Predictions: lapse, fraud and claim cost

Actuarial > Predictions scores policies and claims with three models:

• Lapse: the chance that a policy will not be renewed or will lapse for non-payment, from its payment record, payment plan, age of the relationship, claims and channel.
• Fraud: a score for each open claim from how soon after the start it happened, how late it was reported, the amount against the sum insured, earlier claims and other signs.
• Severity: the expected final cost of an open claim.

How the models learn

Each model starts from starter weights. Once there is enough history (enough ended policies, settled or rejected claims), the weekly run trains it on your own data and the new version is used from then on. The page shows which version is in use, how many records it learned from and how well it separates the outcomes (AUC, from 0.5 for no better than chance to 1).

Retrain and score runs it on demand.

Using the scores

Each score lists the factors that pushed it up or down. The lists on this page put the highest first: policies most likely to lapse for the renewal team, claims with the highest fraud score for review.

Limits

The models are simple statistical models (logistic and linear regression), explained factor by factor. They point to where to look; they do not decide anything on their own.