Privacy and preparation
Insurance loss run data quality: the context a training set needs
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
Loss runs become useful training data only with the context that explains each snapshot. Record the valuation date, policy period and coverage trigger, the policy key and the submission or renewal the run informed, whether reserves are shown, which years are missing, and where the document came from. Keep conflicting runs and label why they differ instead of overwriting them.
Checklist
Insurance loss run data
Key takeaways
- A loss run is a snapshot as of its valuation date; without that date, development looks like error.
- Paid-only runs understate open claims and must be marked as lower bounds.
- A missing year is unknown until the carrier confirms no claims in writing.
- Link every run to a policy key and to the underwriting decision it informed, or label it extraction-only.
- Keep disagreeing runs side by side with a reason label rather than replacing one with the other.
What does a loss run actually record, and what does it leave out?
A loss run is a carrier's snapshot of claims on a policy as of one date, so on its own it tells a model what was paid and reserved at that moment, not how the account actually performed. Before it becomes training data, it needs the context that explains the snapshot: when it was valued, which policy and term it covers, who produced it, and what it omits.
Florida's loss run statute gives a useful floor for what the document must contain. Section 627.444, as published in the 2022 Florida Statutes, defines a loss run statement as a report with the policy number, the period of coverage, the number of claims, the paid losses on all claims and the date of each loss, and says the term does not include supporting claim file documentation. The same section says an insurer is not required to provide loss reserve information.
That last point matters for AI work. A loss run that shows paid losses without reserves understates open claims, and a model trained to predict loss experience from it will learn a systematically optimistic picture of young policy years. The dataset has to record which runs include reserves and which do not, instead of assuming every report carries incurred values.
Which dates decide what a loss run means?
Four dates control the meaning of every loss run row: the policy period, the date of loss, the date the claim was reported and the valuation date of the report. If any of the four is missing, a model cannot tell whether two numbers differ because the claim developed or because the reports were cut at different times.
Actuarial practice defines these precisely. The Casualty Actuarial Society's reserving text by Friedland describes the accident date as the date of the event that triggered potential coverage, the report date as when the insurer learned of the claim, and the valuation date as the date through which transactions are included, regardless of when the analysis is done. It also notes that for claims-made coverage the report date can stand in for the accident date, which is why the coverage trigger belongs in the record.
Development is the reason valuation dates matter. NCCI's unit statistical reporting for workers compensation values a policy's first report 18 months after the policy effective date and each later report 12 months after the previous one, so the same policy legitimately shows different losses at each valuation. A training example that pairs an 18-month value with a 42-month outcome without saying so teaches the wrong lesson.
What should the source quality and linkage audit worksheet include?
The worksheet checks each loss run for completeness, dating, linkage and lineage before it is accepted, and records a flag rather than silently dropping a report that fails. Use one row per loss run document; the decision column is what an audit reviewer fills in.
| Check | Input | Decision rule | If it fails |
|---|---|---|---|
| Valuation date present | Printed as-of date on the report | Every report has one valuation date | Flag undated; exclude from development analysis |
| Policy period and term | Effective and expiration dates | Every claim's loss date falls inside the stated period, or the coverage is claims-made and report date is used | Flag out-of-period rows for review |
| Issuer identified | Carrier, MGA or broker summary | Carrier-issued runs ranked above broker-compiled summaries | Keep but mark as secondary source |
| Reserves shown | Reserve or incurred columns | Record whether reserves are included | Treat paid-only runs as lower bounds |
| Years covered | List of policy terms in the file | Requested look-back is complete or each gap is explained | Record missing term; never fill with zero |
| No-loss confirmation | Carrier statement of no claims | A term with zero claims has written confirmation | Mark as unknown, not as loss-free |
| Totals reconcile | Claim rows and printed totals | Sum of rows equals printed totals per term | Flag extraction or report error |
| Policy link | Policy number and effective date | Matches a policy record in the source system | Hold out of linked analyses |
| Decision link | Submission or renewal ID | Report is attached to the underwriting decision it informed | Usable for extraction only, not for decision modeling |
| Lineage | File origin and receipt date | Original document and receipt date retained | Flag as unverifiable copy |
How do you link a loss run to the policy and decision it informed?
Link each loss run to a policy key and to the submission or renewal that used it; without the second link, the dataset can teach extraction but cannot teach underwriting judgment. The policy key needs at least the carrier, the policy number and the policy effective date, because policy numbers repeat across carriers and sometimes across renewal terms.
At a glance
How do you link a loss run to the policy and decision it informed
Statistical reporting uses the same idea. NCCI's unit statistical guidance treats carrier code, policy number and policy effective date as key fields, and accepts corrections to them only on first reports. A dataset that stores the same three fields can join a loss run to its policy without guessing, and can show when a later report quietly changed a key.
The decision link is where operational archives often fall short. Brokers and MGAs often keep loss runs in a document folder separate from the submission record, so the run that an underwriter actually saw is not recorded. Ask three questions of each file: which submission or renewal requested this report, which version did the underwriter review before quoting, and what was the outcome. When those answers exist, the record supports a model that learns how loss history changes terms; when they do not, label the record extraction-only.
- Policy key: carrier identifier, policy number, policy effective date, line of business.
- Report key: issuer, valuation date, receipt date, document ID.
- Decision key: submission or renewal ID, quote version reviewed, final disposition such as quoted, declined or referred.
- Claim key: carrier claim number pseudonymized consistently across every valuation of the same claim.
Two loss runs for the same policy disagree: which one do you keep?
Keep both and label the reason they disagree; many disagreements are legitimate development, not errors. A later valuation should replace an earlier one only for the question of current incurred losses, while both stay in the record for any task that learns how claims develop.
Overwriting is an easy way for loss history to get corrupted on its way into a dataset. A records team receives a fresh run at renewal, replaces the file in the document system, and the earlier valuation that the underwriter actually relied on disappears. If the dataset is later used to learn how underwriters weigh loss history, the model sees a number the underwriter never saw.
Claim identity across reports needs the same care. Carriers sometimes renumber claims after a system migration or split one occurrence into several claimant-level claims, so a claim that vanishes from a later report may still be there under a new number. Match on loss date, cause and amounts before labeling a claim as removed, and record the match method.
Use a short decision rule for the common cases:
| Situation | Likely cause | Keep | Label |
|---|---|---|---|
| Same policy, different valuation dates, higher incurred later | Normal reserve development | Both, ordered by valuation | development |
| Claim on earlier run missing from later run | Claim closed without payment, merged or moved | Both; mark the claim as dropped | claim_removed_unexplained until confirmed |
| Carrier run and broker summary differ | Manual transcription or different cut-off | Carrier run as primary | secondary_source_conflict |
| Closed claim reappears with new reserve | Reopened claim | Both; link by claim key | reopened |
| Totals do not match rows on one report | Report or extraction error | Report, flagged | reconciliation_failure |
Illustrative audit: five years of commercial auto loss runs
This scenario is illustrative and invented to show how the worksheet is applied; no real insured, carrier or licensed dataset is described. A broker file for a commercial auto renewal contains loss runs requested for five prior policy terms from two carriers, plus a one-page summary the broker typed for the underwriter.
The audit takes the file as it arrived, without asking the broker to tidy it first, because the mess is part of what the dataset must represent faithfully.
Applying the worksheet produces these findings, all counts illustrative:
- Six documents in total: five carrier runs and the broker summary. The five runs cover four of the five terms, because the most recent term has runs at two valuation dates. Four runs have valuation dates; one has none and is excluded from development analysis.
- One policy term has no run and no no-loss letter. It is recorded as missing, and the summary's statement of no losses for that year is labeled unverified.
- One carrier run shows paid losses only. Its open claims are treated as lower bounds, and the account's loss ratio for that term is marked incomplete rather than calculated.
- The broker summary's total for the most recent term differs from the carrier run, which carries a later valuation date. The carrier run is primary; the difference is labeled secondary_source_conflict, not corrected.
- The renewal submission links to the summary, not the carrier runs, so the decision record shows the underwriter reviewed the summary. That fact is kept, because a model learning underwriting decisions should see what the underwriter actually saw.
- Result: the file is accepted for extraction training with four flags, and accepted for decision modeling only with the missing term and the paid-only term marked as uncertain inputs.
Reserves, missing years and young valuations: which flags should each record carry?
Three assumptions can change the decision: whether reserves are included, whether a missing year means no losses, and whether the valuation is mature enough to compare. Each should be written down per record as a flag with a reason, not resolved by a default value.
Show uncertainty as structure, not as a single score. A record can carry an input-completeness tier (complete, partial, minimal), a list of the flags that triggered it, and the valuation age in months for every term. That lets a buyer filter for mature, complete records for evaluation while still using partial ones for extraction.
Document these choices where a dataset consumer will read them. The Datasheets for Datasets proposal by Gebru and colleagues asks dataset creators to describe composition, including whether instances are missing information, and the collection process. For loss runs, that section is where the valuation policy, the treatment of paid-only reports and the missing-year rule belong.
Stop and ask for specialist review when a file mixes occurrence and claims-made terms, when reserves were restated after a carrier change, or when the account was transferred between programs mid-term. Those cases need an underwriter or actuary to define the right comparison before any label is applied.
Claimant details and injury descriptions inside loss runs
Loss runs often name claimants and describe injuries in plain language, so they need the same de-identification care as claim files even though they look like summary reports. Workers compensation and bodily injury runs are the most sensitive.
Whether federal health privacy law applies depends on who holds the record. HHS guidance on workers compensation says the HIPAA Privacy Rule does not apply to workers compensation insurers, administrative agencies or employers, except to the extent they are otherwise covered entities. That does not make the data free to share. The NAIC Insurance Data Security Model Law treats information concerning a consumer, and health-related information created by or derived from a provider or consumer, as nonpublic information, and many states have adopted versions of it.
Practically: replace claimant names and claim numbers with pseudonyms that stay stable across every valuation of the same claim, generalize injury descriptions to a cause and body-part category when the task does not need the text, and keep the loss date, report date and amounts that make the run useful. Counsel reviews the specific records and the states involved.
Whose loss runs are they to license?
A loss run held by a broker or an insured was produced by the carrier, describes the insured's claims and may name third parties, so three parties have a stake before anyone licenses it. Holding a copy in your document system is access, not a licensing right; the agency agreement, the client relationship and any carrier terms decide what reuse is allowed.
Where those rights are clear, loss runs linked to policies and decisions can become a licensed dataset, and SourceX supports that work across sourcing and qualification, rights review, contracting, data preparation, delivery and payment. The supplying company approves the scope and terms and signs before delivery. As a general guide, buyers look for at least two years of consistent history and tens of thousands of records in one system, though smaller sets of rare work can qualify and each case is assessed on its own.
Frequently asked questions
Do loss runs include reserves?
Not always. Florida's statute, for example, says an insurer is not required to provide loss reserve information in a loss run statement. Record whether each run includes reserves and treat paid-only runs as lower bounds for open claims.
Why do two loss runs for the same policy show different numbers?
Usually because they were valued on different dates and the claims developed in between. Keep both with their valuation dates, and only treat the difference as an error when the reports share a valuation date or the totals do not reconcile.
Is a missing year the same as a year with no losses?
No. A term with no report is unknown until the carrier confirms in writing that there were no claims. Labeling a missing year as loss-free makes the account look better than the evidence supports.
Which keys link a loss run to its policy?
At minimum the carrier, the policy number and the policy effective date, plus the line of business. NCCI's statistical reporting treats carrier code, policy number and policy effective date as key fields for the same reason.
Are workers compensation loss runs covered by HIPAA?
HHS says the Privacy Rule does not apply to workers compensation insurers, agencies or employers unless they are otherwise covered entities. The records still contain sensitive health-related information, and state insurance privacy and data security laws can apply, so counsel should review them before any reuse.
Sources
- Florida section 627.444 defines a loss run statement by policy number, coverage period, claim count, paid losses and loss dates, excludes claim file documentation and does not require reserve information. Source
- The CAS reserving text by Friedland defines accident date, report date and valuation date, and notes report date can serve as the accident date for claims-made coverage. Source
- NCCI values first unit statistical reports 18 months after policy effective date and later reports at 12-month intervals, with carrier code, policy number and effective date as key fields. Source
- HHS says the HIPAA Privacy Rule does not apply to workers compensation insurers, agencies or employers except where they are otherwise covered entities. Source
- The NAIC Insurance Data Security Model Law defines nonpublic information to include consumer information and health-related information from providers or consumers. Source
- Datasheets for Datasets proposes documenting a dataset's motivation, composition, collection process and recommended uses. Source
See if your company qualifies
A short company assessment. No data uploads are needed.