Floods and dam overtopping risks, especially for earth fill dikes, during strong rainfall, preoccupy the authorities since a long time. In Switzerland, to assure a good protection, the extreme floods associated with very low probabilities must be estimated. Their determination is usually based on the statistical extrapolation of data available for long periods of time. The difficulty of this probabilistic method led to adaptation the concept of PMP-PMF (Probable Maximum Precipitation-Probable Maximum Flood) to Swiss conditions. This paper shows how the PMP-PMF method applied without regard the probability associated to the initial conditions of the simulation may lead exceptional flow values that seem unrealistic. A solution is proposed to estimate such probabilities for various initial conditions such as the rain-snow boundary or the wind direction and speed. These occurrence probabilities can then be included in the PMP-PMF calculations to lend more clarity to the likelihood of the extreme conditions that lead to such an event. For example, a very rare initial condition on the terrain could be combined with not so rare rainfall, so that the joint probability will not be too small, and vice-versa. A linear fit is then applied to the various simulation results in logarithmic scale, allowing for the estimation of the return period for extreme floods based on the estimated probability.