The transformation of floating-point programs into fixed-point equivalent programs is a challenge because the behaviors of these arithmetics are different. Generally, this transformation is facilitated thanks to a two steps analysis of the design. The first step consists in inferring the range of values for each variables. The second step is the analysis of the precision of the computations taking into account the rounding-errors. In our talk, we will present a new statistical method for inferring dynamic range of floating-point variables in Simulink models. Our approach is based on the extreme value theory. Experimental results will also be presented.