When a segment or geography starts producing adverse loss experience, the instinct is to act at the portfolio level — tighten criteria, increase rates, or non-renew broadly. The logic is sound: the segment is underperforming, and action is required. The problem is that poor aggregate performance in a segment is almost always produced by a minority of risks within it. The other risks in that segment are performing adequately or well. A blanket action applies equally to both.
The carrier that non-renews a postcode after a period of elevated weather losses removes the risks that were driving the elevation and the risks that were not. The competitors that have more granular models take on the latter, at standard rates, and the non-renewing carrier has shed profitable business while its remaining book is now composed of a worse-than-average selection from the segment it kept. The portfolio action designed to improve book quality can deteriorate it.
Where the non-renewal decision breaks down
The data to distinguish performing from non-performing risks within a segment already exists in the carrier’s own claims and policy records. Loss history at the individual policy level, property characteristics, prior claim frequency, credit-based insurance score, and inspection data all contribute to a risk-level view that the segment average obscures. A model that scores each expiring policy against its own risk profile — not the segment average — identifies which policies are contributing to the adverse experience and which are incidental to it.
The distinction matters because the adverse experience in most segments is concentrated. A relatively small proportion of policies drives the majority of the loss, and those policies have identifiable characteristics that distinguish them from the profitable majority. Acting on those characteristics at the individual policy level produces a better outcome than acting on the segment label.
The billing and collections process creates a separate non-renewal trigger that is also applied bluntly in most carriers. A missed premium payment generates an automatic lapse action regardless of the policyholder’s claim history or long-term value. A risk-stratified collections approach holds lapse action on high-value, low-loss policyholders who miss a payment while applying it promptly to high-risk policyholders doing the same.
The regulatory dimension
Non-renewal decisions are subject to oversight by state insurance departments in most US markets. A carrier that cannot produce an individual risk-based justification for a non-renewal is vulnerable to a finding that the decision was arbitrary or discriminatory. Blanket geographic or segment-level non-renewals are harder to defend than individually justified decisions even when the underlying data supports the action at the segment level.
A risk-scoring model that produces an individual risk justification for each non-renewal decision creates both better outcomes and a more defensible audit trail. The regulatory obligation and the commercial objective align: the model that identifies which individual policies warrant non-renewal based on their own risk profile is the same model that produces the documentation required to defend those decisions.
The technology dimension
Non-renewal scoring draws on the full policy and claims history held in the policy administration system. For carriers running their policy administration on IBM Z, deploying the non-renewal scoring model via IBM Machine Learning for z/OS accesses the complete risk profile of each expiring policy within the renewal processing workflow, returning a risk score and the key contributing factors before the renewal offer is generated. The decision is made at the individual policy level, with the data that supports it, without requiring a separate analytical process outside the production environment.
What success looks like
The metrics are retention rate by loss ratio decile, renewal book combined ratio, and regulatory finding rate on non-renewal decisions. The retention rate by loss ratio decile is the most direct measure: a programme working correctly retains high proportions of the low-loss deciles and low proportions of the high-loss deciles. Blanket approaches produce similar retention rates across deciles. The gap between those two profiles, measured over two or three renewal cycles, captures the book quality improvement that targeted non-renewal produces.
