Extreme floods are caused by special meteorological conditions matching critical space-time scales of flood generation processes in a catchment. Fortunately for most of the floods these conditions do not meet. Objective of this study is to investigate how big could have historical floods become under “optimal” rainfall conditions and which are the main factors driving flood maximization. For that, observed storms are stochastically modified. First the intensities are amplified, then the spatial patterns are changed and finally both conditions are varied together. The rainfall intensities are modified with moisture maximisation considering the observed saturation deficit. The spatial patterns are changed by conditional rainfall simulation considering temporal correlation and advection. The simulated rainfall realisations are then used as input for a rainfall-runoff model to simulate corresponding floods. This case study uses data from the Mulde river basin in Germany and applies the methodology to a set of selected large flood events. The results show that observed events could have been much worse with average amplification factors of 2.9 for rainfall and of 4.6 for peak flow for the worst-case scenario of stationary storms. The developed method could also be used as alternative for the estimation of probable maximum precipitation and probable maximum floods.