Common procedures for flood frequency modeling are based on analysis of series of annual maxima (AMS) of the instantaneous discharge. A practical alternative to AMS series is usually considered when the record length is short. In this case, analysis of partial duration series (PDS) on continuous streamflow records allows one to increase the available information by using more than one flood peak per year. In this paper, the analysis of continuous streamflow data is reconsidered from the beginning: a filtered peaks over threshold (FPOT) procedure is proposed as an alternative to the PDS approach, and objective criteria are devised for choosing a reasonable threshold and for determining the average annual number of events λ. The revised procedure demonstrates that there is no need for specific limitations on the magnitude of λ to preserve the basic hypotheses of the marked point process build in the PDS procedure. The proposed FPOT method is applied to 33 time series of daily runoff from rivers of northwestern Italy. Significant advantages over the classical PDS method are demonstrated, both in terms of physical interpretability of the parameters and efficiency of the estimates.