Loan-level sampling helps a buyer test whether a trail-book data file is sufficiently accurate for an acquisition decision. It does not mean checking a convenient handful of loans. The buyer should first define the complete population, document every field, select both an unbiased sample and risk-targeted items, then trace each selected record to independent source evidence.

The result is an exception register, not a declaration that the entire file is correct. Unless the sampling design, sample size and evaluation support population-wide conclusions, findings should not be presented as statistical assurance. Material or patterned errors should expand the work, reduce reliance on the dataset, or change the transaction terms.

Freeze the population before selecting anything

Obtain an unedited loan-level export at a specified cut-off date and reconcile its control totals to the relevant aggregator statements. Record the row count, unique facility count, total current balance, total trail-eligible balance and total trail for the period. Preserve the original file read-only and calculate a file hash or retain other evidence that the tested version did not change.

Agree a data dictionary with the seller. It should define each field, source system, format, permissible values, calculation and cut-off rule. At minimum, consider a stable masked loan identifier, masked borrower or household identifier, lender, product, settlement date, original and current balances, trail-eligible balance, trail rate where available, gross trail, deductions, net trail, status, arrears or payment flag, and discharge or refinance date. Distinguish blanks from zeros and “not available” values.

This step matters because a sample cannot repair an incomplete population. A list omitting discharged loans, zero-trail loans or a lender channel may appear accurate row by row while still misrepresenting the book.

The RBA’s securitisation framework is not a data standard for a broker sale, but it illustrates useful data discipline: its reporting guidance requires defined loan-level fields, explicit no-data codes rather than blanks, monthly reporting and validation (RBA). Buyers can adopt that discipline without suggesting that a trail-book file is an RBA submission.

Combine random and targeted selections

Use two clearly separated components.

Random component. Select items from the frozen population using a reproducible random method and retain the seed, formula or selection record. Every eligible item in the defined population should have a chance of selection. Do not let the seller substitute easier files when evidence is missing.

Targeted component. Add items because they have heightened acquisition risk. Useful strata may include:

  • the largest balances and largest trail contributors;
  • each lender, aggregator channel and material product group;
  • recent settlements and older, heavily seasoned loans;
  • zero, negative or unusually high trail entries;
  • arrears, hardship, complaint, discharge and refinance flags;
  • duplicate identifiers, missing fields and inconsistent dates; and
  • loans affected by manual adjustments, splits or unusual commission terms.

Targeted findings describe the chosen risk group; they do not estimate the error rate of the entire book. Likewise, a purely random sample may miss a small but financially important stratum. Combining the two supports coverage and discovery, provided the results remain separately labelled.

ASA 530 defines audit sampling and discusses random, systematic and value-weighted selection, stratification, anomalies and evaluation of results (AUASB). Those principles can inform a disciplined acquisition test, but applying selected concepts does not turn buyer due diligence into an audit. A suitably qualified assurance practitioner should design any work intended to produce formal assurance or statistically projectable conclusions.

Trace each selected loan through the evidence chain

Prepare a standard test sheet before opening the files. For every selected item, record pass, exception or not testable against the same procedures:

  1. Identity and existence: match the masked identifier across the export, CRM or aggregator portal and the source statement.
  2. Key attributes: compare lender, product, settlement date, status and balances with source-system evidence as at the same date.
  3. Commission: recalculate where the operative schedule and data permit, or trace gross trail, deductions and net trail to the original statement.
  4. Cash: link the statement total containing that item to the corresponding bank receipt; do not pretend a bank deposit proves the allocation of every loan.
  5. History: follow the same loan across selected monthly snapshots to detect overwritten identifiers, unexplained balance jumps or missing periods.
  6. Client grouping: where concentration is being assessed, test whether related borrowers and facilities have been grouped consistently.

ASIC’s historical mortgage-broker remuneration review found that loan-level trail amounts varied and noted that calculations could be affected by unavailable opening and month-by-month balances (ASIC). That reinforces a practical point: a closing balance alone may not reproduce trail. The buyer needs the applicable current agreement, rate basis, timing convention and adjustment history.

Turn discrepancies into an exception register

Log every departure, including missing evidence. Capture the loan key, tested field, expected and observed values, financial effect, period affected, likely cause, owner, due date and resolution evidence. Classify the cause rather than simply marking “fixed”: extraction logic, timing, source-system error, manual override, seller explanation, or unresolved.

Look for patterns by lender, month, broker, product, data field and source system. One transposed date may be isolated. Repeated balance mismatches from one lender feed may indicate that the whole stratum is unreliable. If an item cannot be tested because a source record is absent, do not replace it and record a pass; treat the absence as evidence about record quality.

Do not extrapolate a targeted exception rate across the book. Even for a random sample, projection requires a defensible population, method and sample size, plus treatment of sampling and non-sampling risk. Ask an accountant, data specialist or assurance practitioner to determine whether projection is valid. Otherwise state exactly what was tested and quantify only known exceptions.

Control privacy while testing the data

Loan files can contain personal and sensitive commercial information. Start with aggregated and de-identified data, use stable masked identifiers, restrict access by role, prohibit uncontrolled downloads, keep an access log, and set return or destruction requirements. Obtain privacy advice for the transaction structure and the entities involved.

The OAIC says vendors and prospective purchasers must consider privacy obligations during a business sale, recommends de-identified customer information where possible, and identifies controlled review or limited data-room access as protective measures (OAIC). A masked ID is not automatically de-identified if other fields can reasonably identify the borrower. Legal review should determine the lawful basis, disclosure sequence and contractual controls.

Refresh the test before completion

A sound sample can become stale while the transaction proceeds. Near completion, obtain a new population file and bridge it to the tested cut-off: opening loans, scheduled repayments, new settlements if included, discharges, refinances, adjustments and closing loans. Reperform control totals, retest material exceptions and sample changed or newly added records. Confirm that the final settlement schedule uses the same definitions as the purchase agreement.

The buyer’s decision record should state the population tested, cut-off date, sample method, strata, procedures, limitations, exceptions, remediation and any price or condition response. Data work can test internal consistency; accounting work can reconcile commissions and cash; compliance specialists can review regulated records; and lawyers can address privacy, warranties, access and remedies. None should be silently substituted for another.

Frequently asked questions

How large should the sample be?

There is no universal number. Population size, variability, expected errors, materiality, intended reliance and the strength of source evidence all matter. If the buyer wants a statistically stated conclusion, obtain specialist sampling advice rather than selecting an arbitrary percentage.

Does a clean sample prove the book is accurate?

No. It supports only the procedures and population covered, subject to sampling risk. Reconcile full-population control totals, investigate exclusions and apply separate targeted tests to material risks.

What should happen when exceptions are found?

Identify the cause and affected population, expand testing where appropriate, quantify supported effects, and decide whether remediation, a price adjustment, warranty, holdback, condition precedent or withdrawal is appropriate with advisers.

Sources

Disclaimer

This article provides general information only. It does not constitute legal, tax, accounting, financial, credit, valuation or other professional advice. Circumstances differ. Obtain advice from appropriately qualified advisers before acting or entering a transaction.