Thesis title: Improving mortality diagnostics and estimation through Contrast Trees
Contrast Trees are an iterative input space partition technique introduced by Friedman (2020) in order to automatically uncover regions in the input space itself where two target variables differ the most.
In case inaccuracies are detected, Boosted Contrast Trees can be used in order to reduce differences. The Distribution Contrast Boosting provides an assumption-free method of estimating the full probability distribution of an outcome variable on the same input space.
By applying for the first time such techniques in the context of mortality modeling, the aim of this work is threefold. Firstly, to test if Contrast Tree based diagnostic can be applied to mortality description and projection models, and if results obtained from this new technique are consistent with those give by some traditional indicators. Secondly, to utilize Estimation Contrast Boosting techniques, for building mortality tables for small populations.
Thirdly, to generalize the Italian actuarial practice of reproportioning mortality rates using both Estimation Contrast Boosting and Distribution Contrast Boosting.