Assessing the Performance of Random Forests for Modeling Claim Severity in Collision Car Insurance

For calculating non-life insurance premiums, actuaries traditionally rely on separate severity and frequency models using covariates to explain the claims loss exposure. In this paper, we focus claim severity. First, build two reference models, a generalized linear model additive model, relying log-normal distribution of including most significant factors. Thereby, relate continuous variables response in nonlinear way. second step, tune random forest one for log-transformed severity, where latter requires transformation predicted results. We compare prediction performance different relative error, root mean squared error goodness-of-lift statistics combination with goodness-of-fit statistics. our application, dataset Swiss collision portfolio covering exposure period from 2011 2015, observations 81 309 settled total amount CHF 184 mio. analysis, use data 2014 training 2015 testing. Our results indicate that is not leading gains forests. However, forests are favorite choice explaining right-skewed claims. Finally, when considering all indicators, conclude has best overall performance.