Is considering runs (in)consistency so useless for weather forecasting?

Abstract. This paper addresses the issue of forecasting the weather, using consecutive runs of one given numerical weather prediction (NWP) system. In the literature, considering how forecasts evolve from one run to another has never been proved relevant to predicting the upcoming weather. That is why the usual approach to deal with this consists of blending all together the successive runs, which leads to the well-known "lagged'' ensemble. However, some aspects of this approach are questionable, and if the relationship between changes in forecasts and predictability has so far been considered weak, this does not mean that the door is closed. In this article, we intend to further explore this relationship by focusing on a particular aspect of ensemble prediction systems, the persistence of a given weather scenario over consecutive runs. The idea is that, the more it persists over successive runs, the more it is likely to occur, but its likelihood is not necessarily estimated as it should be by the latest run alone. Using the regional ensemble of Météo-France, AROME-EPS, and forecasting the probability of certain (warning) precipitation amounts being exceeded in 24 hours, it has been found that reliability, an important aspect of probabilistic forecast, is highly sensitive to that persistence. The present study also shows that this dependency can be exploited to improve reliability, for both lagged ensembles and individual runs. From these results, some recommendations for forecasters are made, and the use of new predictors for statistical post-processing, based on consecutive runs, are encouraged. The reason for such sensitivity is also discussed, leading to a new insight on weather forecasting using consecutive ensemble runs.