When it comes to insurance modelling, there is plenty of material and training on how to build statistical models. We can use these resources to learn about generalised linear models and gradient boosting machines (see feature, overleaf), understanding their advantages and weak points. The same applies to different transformations and techniques, such as splines, variables mapping, geographical classification, finding significant interactions and mitigating adverse selection. The statistics background and modelling best practices are similar across various industries, so a general data science approach is usually good enough for entry-level actuaries – especially given that, in insurance pricing, we usually use commonly known distributions such as Tweedie, Poisson or gamma. But there is one insurance-specific area in predictive modelling: how to structure our actuarial analyses in the first place. Pricing actuaries tend to put a lot of effort into building the most accurate statistical models and optimising their Gini scores, root mean squared error and/or Akaike information criterion, but it’s equally (if not more) important to understand the product and structure risk modelling in the first place. So, how should we split our risk models?
Authors
Additional information
- Category
- Publikacja w czasopiśmie
- Type
- artykuły w czasopismach
- Language
- angielski
- Publication year
- 2024
