Numerical weather prediction has seen, in the past 25 years, a shift from a deterministic' approach, based on single numerical integrations, to a probabilistic one, with ensembles of numerical integrations used to estimate the probability distribution function of forecast states. This shift to a probabilistic approach enabled a better extraction of predictive signals at longer lead times and provided a meaningful framework for extending the forecast length beyond 10 days. In this work, the limit of predictive skill is assessed for ECMWF monthly ensemble forecasts at different spatial and temporal scales. The forecast skill horizon is defined as the lead time when the ensemble ceases to be more skilful than a climatological distribution, using a continuous ranked probability score as metric. Results based on 32-day ensemble forecasts indicate that the forecast skill horizon is sensitive to the spatial and temporal scale of the predicted phenomena, to the variable considered and the area analysed. On average over 1 year of forecasts, the forecast skill horizon for instantaneous, grid-point fields is between 16 and 23 days, while it is considerably longer for time- and spatial-average fields. Forecast skill horizons longer than the 2 weeks that were thought to be the limit are now achievable thanks to major advances in numerical weather prediction. More specifically, they are possible because forecasts are now framed in probabilistic terms, with a probability distribution estimated using ensembles generated using forecast models that include more components (e.g. a dynamical ocean and ocean waves) and more faithfully represent processes. Moreover, the forecasts start from more accurate initial conditions constructed using better data-assimilation methods and more observational data.
