An insurer's agency network looked healthy at the top line. Premium was growing, retention was broadly stable, and the annual league table ranked agencies by volume. What the table could not answer was whether the network was creating sustainable underwriting profit — or which apparent outliers were performance problems rather than reporting artefacts.
The decision
The chief underwriting officer needed to decide where to intervene: which agencies to review, which product lines to reprice or support, and whether retention programmes should be aimed at agencies with weaker loss experience. The last question contained an assumption — that better underwriting quality produces stickier customers — which we treated as a hypothesis to test, not a premise to build on.
What we assembled
We joined three views at agency, product line, state and year grain:
- Production and in-force policy counts, including new and written premium and three-year growth.
- Underwriting results, including earned premium, incurred losses, single-year loss ratio and three-year loss ratio.
- Retention and distribution context, including retention ratio, producer counts and agency tenure.
The join looked simple and was not. Prior-year premium fields carried sentinel codes; some loss ratios were negative or implausibly large; the three-year loss ratio was missing for a material share of agencies. A ranking built before those issues were resolved would have been a ranking of data quality.
How we approached it
- Set the rules before the ranking. We agreed a minimum premium volume, separate treatment for personal and commercial lines, and a rule that an agency would be judged on growth, retention and loss ratio together rather than on any single metric.
- Triage the data. Sentinel codes became nulls, not zeros. Extreme values were flagged and winsorised for comparison. Missing three-year loss ratios were reported explicitly, because an agency with no history should not be treated as an agency with perfect history.
- Use robust comparisons. Medians and percentiles replaced means where the distribution was contaminated. One large claim can move a small agency's loss ratio by an order of magnitude; that is a fact about the book, not a fact about management quality.
- Compare within segments. Agencies were compared inside product line and state, with a volume threshold that stopped tiny books from dominating either end of the table.
- Test the retention assumption. We measured the relationship between loss experience and retention directly. The data did not show the strong positive relationship the team expected, and the pattern was not stable across segments.
- Deliver a scorecard, not a spreadsheet. The dashboard opened on the network view, allowed drill-through to agency and product line, and carried a data-quality panel so the business could see which records were excluded and why.
What this changes
Three things changed in the way the network was reviewed:
- The league table became volume-adjusted and segmented. Small agencies with one bad claim no longer occupied the bottom of the list, and personal and commercial lines were no longer compared as if they had the same economics.
- Genuine outliers were separated from reporting artefacts. The underwriting review list became shorter and more credible, which made it easier for the field team to accept.
- Retention stopped being justified by underwriting quality. Because the assumed link was not supported, the retention team moved to direct drivers of customer behaviour instead of using loss ratio as a proxy.
What you receive
- An agency scorecard with volume-adjusted growth, retention and loss-ratio bands.
- Segment comparisons at network, state and product-line level.
- A data-quality register documenting sentinel handling, missingness and exclusion rules.
- A dashboard with drill-through and a short methodology note for internal review.
Where this applies
The same approach fits insurers, MGAs and distribution networks where agency-level production and claims data are reported separately. The minimum useful inputs are agency and product identifiers, written and earned premium, incurred losses, policy counts and retention over at least two comparable periods. The first deliverable is usually not a model; it is a shared definition of what "good" means and a clean comparison.
This is an illustrative case study built on an anonymized dataset; it does not report results from a specific client engagement. It reflects our standard method for distribution and underwriting performance analysis.
