of correlation coefficient, were chosen for matching forecast and observed features. The ability to forecast correctly the precipitation patterns was then quantified by means of a summary measure, the CRA mean shift (CMS). It condenses the outcomes of the twenty-month CRA analyses with a unique value. A bootstrap procedure was applied to test the statistical significance of differences among CMS indices of LAMs and PMEs. Despite the ensemble forecasts display a general improvement, which results in a lower CMS index, with respect to the single LAMs, such improvement was not statistically significant for most ensembles. When the best-fit criterion is the maximization of correlation coefficient, no ensemble was statistically significant better than single models. Instead when the minimization of mean squared error was chosen as best-fit criterion, two out of four PMEs were significantly better than at least a LAM.