A sample of loan files is not a miniature bank unless it has been designed and validated for that purpose. It may aim to identify higher-risk cases, understand a process or estimate a population rate. Each objective calls for a different selection and supports different conclusions. I would therefore write the question before selecting records. A list of the largest loans is valuable for one perspective but does not automatically explain controls on smaller transactions or the frequency of a problem across the institution. A clear sampling rationale is part of the evidence, not administrative background.
A documented large-scale review
On 26 October 2014 BaFin and the Bundesbank published German results from the European comprehensive assessment. The official account records more than 18,000 loan files and 15,000 collateral items reviewed in Germany, with risk-based portfolio selection. This was a specific historical exercise with its own methodology. Its scale does not establish a universal sample size or prove that an unrelated selection is representative.
The case prompts me to distinguish selection of portfolios from selection of individual files. Even a large sample may miss a material weakness if a relevant group is excluded. A small targeted sample can reveal a genuine process defect without allowing estimation of its prevalence across the bank. Both are useful, but they need accurate descriptions. My conclusion should explain whom the sample included, whom it excluded and what question it was designed to answer. That gives readers a basis for judging the evidence rather than relying on a large headline number.
Validate the population first
Before selection, I would establish the complete eligible population, reference date and inclusion rules. Are we counting contracts, borrowers, borrower groups or drawings? Does a closed loan remain in the review period? How are restructured cases handled? A technically random selection can still be misleading if its starting population is incomplete. I would reconcile counts and values against independent available summaries and explain differences. Sampling cannot repair an error introduced while constructing the population, so defining the correct frame deserves as much care as the later selection mechanism.
I would also check duplicates and broken relationships in the export. A contract with several collateral items may appear repeatedly after joining tables. Randomly selecting rows then unintentionally favours contracts with more collateral records. Conversely, unsecured loans may disappear from an inappropriate join altogether. No deliberate manipulation is required for this technical defect to change the result. I would verify the extraction logic using identifiable examples and retain enough information to reproduce it. The relevant unit of selection must survive the data preparation, rather than merely be described correctly in the sampling note.
A hypothetical sample with three purposes
Imagine a population of 10,000 contracts. I want to understand affordability assessment, exceptions and large exposures. A combined selection might include thirty ordinary contracts, twenty exception cases and ten largest exposures. That produces sixty selected positions but is not a simple random sample. Each component answers a separate question. If the same contract appears in two components, I would retain the overlap information and report the number of unique files correctly. Otherwise a summary can exaggerate coverage before any substantive review has even begun, simply by adding overlapping lists together.
Suppose six of the twenty targeted exceptions lack traceable justification. I can report six cases in that component, not assert that thirty per cent of all bank contracts share the defect. Estimating prevalence requires an appropriate selection design, inclusion probabilities and uncertainty assessment. For a targeted review, the next step might instead be extension to similar cases or a population-level data test if the requirement can be expressed reliably. The choice depends on whether we need to understand a cause, establish breadth or estimate frequency; those are related but distinct analytical objectives.
How I would document selection and review
I would retain the source version, extraction time, inclusion filter, selection rule and actual chosen identifiers. Replacing an unavailable file needs an explicit reason. If difficult cases are silently replaced by accessible ones, the sample can shift materially. Unavailability may itself raise a question about evidence management and retrieval, although it does not automatically prove that the underlying loan decision was wrong. The report should distinguish inability to obtain evidence from evidence that a required action was not performed. Those statements have different meanings and may justify different remedies.
I would use common core review questions alongside case-specific follow-up. Comparable results require reviewers to understand criteria consistently. A small calibration exercise can reveal disagreement before the main review and help improve ambiguous wording. I would record the observation separately from judgement of severity. An incomplete date, absent approval and incorrect calculation are not interchangeable findings. Clear classification supports aggregation and prevents a serious problem being obscured by numerous minor deviations. It also helps the reviewed team understand what evidence could resolve a question instead of repeatedly supplying documents that do not address it.
When I would extend the sample
If a problem clusters in one system version or period, I would design the extension to test that relationship, first checking that version and date data are reliable. In a manual workflow, the relevant characteristic might be application type or a handover between teams. More files should answer a specific question. Adding random cases without a clear purpose increases effort without necessarily strengthening the conclusion. I would state in advance what observation would support the proposed explanation and what would make me reconsider it, rather than adjust the interpretation only after seeing the results.
I would also examine cases where no defect was found. They can show the conditions under which a control works. If correct cases come from one channel and defective ones from another, that suggests a useful comparison rather than proof that the channel is the cause. Product and borrower mix may differ too. Appropriate matching or further evidence is needed. A sample is a starting point for reasoning; it does not automatically turn co-occurrence into causation. Keeping the alternative explanations visible makes an extension more focused and reduces the risk of an unfair generalisation.
Summary and recommendations
I recommend connecting purpose, population, selection method and conclusion scope in every review. Validate extraction completeness and record uniqueness, retain the selection trail and report components separately. A targeted finding can be important without statistical generalisation; explaining its limits makes it stronger. The historical European AQR illustrates a large review, not a universal template. For a particular task, an appropriate method and a well-defined question matter more than an impressive file count. The eventual recommendation should follow what the evidence can support and identify clearly what still needs wider investigation.