A travel insurance distributor can report net sales, commission and claims and still not know which products deserve more attention. The three measures tell different stories: a product can sell well and pay little commission, or generate commission while producing claims that erase the margin. Channel mix adds another layer, because the same product can be sold through agencies with very different data quality.
The decision
The commercial team needed to decide where to focus distribution and product development: which products and destinations to grow, which channels to investigate, and where claim activity was high enough to warrant a pricing or underwriting conversation. The review also had to establish which comparisons were supportable. A claim rate on a small product is not a reliable signal, and a missing customer attribute is not neutral when it is concentrated in one channel.
What we worked with
Policy-level records covering roughly 63,000 travel insurance sales:
- Distribution: agency, agency type and online versus offline channel.
- Product: plan name, trip duration and destination.
- Economics: net sales and commission in value.
- Outcome: whether a claim was recorded.
- Customer: age and gender, where recorded.
The traps we designed around
- Grain before totals. The extract contained thousands of fully duplicated rows. Before treating them as errors, we checked whether they represented separate policies with identical attributes. Only after confirming the grain did we deduplicate — otherwise the claim rate and premium totals would both be wrong.
- Net sales is not premium. The field includes negative values and zeros. A negative "sale" may be a refund, an adjustment or a claim-related entry. We confirmed the sign convention and separated the measures instead of summing everything into one revenue figure.
- Claim rate is a rare event. Around one and a half percent of policies record a claim. At that frequency, a product with a handful of claims can move to the top or bottom of a ranking by chance. We applied minimum-volume rules and reported intervals rather than point estimates alone.
- Missing gender is not random. Gender is missing for most records, but the pattern is concentrated: it is almost entirely absent for travel-agency sales and largely present for airline sales. Analysing "gender" would mostly be analysing agency type. We treated the field as unavailable for portfolio segmentation and documented why.
- Duration is exposure. Claimed policies have much longer average trip durations. That is expected: a longer trip creates more opportunity to claim. Comparing claim counts without normalising for exposure would overstate the risk of long-duration products.
- Product names are not a taxonomy. Dozens of plan names overlap in cover and target customer. Grouping them requires a documented mapping, not string matching on "Gold" or "Comprehensive".
How the review was built
- Reconcile the grain. We compared row counts with policy counts, identified duplicate rows and confirmed which fields made a record unique.
- Define the measures. Net sales, commission and claim indicator were documented separately, including sign conventions and the treatment of zero and negative values.
- Build a product hierarchy. Plan names were mapped to a smaller set of cover types, with the original label retained for traceability.
- Compare on exposure and volume. Claim frequency was reviewed alongside trip duration, policy count and the number of claims, with a minimum-volume threshold for any ranking.
- Audit missingness by source. Each field's missing rate was checked by agency type and channel. Fields that were effectively channel-specific were excluded from cross-segment comparisons.
- Present ranges, not league tables. Product and channel views showed uncertainty and sample size next to the estimate, so a small product could not masquerade as a top performer.
What this changes
- "Best product" becomes "best under which measure". Commission, net sales and claims produce different rankings, and the team now chooses the measure that matches the decision.
- Channel data quality becomes part of the decision. A channel with systematically missing customer attributes cannot support the same analysis as one with complete records.
- Small products are treated as uncertain, not promising. Minimum-volume rules stop a handful of claims from driving a portfolio recommendation.
- Duration is built into risk conversations. Claim comparisons are framed per unit of exposure, not as raw counts.
What you receive
- A documented data grain and de-duplication decision.
- Separate premium, commission and claims views with sign conventions.
- Product and channel performance with volume and uncertainty shown.
- A missingness audit by distribution source.
- A short list of questions to resolve with pricing and underwriting before any repricing decision.
Where this applies
The method fits insurers, brokers and travel distributors with policy-level sales and claims data. The minimum inputs are a policy identifier, product and channel, premium or net sales, commission, a claim indicator and trip duration. The first deliverable is usually a grain and measure-definition note, because most portfolio disagreements start there.
This is an illustrative case study using an anonymized public dataset; it does not report results from a specific client engagement. It reflects our standard method for insurance portfolio analysis.
