Extreme flood simulation with synthetic extreme precipitation events raises unavoidable questions about the choice of initial conditions. State-of-the-art extreme flood estimation frameworks propose to address these questions with the help of semicontinuous modeling and reanalysis of simulated state variables. In this context, the present work proposes a new method for the selection of initial conditions for extreme flood simulation. The method is based on generating sets of initial conditions from the matrix of state variables corresponding to a long simulation run of the selected hydrological model. Two sets of initial conditions are obtained: a deterministic set composed of selected state variable quantiles and a stochastic set composed of state variable vectors randomly drawn from the complete state variable matrix. The extreme flood simulations corresponding to both sets are compared in detail, and the stochastic simulations are used in a sensitivity analysis to identify the dominant state variables and possible interactions. The aim hereby is to provide a tool to analyze the role of initial conditions and the importance to account for state variable interactions in extreme flood estimation. The proposed method is applied to probable maximum flood estimation for the Swiss Mattmark Dam catchment with a semilumped hydrological model. The obtained results for this case study show that for high flood peak quantiles, the initial soil saturation is dominating other state variables, and deterministic initial conditions are sufficient to generate extreme floods.