This is an empirical work showing numerically the importance of taking into account non-stationarities in CGMY and similar models for computing VaRs. More precisely we model S&P 500 logs with a CGMY model, but where the parameters are allowed to vary. We show that C, G, M and Y in- deed display significant variations across time. Most relevant is the variation of Y , which implies that the correct modelling frame is the one of tempered multistable motions, a recently introduced class of stochastic processes. We provide experimental evidence that these variations have a notable impact on VaR estimation, and thus should be taken into account for proper risk evaluation.