Optimal personalized treatment learning models with insurance applications

摘要: In many important settings, subjects can show significant heterogeneity in response to a stimulus or “treatment". For instance, treatment that works for the overall population might be highly inefiective, even harmful, subgroup of with specific characteristics. Similarly, new may not better than an existing population, but there is likely who would benefit from it. The notion “one size fit all" becoming increasingly recognized wide variety fields, ranging economics medicine. This has drawn attention personalize choice treatment, so it optimal each individual. An personalized one maximizes probability desirable outcome. We call task learning (PTL). From statistical perspective, building PTL models imposes challenges, primarily because unknown on given training data set. this thesis, we formalize problem causal inference perspective and provide comprehensive description methods solve problem. contribute literature by proposing two novel methods, namely uplift random forests conditional forests. Our proposal outperforms based extensive numerical simulation real-world data. Next, introduce concept insurance marketing pricing applications. particular, Insurance these areas optimize client retention cross-selling experimental also illustrate application price-elasticity estimation economic price optimization context observational field, selection requires consideration expected losses individual policyholder within portfolio. non-life ratemaking gradient boosting estimate loss cost, key advantages over conventional generalized linear model approach. A facing research been lack publicly available software models. implement most fitting models, as well our proposed ones, package named uplift, which now released freely CRAN (Comprehensive R Archive Network) repository under computing environment.