Decision tree
A model that predicts by asking a sequence of binary questions about the variables, readable by a human, the basis of ensemble methods and a natural tool for underwriting rules.
A decision tree predicts a value or a class by successively splitting the data according to simple questions: is the vehicle more than ten years old, is the driver under twenty-five, does the claim amount exceed a threshold. Each split is chosen to make the resulting groups as homogeneous as possible with respect to the variable to predict, and the process stops according to size or minimum gain rules. The best-known method, CART for classification and regression trees, was formalized in 1984. The tree has a rare quality: it is readable. Each leaf corresponds to an explicit rule that an underwriter, handler or regulator can understand and discuss. It naturally captures interactions between variables, without their having to be specified. Its weakness is instability: small variations in the data can profoundly change its structure, and a deep tree learns the noise in the sample. That is why ensemble methods, random forest and gradient boosting, combine hundreds of them. In insurance, the single tree remains the tool of choice for formalizing underwriting or claims triage rules that must be explainable.
Textbook case: a four-level tree trained on liability claims isolates a leaf grouping construction firms with fewer than ten employees that have already had a claim above 50,000 euros: their frequency is three times the average. The readable rule is added to the underwriting guide as a referral criterion.
Follow the research
One letter a week. First, the week's underwriting note, when one is published: a dated, argued decision, signed by Paul Bernardski. Then what the observatory published on cyber risk, political risk and insurance. No account, no tracking.
No spam, one-click unsubscribe, no data ever sold.