Inspiration

Millennium Mutual's leaders described one problem from different seats: the business coming in isn't turning into profit, and customers are waiting too long on their claims. We wanted to find exactly where the money goes, then build something a claims team could act on the same day, with a person still making the final call.

What it does

Millennium Signal runs DATA → DETECT → DECIDE → ACT on Millennium Mutual's policy and claims data (4,680 policies, 847 claims).

What we found

  • One broker explains most of the problem. The overall loss ratio is 64%. Take out broker B07 and it falls to 56%. B07 sold 173 policies and has a 429% loss ratio, a 40% claim rate (typical broker: 17%), and 90% of its claims arrive in the first 60 days of the policy (typical: 28%). Many land just under the $10,000 line where a senior reviewer would step in.
  • Slow claims cost renewals. Customers who waited more than 30 days for their claim renew at 45%, compared with 83% for a 15–30 day wait and 85% for customers with no claim. Team C's median time to close roughly doubled (about 15 to 32 days), and it holds 38 of the 48 open claims.

What it detects (rules we wrote ourselves; no AI model decides anything)

  1. Aging Claim Alert: an Open or Under Review claim more than 30 days past its reported date. That is 15 claims today. The slider shows 34 at 7 days and 8 at 60.
  2. Near-threshold early claim: a claim made in the first 60 days of a policy that sits within 25% below the $10,000 senior-review threshold.
  3. Broker scorecard: a broker is flagged when two of three metrics (loss ratio, claim rate, early-claim share) are at least 2× the median broker, with at least 10 claims. No broker is named in the code; B07 is the only one that trips it.

What it does about it

Each flag becomes a task in a review queue with a plain-English reason and a drafted alert to the right owner (Senior Claims Review, Head of Distribution, VP of Claims, Retention). A person confirms a hold or releases the claim, and every action and decision is written to a log. Messages are drafted, not sent. A flag is a request for review, not a fraud verdict.

Sliders on every rule let a manager ask "what if we change the window?" and see how many claims would be flagged. A new-claim form shows a claim being routed live.

How we built it

  • Cleaning in Python and pandas. We removed 15 exact duplicate claims and 4 test policies, standardized 16 spellings of 5 cities, converted premiums stored as text, parsed mixed date formats, and marked impossible values (3 ages under 16, 6 claims reported before the loss, 8 missing amounts) instead of silently dropping them.
  • Earned premium, not written premium. Loss ratio = claims paid ÷ premium earned up to the October 31, 2026 extract date, so policies sold at different times compare fairly.
  • Rules over a black box. Every flag carries a reason a claims manager can read, and every threshold can be tuned. It runs locally.

Challenges

  • Choosing the 60-day window with evidence. We tested 30 to 120 days. A 60-day window finds all 58 B07 claims that a 120-day window does, while flagging 15 claims elsewhere instead of 27.
  • Not punishing good brokers. Only one broker trips the scorecard, so the response is targeted instead of slowing everyone down.
  • Messy data, one afternoon. Cleaning decisions change the answer, so we logged every one of them in the app.

What we learned

Higher sales are not the same as higher profit. The problem was hidden by the portfolio average and sat in one sales channel and one claims team.

What's next

Send the drafted alerts by real email or text, connect to the live claims system, back-test the thresholds on older data, and add an AI summary of each flagged claim to help the reviewer (the decision stays human).

Built With

Share this project:

Updates

Submission history