The Boundary Problem: What Happens at a Cutoff

Two applicants sit one point apart on a score. Their predicted risk differs by an amount too small to act on, and their outcomes differ completely: one is approved at a good rate, the other is declined or priced several points higher. Every lending system contains lines like this, and they are not a design flaw, because a binary decision has to come from somewhere. But the line has three properties worth knowing. It concentrates the cost of every data error. It creates the strongest possible incentive to move the measured number. And it is the cheapest experiment in lending, sitting unused in almost every portfolio.

Where the lines are

Boundary	What changes when you cross it
Approval cutoff	Approved or declined
Pricing tier	The rate, in a discrete step
Limit band	How much you can borrow
Policy knock-out rule	Everything, regardless of score
Verification threshold	Whether documents are required
Collections treatment band	Who contacts you, and how
Retention period ending	Whether an old item is visible at all

Most of these are invisible to the person they affect. Per our decline guide, an applicant learns the outcome and not how close it was, which means nobody on the wrong side of a line knows they were on the wrong side of a line.

The last row is the subject of our record retention analysis, and it is the same structure in a different variable: full weight on one day, nothing the next. Per our benefits analysis, the pattern also runs through public programmes, where crossing an income line can cost more than the extra income.

Why they are unavoidable

Worth establishing before criticizing anything, because a report that treated cutoffs as an error would be wrong.

• Many decisions are binary in nature. An application is approved or it is not. There is no way to approve 63 percent of a card application.
• Products exist at defined terms. Rates and fees are usually set in tiers because a product has to be described, disclosed, and administered.
• Operational systems need rules. Per our policy analysis, a decision process built from continuous judgments is hard to run consistently and harder to review.
• Consistency is itself a protection. A rule applied identically to everyone is testable for disparate impact in a way that case-by-case judgment is not.

So the question is never whether to have boundaries. It is how much rides on each one, and whether anyone is measuring what happens around it.

The question is not whether to draw a line. It is how much should change when someone crosses it.

Errors concentrate at the line

The first consequence, and it follows directly from how a cutoff works.

A data error changes an outcome only if it moves someone across a line. An applicant with a strong file who has one incorrect item is still approved. An applicant with a weak file who has one incorrect item is still declined. The only people whose outcome changes are those near the boundary.

Which means:

• The cost of file errors falls almost entirely on near-boundary applicants, per our errors analysis.
• The same is true of attribute definition problems, per our attribute analysis: a changed definition moves scores slightly, and slightly is enough at the line.
• Stale data has the same property. A file that has not caught up with a paid-off balance matters only for people close to a threshold.
• Model uncertainty is largest exactly there. A prediction of 5.9 percent against a cutoff of 6 percent is a distinction the model cannot really support, and the decision treats it as decisive.

The fourth point deserves more weight than it usually gets. A score is an estimate with error around it. At the boundary, the difference between two applicants is smaller than the uncertainty in either estimate, so the decision is being driven by a difference the model cannot actually distinguish from noise. That is not an argument against cutoffs. It is an argument for knowing how wide that zone is, and for treating it differently from the rest of the distribution.

The incentive to cross

The second consequence, and it connects to a finding this desk has made elsewhere.

Per our score optimization analysis, when a number decides access, people and products work to move the number. A cutoff makes that incentive as sharp as it can be: a small, cheap change in the measured variable is worth a large change in outcome, and only for people near the line.

What this produces:

• Effort concentrated just below the line, where the return per point is highest.
• Bunching just above it, as people who optimized land where they aimed.
• Intermediaries who know the thresholds and coach applicants toward them, per our channel analysis.
• Timing effects, where applications are held until a file crosses a known line.

Bunching is measurable, and it is diagnostic. If the distribution of applicants is smooth everywhere except for a pile immediately above a cutoff, something is moving people across it, and the something is usually presentational rather than substantive. A lender can detect gaming without knowing anything about the methods being used, simply by plotting the density of applicants around its own thresholds.

The experiment nobody runs

The constructive core of this report, and the reason boundaries are worth more attention than they get.

Per our measurement analysis and our reject inference analysis, the central difficulty in credit measurement is that declined applicants produce no outcomes, so a lender cannot see what would have happened had it approved them.

At the boundary, that difficulty partially dissolves. Applicants just below and just above a cutoff are similar in every respect that matters, because which side of a line a continuous score falls on is close to arbitrary within a narrow band. So:

1. Take applicants within a narrow band around the cutoff.
2. Compare those just above with those just below.
3. The difference in their later outcomes estimates the effect of the decision, not the effect of being a different kind of applicant.

That comparison answers questions lenders otherwise guess at:

• How would near-miss declines have performed? Locally, this is answerable.
• What does a price step actually do to default and to acceptance, per our acceptance analysis?
• Does a higher limit change repayment behaviour, per our line analysis?
• Does a verification requirement improve outcomes enough to justify the applicants it loses?

Two honest limits. The estimate is local: it tells you about applicants near that cutoff, not about applicants far from it, so it cannot justify moving the line a long way. And it fails if people control which side they land on, which is exactly what bunching indicates. The two analyses therefore go together: check the density first, and if it is smooth, the comparison is trustworthy.

Per our testing analysis, a deliberate test in a narrow band produces cleaner evidence still, though randomizing decisions raises fairness and disclosure questions that belong with counsel before anything is designed. The observational version requires no change to the process at all, which is why it should come first.

Near-identical, treated differently

The fairness dimension, stated carefully because reasonable people differ on it.

A cutoff means two people who are almost the same receive materially different treatment. That is defensible when the line rests on something meaningful and applied consistently, and it becomes harder to defend as the difference in treatment grows relative to the difference in measurement.

What deserves attention:

• Size of the step. A boundary that changes price slightly is easier to justify than one that changes approval entirely.
• Quality of the variable. Per our monoculture analysis, if every lender uses similar inputs, an applicant just below the line is below it everywhere, and the consequence is not one lender's decision but exclusion.
• Distribution around the line. Per our fair lending analysis, if groups are distributed differently around a threshold, moving it has different effects on each, which is a question worth examining before the line moves.
• Whether anyone can ask for a second look. Per our reconsideration analysis, a near-boundary decline is the clearest case for review, and per our override analysis, near-cutoff overrides are where reviewers most often turn out to have been right.

That last point is the practical bridge. If your own override data shows that near-boundary reversals perform well, the boundary is sitting in a place the model does not support, and the evidence is already in your systems.

Making the step smaller

The design response. The aim is not to remove boundaries but to reduce how much depends on any single one.

1. Graduated pricing. Several small steps rather than one large one, so a point costs less.
2. Graduated limits, per our line analysis, which turn a binary into a quantity.
3. A defined review band around the cutoff, routed to a second look rather than an automatic decline.
4. Counter-offers instead of declines, per our acceptance analysis, which convert a boundary into a negotiation about amount.
5. Recency windows instead of presence rules in policy knock-outs, per our retention analysis.
6. Data quality effort focused on near-boundary applicants, since that is where accuracy changes outcomes.
7. Monitoring of the population near each line, including density, as a standing report.

Item three is the most useful and the most resisted, because it costs manual review time. The answer to that objection is item six and the measurement above: if near-boundary applicants perform like those just above the line, the review is recovering business the rule is throwing away, and the size of that recovery is measurable before any commitment is made.

The strongest objections

"Someone always has to be last in the queue." True, and this report accepts it. The claim is not that the unlucky applicant can be eliminated. It is that the magnitude of what changes at the line is a choice, and that most lenders have never examined how large their steps are or what happens either side of them.

"Boundary analysis only tells you about the boundary." Correct, and stated above. Local estimates cannot justify large policy moves. They are valuable precisely for the decision most often taken without evidence, which is whether the line is in the right place by a small margin.

"Softening cutoffs is a way to loosen credit standards." It can be, if it is done without measurement. The argument here is the opposite of loosening on instinct: measure what happens at the line, then decide. That may equally show the cutoff should be tighter, and a lender that finds near-boundary approvals performing badly has learned something just as useful.

Testable implications

1. Applicants just below and just above a cutoff should perform almost identically when both are approved, which is the test of whether the line marks a real difference.
2. The density of applicants should be smooth across the cutoff. A pile just above it indicates that the measured variable is being managed.
3. Near-boundary overrides should outperform other overrides, which is partly evidenced already.
4. File errors should change outcomes disproportionately for near-boundary applicants, making error correction most valuable there.
5. Price steps at tier boundaries should show acceptance effects larger than the risk difference justifies.
6. Graduated structures should produce better aggregate outcomes than single large steps, at equivalent overall risk appetite.

The first two go together and cost almost nothing. Plot the density of your applicants around each cutoff, then compare outcomes for the narrow bands either side. If the density is smooth and the outcomes are similar, the line is not marking the difference it claims to mark, and it can be examined honestly. If the density piles up above the line, you have found evidence of optimization before anyone has told you it exists.

The conclusion we would hold: cutoffs are necessary, their costs fall on a small and identifiable group, and they are simultaneously the best natural experiment a lender has. A boundary that nobody studies is pure cost. A boundary that is measured pays for itself, because it answers the question the rest of the portfolio cannot.

Frequently asked questions

---

This report presents an analytical framework and the authors' interpretation; it is not legal, compliance, or model risk advice. Analyses of decision boundaries, changes to cutoffs, and any randomized testing of credit decisions raise fair lending, disclosure, and model governance questions that vary by product and jurisdiction; consult qualified counsel and your compliance and model risk functions. The implications identified as testable are hypotheses.