In this paper, a stochastic model is used to simulate daily minimum temperature time series coming from observations and two CMIP5 climate models (IPSL-CM5A-MR and CNRM-CM5) in order to analyze the changes in cold wave number and proportions under future climate conditions. The stochastic model allows computing 100 temperature time series for each different source (observation or climate model), and for 22 locations in France, which enables inferring the statistical significance of the changes. Two future time periods, near (around 2010–2060) and far future (around 2050–2100), and two RCPs (RCP4.5 and RCP8.5) are considered, while 3 different thresholds are used to identify cold waves: 0 °C and the 10th and 5th percentiles of observed wintertime (December–January–February) daily minimum temperature distribution. The results show that both models project a significantly lower number of cold waves in the future, all durations considered, but the changes mainly concern the proportion of the longest cold waves (10 days and more). The decreases are higher with IPSL-CM5A-MR than with CNRM-CM5. The main driver of this change is the decreasing frequency of the observation based thresholds in the future, which is higher for IPSL-CM5-MR model because the impact of a higher mean is enhanced by a decrease in the variance.