Why Data Quality Comes Before Credibility
A credibility assessment depends on experience that represents the intended risks, time period, and claim measure. A large volume of records does not make the result dependable if some exposures are missing, claims are counted twice, or records have been assigned to the wrong classification. The first task is therefore to check whether the data are complete, unique, consistently defined, and reconcilable.
These checks do not determine whether experience is mature or whether a credibility method is appropriate. They help establish whether the inputs describe the experience they are intended to describe. For example, before comparing claim frequency across groups, verify that each group’s claim count and earned exposure use the same coverage, period, and classification rules.
For clarity, let \(E\) denote earned exposure, \(N\) the number of distinct eligible claims, and \(L\) the aggregate loss measured on the selected basis. Frequency is \(N/E\), and loss per exposure is \(L/E\). If \(N>0\), average loss per claim is \(L/N\), so loss per exposure can also be written as
This identity follows by cancelling \(N\): \((N/E)(L/N)=L/E\). It is useful in an audit because a defect in exposure can distort both frequency and loss per exposure, while a defect in claim count may affect frequency and average loss per claim differently. The measures are interpretable only when their definitions and populations align. For instance, do not divide a claim count for one coverage by exposure for a different coverage.
A Practical Sequence of Checks
A useful review proceeds from the data’s definition to individual records, then to totals and comparisons. Keep the original extracts and document any adjustments: silently removing records makes a result difficult to reproduce.
Write down the coverage, experience dates, inclusion rules, exposure unit, claim measure (such as paid or incurred), and classification definitions. Confirm that the claim and exposure files refer to the same population.
Look for missing required fields, invalid dates, unexpected exposure values, and records outside the intended period. Compare exposure totals with a policy or premium system where possible.
Use stable identifiers to find repeated records. Compare classification codes with an authoritative mapping, and examine whether each record’s code agrees with its underlying attributes.
Compare claim and loss summaries with claim-system or accounting totals. Examine unusual changes across periods or groups, then trace the difference to source records before deciding whether an adjustment is justified.
A check should have a clear rule and a defined response. For example, a repeated claim identifier may be a duplicate, but it may also identify multiple legitimate transactions for one claim. Whether to count one claim or several payments depends on the chosen measure. Similarly, an unusually high loss total can reflect an actual change in experience rather than a data defect. Flag first; investigate before correcting.
Worked Examples: Finding and Measuring Data Problems
Worked Example: Missing Earned Exposure
A hypothetical small-commercial portfolio has 32 eligible claims and $96,000 of losses for a defined year. The exposure summary reports 360 earned exposure units. A policy-system reconciliation identifies 40 additional earned units that meet the same coverage and period rules. Calculate the frequency and loss per exposure using the unreconciled and reconciled exposure totals.
Method. Reconcile the denominator first, then calculate \(N/E\) and \(L/E\) on each basis. The review confirms that the 40 units were omitted from the summary, rather than outside the intended scope.
Unreconciled results. Frequency is \(32/360=0.08889\) claims per exposure unit. Loss per exposure is \(\$96{,}000/360=\$266.67\).
Reconciled results. Total exposure is \(360+40=400\) units. Frequency is \(32/400=0.08000\) claims per exposure unit. Loss per exposure is \(\$96{,}000/400=\$240.00\).
The claim count and loss total did not change; the corrected denominator is larger. Both calculated rates therefore decrease. The check avoids treating a missing exposure record as if it were evidence of worse experience. Results are rounded to five decimal places for frequency and to the nearest cent for dollars.
Check. The reconciled loss per exposure also equals frequency times average loss per claim: \(0.08000\times(\$96{,}000/32)=0.08000\times\$3{,}000=\$240.00\).
Worked Example: Duplicate Claim Records
A hypothetical liability portfolio has 600 earned exposure units. Its claim extract contains 54 rows and $162,000 of losses. A stable claim identifier review finds three repeated rows that are exact duplicate claim records, with losses totaling $6,000. The business definition counts each distinct eligible claim once. Find the corrected claim count, loss total, frequency, and loss per exposure.
Method. Remove only the confirmed duplicate records, then recompute the measures using the same exposure denominator. Do not assume that every repeated identifier is a duplicate without checking whether the rows represent distinct transactions.
The corrected claim count is \(54-3=51\). The corrected loss total is \(\$162{,}000-\$6{,}000=\$156{,}000\). Frequency is \(51/600=0.0850\) claims per exposure unit. Loss per exposure is \(\$156{,}000/600=\$260.00\).
Before correction, the corresponding values would have been \(54/600=0.0900\) and \(\$162{,}000/600=\$270.00\). The confirmed duplicates inflated both measures. As a check, corrected average loss per claim is \(\$156{,}000/51=\$3{,}058.82\) to the nearest cent, and \(0.0850\times\$3{,}058.82\) is approximately \(\$260.00\); the small difference before rounding is removed by using the unrounded average.
The common mistake is to delete all rows sharing a claim identifier. A claim may have multiple legitimate payments. Here, only the three confirmed duplicate records are removed.
Worked Example: Inconsistent Classification Codes
A hypothetical insurer compares two territories. A code-mapping audit finds that 60 exposure units, 9 claims, and $36,000 of losses belonging to Territory B were recorded under Territory A. Before correction, Territory A has 560 exposure units, 49 claims, and $156,000 of losses; Territory B has 240 exposure units, 27 claims, and $108,000 of losses. Calculate frequency and loss per exposure for each territory before and after correction.
Method. Move the identified records to their correct territory, then calculate each territory’s measures. This is a classification correction, not a change in the portfolio total.
Corrected Territory A. Exposure is \(560-60=500\), claims are \(49-9=40\), and losses are \(\$156{,}000-\$36{,}000=\$120{,}000\). Frequency is \(40/500=0.0800\), and loss per exposure is \(\$120{,}000/500=\$240.00\).
Corrected Territory B. Exposure is \(240+60=300\), claims are \(27+9=36\), and losses are \(\$108{,}000+\$36{,}000=\$144{,}000\). Frequency is \(36/300=0.1200\), and loss per exposure is \(\$144{,}000/300=\$480.00\).
For comparison, the uncorrected results are \(49/560=0.0875\) and \(\$156{,}000/560=\$278.57\) for A; for B they are \(27/240=0.1125\) and \(\$108{,}000/240=\$450.00\). The correction changes the territory comparisons even though it does not change the portfolio totals. As a check, total exposure remains \(500+300=800\), claims remain \(40+36=76\), and losses remain \(\$120{,}000+\$144{,}000=\$264{,}000\).
The mistake to avoid is treating a code as reliable merely because it is populated. Compare it with the approved mapping and the underlying territory attribute.
Worked Example: Investigating an Unusual Period Total
A hypothetical portfolio has 500 earned exposure units in each of four years. Its reported loss totals are $100,000, $102,000, $98,000, and $200,000. A reconciliation with the accounting ledger shows that the fourth-year claim extract includes an extra copy of records totaling $100,000. Calculate the reported and corrected fourth-year loss per exposure, and explain what the period comparison establishes.
Method. Use the unusual period total as a trigger for reconciliation, then calculate using the confirmed corrected amount. Do not remove losses simply because a year appears high.
Reported fourth-year loss per exposure is \(\$200{,}000/500=\$400.00\). The confirmed extra copy is removed, giving corrected loss of \(\$200{,}000-\$100{,}000=\$100{,}000\). Corrected loss per exposure is \(\$100{,}000/500=\$200.00\).
The other years’ loss per exposure values are \(\$100{,}000/500=\$200.00\), \(\$102{,}000/500=\$204.00\), and \(\$98{,}000/500=\$196.00\). The corrected fourth-year value, $200.00, is consistent with this nearby range. This comparison supports the reconciliation, but the ledger and duplicate-record review—not the apparent pattern alone—justify the correction.
The common trap is to treat an outlier as an error without checking its source. Genuine changes in the underlying experience can also produce unusual period totals.
Interpreting the Checks
A useful audit keeps separate the signal, the investigation, and the decision. A sudden change in loss per exposure may signal a problem. Comparing source systems and tracing records investigates it. Only then can the analyst determine whether to correct the data, retain a valid change, or qualify the result because an issue remains unresolved.
| Check | Possible warning sign | Follow-up |
|---|---|---|
| Exposure reconciliation | Exposure total differs from the policy or premium source | Compare scope, dates, units, and omitted records |
| Claim uniqueness | Repeated identifiers or duplicate-looking rows | Distinguish duplicate records from legitimate claim transactions |
| Classification consistency | Codes disagree with the approved mapping or underlying attributes | Trace affected records and verify the correct assignment |
| Period-total reasonableness | A sudden change in exposure, claim count, or loss | Reconcile totals and inspect record-level changes before adjusting |
Definitions matter as much as arithmetic. State whether claims are counted by distinct claim or transaction, whether losses are paid or incurred, and what exposure unit is used. Apply those definitions consistently to each group and period being compared. If a mismatch cannot be resolved, preserve it as a limitation rather than presenting a precise-looking rate as fully reliable.
Check Your Understanding
For each problem, identify the relevant check and calculate the requested measure.
- A hypothetical class has 25 eligible claims and 250 exposure units. A reconciliation finds 20 additional eligible exposure units. What is the corrected frequency? (Answer: \(25/270=0.0926\) claims per exposure unit, to four decimal places.)
- A claim file has 31 rows, of which 4 are confirmed duplicate claim records. The exposure is 300 units. What is the corrected claim frequency? (Answer: \(27/300=0.0900\) claims per exposure unit.)
- Two hypothetical classes are reported as A: 220 exposure units and $55,000 loss; B: 180 exposure units and $72,000 loss. A verified coding error moved 20 exposure units and $12,000 loss from B to A. What are the corrected loss per exposure values for A and B? (Answer: A has \( \$43{,}000/200=\$215.00\); B has \(\$84{,}000/200=\$420.00\).)
- A portfolio has 500 exposure units in each of four years, with reported losses of $100,000, $102,000, $98,000, and $200,000. A source reconciliation confirms that $100,000 of the fourth-year total is a duplicated extract. What is the corrected fourth-year loss per exposure, and what does the original outlier establish by itself? (Answer: \(\$100{,}000/500=\$200.00\) per exposure; the outlier alone signals a need to investigate but does not prove an error.)