In Niger, pearl millet is the main staple food and the dominant crop in the traditional agricultural systems with about 75%, of the national total cereal production. The low productivity of the crop in traditional environments and its high variability, combined with the rapid growth of the population, is the principle cause of vulnerability of the population which suffers from recurrent food crises. Since rainfall is the main limiting factor in the productivity of millet in Niger, the present study focuses on the relationships between rainfall and yields through a multi-year 2004-2007 field campaign within the AMMA project and in coordination with the AMMA-CATCH program. We found a high variability of yield from one plot to another and from one village to another within a 10 000 km² area around Niamey. When smoothing the local variability of the yields unlikely linked with rainfall, we found than the variability of yields is not linked with the seasonal rainfall but with the rainfall amount during the critical stages of the crop, i.e. the reproductive and the maturation phases. The rainfall amount during these critical stages explains around 24% of the yield variability. The analysis of both observed and simulated yields from the SARRAH crop model shows that the performances of the crop model depend on its accuracy in simulating the water stress during these critical stages. These results are important in working towards the predictability of crop yields using rainfall information and in particular to highlight the most salient rainfall parameters that arc needed in a forecast system. They are also important to improve the skill of crop models in simulated yields variability.