Customer risk scoring: building a model you can defend.

What a rating is for, and what it is not

A customer risk rating exists to answer one operational question: how much attention does this relationship need? It decides the depth of due diligence at onboarding, how closely activity is monitored afterwards, how often the file is reviewed, and who has to approve the relationship.

That is the whole job. It is worth saying plainly, because ratings are routinely asked to do things they cannot.

A rating is not a prediction that a customer is a criminal. It is an assessment of exposure — how attractive this combination of customer, product, channel and geography would be to someone laundering money, and how hard it would be to see if they did. A high-rated customer is overwhelmingly likely to be entirely legitimate.

A rating is not a credit decision, a profitability measure, or a verdict on a person. Where those are allowed to leak into the model — consciously or through a proxy — it stops measuring financial crime risk and becomes untrustworthy for every purpose.

Inherent risk, controls, residual risk. Most methodologies rate inherent risk: the exposure before the institution's own controls are considered. Some then apply mitigation to reach a residual rating. Both approaches are defensible; what is not defensible is being unclear which one you are doing, because the two mean different things to an investigator reading the file and to a supervisor reading the methodology.

There is no single correct method, and that is official. The Wolfsberg Group notes in its 2024 Country Risk FAQs that there is no universally agreed approach to determining financial crime country risk — and the same is true of customer risk more generally. What exists is a requirement to be deliberate: to choose factors for stated reasons, to combine them in a way you can explain, and to be able to show the result is sensible.

The FATF's risk-based approach is what makes this necessary. Recommendation 1 requires institutions to identify and assess their risks and apply measures proportionate to them. Recommendation 10 requires due diligence calibrated to that assessment. The logic runs in one direction: assess, then apply effort where the risk is. A rating model is simply the machinery that does this consistently across a book too large to judge case by case.

One warning about perfectionism. The Wolfsberg FAQs quote the FATF's 2021 work on risk-based supervision to the effect that a zero-tolerance approach which does not tolerate imperfections — particularly in areas identified as lower risk — is counterproductive to an effective system. A model does not have to be right about every customer. It has to be right about where the risk concentrates, and honest about its own limits.

Choosing factors, and where country risk belongs

A factor earns its place in a model by changing the answer. If every customer scores the same on a factor, it carries no information and only adds the appearance of rigour.

The four families, from the risk-based approach:

• Customer — legal form and complexity of ownership, industry, whether a politically exposed person is involved, adverse media, the plausibility of the stated purpose, cash intensity of the business.
• Product and service — how easily the product moves value, to whom, and how visibly. Correspondent banking, private banking, trade finance, pooled or omnibus accounts and anything bearer-like sit high for structural reasons.
• Channel — face-to-face, intermediated, fully remote, introduced by a third party. Non-face-to-face is not inherently high risk; what matters is what identity assurance the channel actually achieves.
• Geography — domicile, country of incorporation, centre of activity, and other nexus such as country of tax residency.

Country risk deserves its own treatment, because it is the factor most often misused. The Wolfsberg Group is direct: country risk does not, by itself, determine the risk of a customer. It is one factor among many, and the FAQs frame the balance precisely — disregarding country risk discards important context about a customer, while placing too much emphasis on it produces undesirable outcomes across the customer base. That second half is the one institutions underweight. A model dominated by geography will rate an entire nationality or diaspora as high risk, generate enormous volumes of work that finds nothing, and push the institution towards cutting off whole categories of legitimate customer.

Where country ratings come from. Institutions usually build a country score from public sources — FATF public statements and mutual evaluation reports, national risk assessments, sanctions regimes, corruption and governance indices, tax transparency lists. The Wolfsberg FAQs set out sensible constraints: there is no prescribed number of sources; sources with no editorial process, or with serious unresolved allegations about completeness or data quality, should be avoided; and the effect of a change in any data source on the final rating should be clearly identifiable and understandable to internal and external readers. If nobody can say why a country moved, the methodology is a black box.

Missing data is itself a risk factor. The FAQs state the principle that unknown information is a priori a risk factor, and that a data quality assessment should be done where data is missing. Where a country is poorly covered, the options are to elevate it to a higher rating or to exclude it from the automated process and rate it subjectively — with the choice recorded. Silently treating "no data" as "no risk" is the failure mode, and it is common: a country with no index coverage drops to the bottom of the list and nobody notices.

Beware proxies. Nationality, name origin, language and residence status correlate with protected characteristics. A model that uses them as shortcuts for risk will produce discriminatory outcomes and will be very difficult to defend — to a supervisor, to a court, or to the customers it excludes.

From factors to a score

Once the factors are chosen, three decisions determine what the model actually does: how each factor is scored, how the scores are combined, and when a human may overrule the result.

Scoring each factor. The Wolfsberg Group's risk assessment guidance says a methodology should clearly articulate the factors being assessed, the criteria used to score them, the weightings used, and any scoring overrides together with the rationale for them. In practice that means writing down what makes a factor high, medium or low — in terms someone else could apply and reach the same answer. A scoring rule that only its author can apply is not a methodology.

Weighting. Each factor is normally assigned a weight reflecting how much risk it carries and how prevalent that risk is relative to other factors. The weight given to each category is discretionary — it is the institution's judgement, not a given — which is exactly why it has to be justified and recorded. The Wolfsberg FAQs add a useful technical constraint: weight-based models should consider the statistically significant contribution of each factor. A factor weighted at two per cent that never moves an outcome is not a safeguard; it is padding.

Aggregation, and the choice that matters most.

• Additive or weighted average — scores are summed. Smooth and easy to explain, but it allows one genuinely serious factor to be diluted by many benign ones. A customer with a sanctioned-jurisdiction nexus can average down to medium.
• Highest-factor-wins — the overall rating is driven by the worst single factor. Conservative and simple, but it compresses everything upward and destroys discrimination: eventually most of the book is high risk and the rating stops directing effort.
• Hybrid — a weighted score, plus a small set of stated rules that force a rating regardless of the average. This is where most defensible models land.

Overrides, and what they reveal. An override is a deliberate departure from what the methodology produced. Two kinds, and the distinction matters:

• Rule-based, built into the methodology — for example, a defined category always treated as high risk. These are not really overrides; they are part of the model and should be documented as such.
• Discretionary, applied case by case by a person. The Wolfsberg guidance is that any override mechanism should be clearly documented within the methodology and used only in exceptional circumstances.

The most useful idea in this whole area is what recurring overrides mean. Wolfsberg's position is that recurring overrides are typically an indication that, in some respect, the methodology is not working as intended. Treat your override log as a defect report on the model. If the same category is manually downgraded forty times a year, the model is wrong about that category and should be changed — not quietly corrected forty more times. An override log nobody analyses is a missed diagnostic, and a bad look when someone else reads it.

Who approves it. The methodology and its outputs should be approved by an authorised senior manager or a governance forum whose members have appropriate financial crime knowledge and enough seniority to withstand regulatory scrutiny. A model owned by nobody senior is a model nobody will defend.

Keeping the rating true

A rating describes a customer at a moment. Customers change. A model that only recalculates when someone opens the file is describing the past.

Periodic review — reassessing every customer on a cycle set by rating, commonly annually for high risk and every few years for low. It is predictable and easy to evidence, and it is slow: a customer whose business changes the month after review carries a stale rating for a long time.

Event-driven review — recalculating when something happens that could change the answer. This is where the real protection is, provided the triggers are defined and actually wired up:

• A change in ownership, control or beneficial ownership.
• A change in the nature of the business, or activity that does not match the stated purpose.
• A politically exposed person entering or leaving the relationship.
• Adverse media, a law enforcement request, a production order, or an internal suspicion report.
• A change in the country rating that affects the customer's nexus.
• An alert pattern inconsistent with the customer's rating.

That last one deserves emphasis. If monitoring repeatedly alerts on a customer rated low, two systems are disagreeing with each other. Resolving that disagreement is far more valuable than closing the alerts.

Perpetual KYC — continuously recalculating as data changes, rather than on a cycle. It is where the industry is heading and it is genuinely better, but it has a governance cost. If ratings move constantly, you must be able to reconstruct what the rating was on any past date and why, or you will be unable to explain the due diligence applied at the time. Keep the audit trail or the improvement becomes a liability.

Country ratings move too. The Wolfsberg FAQs suggest updating country ratings no less frequently than annually, following the updates of the underlying data sources, many of which publish only once a year — and choosing a date when most of the data is current. They also recommend interim revision when key trigger data lands: FATF public statements showing material deficiencies, new mutual evaluation reports with technical compliance and effectiveness ratings, or a country becoming subject to sanctions. The practical question for an institution is the one that follows: when a country rating moves, what happens to every customer with a nexus to it? If the answer is "nothing until their next periodic review", the country update achieved very little.

Proving the model works

A rating model is a model, and it attracts the obligations of one whatever the institution calls it. The Wolfsberg FAQs note there is genuine debate about whether these should be treated as models, methodologies, tools or applications. The debate matters less than the discipline: you need to be able to show the thing works.

Look at the distribution before anything else. Pull the population across rating bands. The failure modes are visible immediately:

• Almost everything is low. Either the book genuinely is low risk and you can evidence that, or the model is not discriminating and the high-risk cases are hiding in the bulk.
• A large share is high. The rating has stopped directing effort. Enhanced due diligence applied to a third of the book is enhanced for nobody, and the real exposures are queued behind ordinary customers.
• Everything is medium. The model is averaging its way to no answer.

Back-test against what actually happened. Take the customers who generated genuine suspicion — confirmed reports, law enforcement interest, cases that came to light some other way — and look at how they were rated before. If they were mostly medium or low, the model is not capturing the risk that materialises in your institution, and no amount of internal logic will fix that. This is the single most informative test available, and it is the one most rarely performed.

Test the mechanics as well. Recalculate a sample by hand and confirm the system agrees. Change one factor and confirm the rating moves as the methodology says it should. Check that customers with identical attributes receive identical ratings — divergence means undocumented logic or data problems. Confirm that a factor carrying real weight actually moves outcomes in the live population.

Keep the record. Documented rationale for every factor and weight; the methodology version applied to each rating; an approval trail; override logs with reasons, analysed; validation results including the ones that were unflattering; and the ability to reconstruct any customer's rating on any past date.

The cost of getting it wrong runs both ways. A model that under-rates leaves real exposure unexamined. A model that over-rates is not merely cautious — it buries the genuine cases in volume, and it pushes the institution towards de-risking: exiting whole categories of customer because assessing them individually costs too much. The FATF has been explicit that wholesale de-risking is not a risk-based approach; it is the opposite, because the business does not stop, it moves somewhere with less visibility. A rating model that cannot tell a legitimate money remitter from a high-risk one will eventually recommend abandoning all of them, and that outcome is worse for the financial system and for the people cut off from it.

What good looks like. The model is understood by the people who use it, its reasoning is written down, its output is checked against reality, its overrides are read as information, and the institution can explain — to a supervisor, in a sentence — why this customer is rated as they are.

Your certificate of completion

Enter your name as you want it to appear, then save the certificate as a PDF. In the print window, choose Save as PDF. A certificate is issued once per completion of the course.

Saolix does not record who takes this course, so it cannot verify these certificates. The certificate confirms completion of a free self-paced course and is not an accredited qualification.