The performance of the ARPA-SMR Limited-area Ensemble Prediction System (LEPS), generated by nesting a limited-area model on selected members of the ECMWF targeted ensemble, is evaluated for two flood events that occurred during September 1992. The predictability of the events is studied for forecast times ranging from 2 to 4 days. The extent to which floods localised in time and space can be forecast at high resolution in probabilistic terms was investigated. Rainfall probability maps generated by both LEPS and ECMWF targeted ensembles are compared for different precipitation thresholds in order to assess the impact of enhanced resolution. At all considered forecast ranges, LEPS performs better, providing a more accurate description of the event with respect to the spatio-temporal location, as well as its intensity. In both flood cases, LEPS probability maps turn out to be a very valuable tool to assist forecasters to issue flood alerts at different forecast ranges. It is also shown that at the shortest forecast range, the deterministic prediction provided by the limited area model, when run in a higher-resolution configuration, provides a very accurate rainfall pattern and a good quantitative estimate of the total rainfall deployed in the flooded regions.