Buying External Property Data for SME Underwriting
An MGA writes 8,000 SME property policies with $24 million GWP. The current loss ratio is 68% and the expense ratio is 29%, producing a 97% combined ratio. A vendor offers an external data feed for $450,000 a year. In a retrospective match, 6% of policies are flagged. Flagged policies have a 180% loss ratio; unflagged policies have a 60% loss ratio. The vendor claims a 2% false-positive rate among flagged policies. The MGA's largest broker controls 22% of submissions and threatens to move business if declinations rise. The CEO wants a yes or no answer this quarter.
1. Calculate the underwriting impact of using the feed to decline all flagged policies, and explain why that calculation is not a sufficient decision rule.
Read the worked answer
Flagged policies: 6% × 8,000 = 480 policies, premium 480 × $3,000 = $1.44 million, losses at 180% = $2.592 million. Unflagged: 7,520 policies, premium $22.56 million, losses at 60% = $13.536 million. Total losses are $16.128 million on $24 million premium, or 67.2%, close to the stated 68%.
If you decline all flagged policies, the remaining loss ratio is 60% and the combined ratio is 89% before the data cost. Losses avoided are $2.592 million; premium lost is $1.44 million; data cost is $0.45 million. Net improvement before fixed-expense effects is $0.702 million. But fixed expenses remain, so the expense ratio rises on a smaller premium base.
The calculation is not sufficient because the 2% false positive rate represents about 10 good policies, the largest broker controls 22% of submissions, and retrospective match is not the same as data available at quote. Blanket decline also invites adverse selection and regulatory scrutiny. The feed is better used for referral, pricing, or terms.
2. Design a 12-month pilot that tests the feed without destabilizing the broker relationship and gives a credible ROI estimate.
Read the worked answer
Randomize at the quote level within broker, with a 20% holdout, or use a stepped-wedge by broker if quote-level randomization is operationally difficult. Use the feed only to trigger referral or a rating surcharge above a defined threshold, not automatic decline. That limits broker disruption while testing value.
Track flagged rate, precision, quote-to-bind, retention, loss ratio by segment, average premium, and broker attrition. Pre-register the primary metric as incremental underwriting margin, not loss ratio alone. Include a broker communication plan and a data-use review so the largest broker sees the test as controlled.
At month 12, buy only if the lower confidence bound on incremental margin exceeds the $450,000 annual cost and broker attrition is not materially worse. Otherwise, negotiate a narrower use case or walk away.
3. What decision rule would you recommend at the end of the pilot?
Read the worked answer
Recommend a segmented rule: accept clean risks, refer or surcharge the highest-risk flagged segment, and decline only a narrow tail with confirmed high loss history. This captures most of the loss-avoidance value while preserving premium and broker relationships.
The go decision should require credible incremental margin, operational data latency under one day at quote, and no material broker attrition. If the data is only available after binding, the value case collapses, so confirm timing before renewal.