These analytical notes explore a field I want to develop in. The examples are hypothetical and do not describe inspections I have carried out.
An output has a history
A rating summarises information using a particular methodology. Understand model version, input period, data freshness and scale. Identical symbols can represent different risk levels across systems.
An override needs a trail
Expert adjustment may capture information absent from the model. Trace the rationale, approving person and supporting evidence. Frequent overrides in one direction warrant investigation but do not automatically prove an error.
Hypothetical example
A borrower receives a better final rating because of expected group support. I would examine the evidence for that support and how the methodology permits its use. A traceability check does not replace model validation, but can expose missing support for a decision.
Checks that do not require model development
Even without developing a rating model, an analyst can trace inputs and outputs. Does the assessment concern the right borrower and period? Did refreshed data leave an outdated result in place? Which information supported the override? These checks address operational weaknesses without claiming to validate the mathematics.
Separate model output, final assessment and subsequent use. One report may display the initial rating and another the adjusted one for a valid reason. The difference must fit the purpose and be understandable; unclear field names can make unlike measures appear comparable.
Reading the override overview
For group support, examine the argument’s freshness and consistency across similar cases. Different outcomes may be justified by different circumstances. Documentation should explain that distinction rather than leave it to guesswork.
An overview can suggest areas for investigation: a particular segment, frequent changes shortly before approval or repeatedly generic reasons. Statistics alone establish none of these weaknesses, but help select cases for examining decision quality.
A real case: Moody’s and departures from model-implied ratings
On 28 August 2018 the US SEC announced settlements with Moody’s concerning rating-model controls and rating symbols. Its findings included instances where reasons for material departures from model-implied ratings were not documented. The firm neither admitted nor denied the charges. The case concerns agency ratings of US securities, not Czech banks’ internal ratings; the transferable lesson is decision traceability, not automatic application of identical rules.
For me, the case raises a question about understanding expert judgement after the event. A departure need not itself be wrong: a model may omit important information that an authorised specialist assesses. The difficulty arises when the record does not show what information mattered, who decided and why a particular argument outweighed the original result. Without that trail, distinguishing a qualified adjustment from inconsistent decision-making becomes much harder.
Hypothetical example: a better rating after a one-off gain
Imagine a company whose model result indicates elevated risk because of debt and weak operating cash flow. The business proposes an improved category based on a strong annual profit. An analyst finds that part of the profit came from selling a property. This does not determine the correct rating grade. It creates a need to explain whether the argument demonstrates sustainable repayment capacity or a one-off accounting outcome.
In fictional figures, the company generates CZK 8 million of operating cash flow against CZK 10 million of annual debt service. An asset sale contributes another CZK 15 million. That receipt may improve liquidity or permit debt reduction, but it should not silently become recurring operating income. These numbers illustrate a question for analysis; they are neither a rating model nor a recommendation about lending to a real borrower.
How I would examine the adjustment rationale
I would obtain the original model result, inputs, proposed adjustment and documented justification. I would distinguish correcting erroneous data from applying expert judgement to a result based on correct inputs. Using the wrong financial statements is different from deliberately assigning greater weight to a qualitative factor. A single “manual change” category can conceal these differences and make later monitoring difficult to interpret.
Next I would verify the decision-maker’s authority and the methodology in force. The relevant specialist would assess the argument’s substantive adequacy. My analytical contribution would be to reconstruct what information was actually available at the decision date. A later favourable outcome does not by itself demonstrate adequate original evidence; similarly, a later loss does not automatically prove that the earlier judgement was carelessly made.
From individual files to adjustment patterns
Across a population I would examine frequency by segment, direction, approver and period. A concentration is not automatically evidence of misconduct: it may reflect a team’s specialisation or portfolio mix. It can nevertheless guide the selection of files for review. Consistent change definitions are important, as is including applications that were ultimately declined, rather than examining only successful approvals and overlooking the selection effect.
I would avoid ranking individual employees without context. Someone handling all unusual cases may naturally have more adjustments. The purpose is to understand the decision mechanism, not assign blame from a simple table. If the wording “good knowledge of the customer” recurs, I would ask what specific information that knowledge represents in each file and whether it can be independently evaluated.
Checking the quality of judgement over time
For a hypothetical set of adjustments, I would propose monitoring subsequent borrower development and comparing it with the reasoning recorded at the time. This requires a defined horizon and an appropriate comparison group. Otherwise unlike cases may be compared and a persuasive-looking conclusion produced. Findings should inform methodology, training or controls rather than assume every adverse outcome was caused by an inappropriate manual intervention.
I would retain rejected adjustment proposals as well as accepted ones. Without them, the evidence omits part of the decision activity and may falsely suggest an absence of challenge. A precise record supports review and learning, while access and retention should match the sensitivity of the customer information. The aim is enough evidence to understand the decision, not unnecessary duplication of entire confidential files.
An additional check would compare the recorded reason with the direction and size of the change. A small data correction that produces a very large rating movement deserves explanation, as does an improved grade justified by a factor that has deteriorated. These are investigation prompts, not automatic error rules. They help select the cases where a specialist’s limited time can produce the clearest understanding of whether the process is functioning consistently.
Summary and recommendations
Distinguish data correction, model change and expert adjustment. Require a specific rationale, authorised decision and traceable supporting information for judgement. Use population patterns to target review, and interpret subsequent outcomes in context. My proposed approach would not replace specialist judgement with a model number; it would help establish why the team relied on a result and how the reasoning can be independently understood, tested and challenged when the evidence warrants it.