Actuarial Data Quality and ASOP 23

Every actuarial estimate is only as good as the data beneath it. ASOP No. 23 does not require the actuary to audit data like an accountant, but it does require a documented, risk-based review appropriate to the intended use — and in practice this means reconciliations, completeness checks, and anomaly screens performed before methods are applied, not after results look strange.

What ASOP 23 actually requires

ASOP 23 applies whenever an actuary uses data (the actuary's own or supplied by others) to perform actuarial services. It requires the actuary to:

1. Take reasonable steps to review the data for reasonableness and consistency, given the intended use, though the actuary is not required to audit the data.
2. Disclose the extent of review performed, including reliance on data supplied by others (per ASOP 41's disclosure norms as well).
3. Disclose any known material limitations, defects, or biases discovered in the data that were not corrected before use.
4. Consider whether appropriate alternative data exists if the data provided is insufficient for the assigned task, and document that consideration.
5. Make a reasonable effort to determine data is appropriate, including confirming with the data provider the definitions used (e.g., is "paid loss" gross or net of salvage/subrogation, is a claim count per occurrence or per claimant).

Critically, ASOP 23 shifts responsibility: if the actuary relies on data supplied by another party without independent verification, the actuary must still disclose that reliance and any resulting limitation on the conclusions — reliance does not eliminate the disclosure obligation.

The three pillars: completeness, accuracy, timeliness

A practical data-quality review organizes checks around three dimensions:

• Completeness — are all expected policies/claims/transactions present? Compare record counts and control totals against an independent source (general ledger, policy admin system extract counts, prior period's data plus known changes).
• Accuracy — do individual fields make sense (no negative earned premium, dates in logical order, claim status codes valid, no duplicate claim numbers)? Field-level validation rules and outlier/anomaly screens live here.
• Timeliness — is the data current enough for the "as of" date being analyzed, and consistent in cutoff across sources (e.g., claims data as of 12/31 vs. premium data as of 12/31, not 12/25)?

Control totals and reconciliation to the ledger

The single most common and most effective ASOP 23 check is a reconciliation of the actuarial extract to an independent financial control total, typically the general ledger or statutory annual statement exhibit. The mechanics:

$\text{Break} = \text{Source total (ledger)} - \text{Actuarial extract total (claims system)}.$

A break is not automatically disqualifying — it must be explained (timing cutoff differences, LAE classification differences, currency/rounding, ceded vs. gross scope) and judged against a materiality threshold:

$\left|\frac{\text{Break}}{\text{Source total}}\right| < \text{tolerance (e.g., 0.5 percent)}.$

Unexplained breaks above tolerance must be either resolved before the data is used, or disclosed as a known limitation if the actuary proceeds anyway (e.g., under a deadline, with client sign-off).

Worked reconciliation example

An actuary receives a claims extract to support a year-end reserve review, with a general-ledger paid-loss control total of 84,236,905 as of 12/31/2023.

Unexplained residual as a fraction of the control total is $1{,}445 / 84{,}236{,}905 = 0.0017\%$, well inside a typical 0.25%–0.5% tolerance — the actuary can proceed and document the reconciliation, including the three explained items, in the workpapers.

Anomaly detection

Beyond reconciliation, a data-quality pass should screen for outliers using simple, transparent tools rather than opaque black boxes, since results must be explainable to reviewers:

$z_i = \frac{x_i - \bar x}{s},$

flagging individual claim payments, loss ratios by segment, or triangle diagonal-to-diagonal ratios with $|z_i|$ above a chosen threshold (commonly 3) for manual review — not automatic exclusion. Development-triangle-specific anomaly checks include large negative incremental values, sudden claim count decreases (suggesting void/reopen coding issues), and a diagonal (calendar-period) ratio that jumps out of line with adjacent diagonals, which often signals a systems conversion or claims-practice change rather than a real trend.

Documentation

At minimum, the workpapers should retain: the source and vintage of each data file, the reconciliation performed and its results (including unresolved breaks and their disposition), completeness/timeliness checks performed, any data corrections made and by whom, and a statement of reliance where data was supplied by others and not independently verified.

Pitfalls

• Reconciling to a different control total than the one relevant to the analysis (e.g., calendar-year paid vs. accident-year paid).
• Treating "the data reconciles" as equivalent to "the data is fit for purpose" — reconciliation checks totals, not definitions (e.g., ALAE classification consistency over time).
• Silently correcting data without documenting what was changed and why, breaking the audit trail for a future reviewer or successor actuary.
• Applying anomaly screens only to the current period, missing gradual drift that a single period's z-score would not flag.

Exam relevance

ASOP 23 and data-quality practices are core professionalism content on CAS Exam 6 and appear in the case-study components of CAS Exams 7 and 9.

Further reading

• ASOP No. 23, Data Quality (ASB)
• American Academy of Actuaries, Practice Note on Data Quality Considerations
• CAS Statement of Principles Regarding P&C Loss and LAE Reserves (data section)