A credit standard describes the conditions under which a bank is willing to accept risk. It is more than a minimum rating or a list of mandatory documents. What matters is how rules affect decisions, how exceptions are handled and whether the institution understands the quality of its new lending. For analytical work, the interesting boundary is between a well-written policy and its operation. I would examine the data, decision path and evidence of challenge as well as the existence of the rule itself, without assuming that a completed checklist proves a sound lending decision.
A verified development in 2026
On 24 February 2026 ECB Banking Supervision in Frankfurt published Sound credit, sustainable growth, announcing an underwriting thematic review and data collection planned for the first half of 2026. This is a documented supervisory initiative, not published findings against an individual bank. It does not establish a breach by a Czech institution or automatic application of a particular measure in the Czech Republic.
My analytical interpretation is that growth in new lending requires more than a review of the outstanding balance. New vintages may differ in income, purpose, maturity or exception use. I would ask whether the risk actually accepted is changing while the written standard remains unchanged. Answering that requires information about applications and decisions, not only loans that were drawn. Rejected applications and changes made before approval can help explain whether an apparently stable portfolio reflects consistent standards or a changing selection process. Those are hypotheses to investigate, not conclusions from the announcement.
Turn a principle into an observable control
A policy may state that income must be evidenced and lending consistent with repayment capacity. For testing, that is still too broad. I would establish which evidence is accepted, how freshness is verified, how irregular income is treated and where approval responsibility is recorded. This does not replace judgement with tick boxes. It makes it possible to understand retrospectively which information was available and why a decision was considered acceptable. A reviewer should also know what the decision-maker could see at the time, rather than rely on documents added later.
Exceptions need particular attention. An exception can be justified and properly authorised; its presence alone is not a defect. The concern is an undocumented bypass or a practice that diverges from policy. I would examine reason, extent, approver and repetition by product, branch and distribution channel. Individual approval and aggregate monitoring answer different questions. A collection of reasonable-looking decisions can still create a concentrated exposure. Equally, a higher exception count may reflect improved recording rather than looser underwriting. The definition and denominator must be checked before interpreting a change.
A hypothetical example: unchanged volume, different composition
Imagine a bank originating CZK 1 billion in each of two quarters. Recorded exceptions rise from five to fifteen per cent of contracts. My first question would not be whether fifteen per cent is inherently excessive. I would verify that definition, denominator and recording practices are unchanged, then compare purpose, maturity, individual amounts and approval reasons. The movement could reflect a new product or a concentrated weakness in one channel. A clear analytical result would show which explanation the evidence supports and what still needs to be established from the underlying decisions.
Suppose ten large exceptions represent forty per cent of new volume. A contract-count measure understates their significance. I would add an amount-based view and investigate common borrower characteristics. For selected cases I would return to the original application and repayment evidence. I would not label this proof of future losses: at this stage it describes a shift in accepted risk and the operation of controls. Specialist assessment and further information remain necessary. This distinction prevents a useful early indicator from being overstated as an established prediction of credit performance.
How I would investigate
I would compare policy, system configuration and actual decisions. For each important rule I would select a normal approval, a rejection and an authorised exception. If a limit is automated, I would test the boundary and changes during approval. Does an increase in amount or maturity trigger reassessment? A system may correctly check the first application version but allow a changed version through without review. Investigating that requires both process analysis and application testing. The evidence should identify the exact version checked and the version ultimately approved, rather than merely a successful control timestamp.
I would then create a consistent new-lending dataset using approval periods. It should connect application, contract and customer identifiers with decision date, rule version, significant inputs and approval history. Performance comparisons should use vintages observed for similar periods. A newly originated loan has had little time to show problems, whereas an older loan has faced more payment dates and business changes. An apparently favourable recent outcome is therefore not automatically evidence of equal quality. The observation period is part of the interpretation, just as important as the headline arrears percentage.
Questions for the specialist team
I would ask which portfolio changes are intentional and how the bank evaluates them. How does it recognise that exceptions are becoming routine? What happens when an employee selects the wrong exception reason? Who monitors combined effects across business units? I would connect each explanation to a document, report or decision. A general assurance that every case is individually approved does not answer questions about concentration or aggregate risk. It also leaves open whether approvers receive consistent evidence and whether the same weakness could occur independently in several channels.
My finding would distinguish missing documentation from an inappropriate lending decision. The former can be evidenced by an absent required record; the latter requires specialist assessment of the whole case. If evidence exists elsewhere, the issue may concern recording and accessibility rather than an assessment never having occurred. That distinction supports proportionate remediation. The aim is not to maximise failed checklist items but to establish whether the process makes risk acceptance deliberate, understandable and open to challenge. A precise limitation is more valuable than a strong conclusion unsupported by the selected evidence.
Summary and recommendations
I would assess underwriting standards through decisions and new-lending data. My recommendations are to retain rule versions and exception histories, compare measures by count and value, and verify that significant application changes trigger reassessment. Combining policy review, system tests and loan-file evidence can reveal a mismatch missed by any one view alone. The 2026 European initiative is a useful source of analytical questions, not a substitute for applicable Czech requirements or a ready-made finding about a selected institution. The practical objective is a traceable connection between the rule, the evidence and the risk accepted.