The problem

A leading US auto insurer periodically reviews policies at the time of renewal and inception. Over 15% of policies are subject to review every year, and from a business efficiency point of view, the insurer wants to reduce the number of referred policies. The team was challenged with two problems viz.

1. How can you identify the optimized premium on the policy?

1. How can you reduce the number of policies up for review without resulting in any underwriter action?

The cognitive solution

To answer the first question, a rule engine would have to be devised, so as to correctly understand which policies to review. For auto-renewals specific information on the insured age, type of plan in-force, the claim severity and claim amount were very critical to devise the rules from a business perspective. However, a generic rule would yield multiple policies for review. We employed the below design to solve the question in the study:

1. Identify top predictors that affect a premium loading

2. Build an ensemble tree model to proxy as a rule engine

3. Using the tree-based model, classify the policies as less likely to be premium loaded vs. more likely to be premium loaded by underwriters

Historically underwriters over-adjusted the premiums for the ones up for review. Amongst the policies that were not referred, missing information on the tenure and previous insurer resulted in a higher loss ratio. These were then optimized with a higher weight in the tree model for prioritization for review.

Key insights

After employing a tree-based classifier, the overall number of yearly reviews dropped by a staggering 40%, resulting in reduced workload for the underwriters. Applying recommended loading to policies with a high likelihood of receiving an underwriter premium load helped in reducing over-adjustment of referred policies and achieving the desired target loss ratio.