22 pages ; International audience ; Using a large set of WRF ensemble simulations at 70-km horizontal resolution over a domain encompassing the Warm Pool region and its surroundings [45°N-45°S, 10°E-240°E], this study aims at quantifying how nudging techniques can modify the simulation of deep atmospheric convection. Both seasonal mean climate, transient variability at intraseasonal timescales, and the respective weight of internal (stochastic) and forced (reproducible) variability are considered. Sensitivity to a large variety of nudging settings (nudged variables and layers and nudging strength) and to the model physics (using 3 convective parameterizations) is addressed. Integrations are carried out during a 7-month season characterized by neutral background conditions and strong intraseasonal variability. Results show that (1) the model responds differently to the nudging from one parameterization to another. Biases are decreased by ~50 % for Betts-Miller-Janjic convection against 17 % only for Grell-Dévényi, the scheme producing yet the largest biases; (2) relaxing air temperature is the most efficient way to reduce biases, while nudging the wind increases most co-variability with daily observations; (3) the model's internal variability is drastically reduced and mostly depends on the nudging strength and nudged variables; (4) interrupting the relaxation before the end of the simulations leads to an abrupt convergence towards the model's natural solution, with no clear effects on the simulated climate after a few days. The usefulness and limitations of the approach are finally discussed through the example of the Madden-Julian Oscillation, that the model fails at simulating and that can be artificially and still imperfectly reproduced in relaxation experiments.