Giovanni Rabitti (Heriot Watt University, Edinburgh, UK)

Fantastic sensitivity measures and where to find them

Abstract. Data science and scientific modeling are extensively used to provide predictions for policy-making purposes. However, resulting communications need to be supported by a proper uncertainty quantification to assess variability in model predictions, by the identification of the key-uncertainty drivers. This information can be provided by statisticians with sensitivity analysis methods. Knowing the drivers of uncertainty supports effective policy-making. This talk is a guided tour in sensitivity analysis techniques and some recent results. In the first part of the talk, starting from the basics, local and global methods will be introduced. When the model is large (i.e. with many input variables), screening techniques are typically used to reduce the dimensionality of the scientific model and will be presented in the talk. In the second part of the talk, I will introduce cooperative games techniques to assess variable importance. I will present recent results on the computation on these importance measures, as well as their application to GAM models, to extreme value problems and to two actuarial pricing models. If there is time, I will introduce a game-theoretic interaction index for discovering subjects at higher identification risk.