Comparative analyses are conducted to assess the impact of streamflow data assimilation on the quality of hydrological forecasts. Hydrological predictions are issued by a lumped soil-moisture-accounting type rainfall-runoff model (GRP model developed at Cemagref). The model was applied to a large set of catchments in France, representative of a variety of climate and physiographic conditions. In this study, results from the evaluation of short-term streamflow forecasts (up to 2-3 days) are presented. Two configurations were considered: 1) the model is driven by "perfect precipitation forecasts", i.e., observed precipitation is used as input, and 2) the model is driven by a weather ensemble prediction system. In the first case, different assimilation techniques are tested to update the model's parameters, states and outputs at an hourly time step. The main results show that state updating proved to be more effective than output updating for the forecasting model studied. Additionally, optimising the model's parameters directly for a forecasting objective proved to be more efficient than optimising the hydrological model in a simulation mode, and then adjoining an updating procedure when forecasting. In the second configuration considered, weather forecasts from the ensemble prediction system PEARP of Météo-France (11 perturbed members and a forecast range of 60 hours) are used. Based on the evaluation of typical forecast verification scores, the impact of the model state updating is assessed at the daily time step. The results highlight the benefits of streamflow data assimilation for ensemble short-term forecasting.