This contribution presents a probabilistic and Bayesian approach for the finite element modelling of a deep excavation in a lacustrine environment in Geneva, Switzerland. Coupling non-linear 3D finite element modeling with highly accurate meta-modelling techniques was shown to be necessary in order to run time requiring analyses like Sobol’ sensitivity and Monte-Carlo reliability analyses. A priori input parameters’ probability density functions were based on geotechnical site tests, literature and our expertise. Displacement measurements from on-site monitoring were subsequently used during the first excavation phases in order to conduct an inverse Bayesian analysis. This helped us to refine our prior knowledge for final settlements’ probability density functions and associated failure probabilities. In particular, the need to install the fourth level of bracing has been evaluated in the light of this risk analysis.