In regions characterized by a great inter-annual variability or by decadal-scale changes of the rainfall regime, the simulation of long series of rainfall events is an efficient way to explore the runoff fluctuations or modifications resulting from this rainfall variability. In a context of great uncertainty regarding the Sahelian rainfall regime in a changing climate, a coherent stochastic framework is presented here to produce high spatial resolution rain fields in order to force a rainfall–runoff model and to perform sensitivity analyses. The focus of the paper is on the comparison of various conditioning methods reflecting the various types of information available for the study of past situations (data from rain gage and satellite) as well as of future scenarios (outputs of atmospheric models). Various types of rainfall simulations are performed over a 13 year period, using four levels of conditioning information obtained from a 15 gauge network covering a 60 × 60 km2 region. These simulations are then used as inputs to a Hortonian rainfall–runoff model. The simulation relevance is first assessed by studying the simulated rain field series (event time-step mean characteristics, seasonal cycle and inter-annual variability) in comparison with reference rain fields estimated by kriging. This shows that the conditioning of the simulations, even by a minimal information provided by a unique station, is of great relevance for constraining the stochastic dispersion and thus to retrieve the rainfall variability at the considered scales. Significant differences are reported between runoff obtained by the different types of created rain fields, one of the most noticeable being that runoff obtained from kriging is 25% lower than runoff obtained from point conditional simulations. The results confirm the sensitivity of Hortonian hydrological systems to rainfall intensity and particularly point out the importance of representing realistic spatial rainfall patterns to force hydrological models.