Line Assignment: The Decision Nobody Validates | HL Hunt
Line Assignment: The Decision Nobody Validates
The approval decision gets a model, a validation, monitoring, documentation, and governance review. The line amount assigned two seconds later is frequently set by a lookup table somebody built years ago and nobody has tested since — despite determining exposure at default, the customer's utilization, whether the product is usable for what they needed it for, and whether they come back. It's the most consequential unexamined decision in most consumer lending operations, and it has a property that makes the neglect worse: the line doesn't merely size the risk, it partly creates it.
What you'll learn
The line changes the risk
The observation that distinguishes line assignment from ordinary exposure sizing.
The line determines utilization, and utilization affects both the customer's profile and their behaviour.
Trace a customer given $700 who needed $2,400:
- They use the full $700 and remain short.
- Utilization is near 100%, which per our utilization guide lowers their score materially.
- Every other lender reviewing them sees a maxed account and a lower score.
- Other lenders reduce lines or decline, per the review dynamics in our timing analysis.
- Available credit falls further, raising utilization again.
- They have no headroom for the next shock — the failure our shock analysis describes.
- Performance deteriorates.
The same customer with a $2,400 line shows 30% utilization, a better score, and headroom. Their circumstances are identical; the assignment produced the difference.
Which means observed performance by line size is contaminated. Small lines are associated with worse outcomes partly because they were given to riskier customers and partly because being on a small line is itself harmful — and no amount of analysis on existing data separates those two, because the assignment rule created the correlation.
What a small line costs
The costs of conservatism are structurally invisible while the costs of generosity are itemized — the asymmetry our measurement analysis identifies as Family A.
| Line too large | Line too small |
|---|---|
| Higher exposure at default — visible in losses | Customer can't use it — no record |
| Attributable to the assignment | Lower revenue — attributed to demand |
| Appears in performance reporting | Customer goes elsewhere — invisible |
| Someone is accountable | Higher utilization worsens performance — attributed to the customer |
Every cost of a too-small line is either unrecorded or attributed to something else. So a manager reading their own reporting sees exposure they can control and no cost on the other side, and the rational response is to tighten.
Which produces the prediction: line assignment in operations that have never randomized should be systematically too conservative. That's the general Family A finding applied here, and it's testable.
The scale is worth noting. On a portfolio where the average customer would use meaningfully more capacity at a larger line, the foregone revenue can exceed the incremental losses by a wide margin — and the only reason that isn't obvious is that one side of it has never been measured.
Why you can't test it observationally
The methodological point that explains why so few lenders know whether their assignment is right.
The line is correlated with everything used to set it. Customers with larger lines have better scores, longer histories, and higher incomes — because that's what the rule conditions on. So a comparison of outcomes by line size measures the rule, not the line.
Statistical adjustment doesn't fix it. Controlling for the variables in the assignment rule removes the variation you're trying to study, and controlling imperfectly leaves the confounding in place. There is no observational answer to this question, which is why it persists as a matter of opinion in most operations.
The one accidental source of variation most lenders already hold: manual line overrides. These are cases where the assigned amount was set by judgment rather than by the table — a biased sample, but per our override analysis, a free and unexamined one. Worth looking at before designing anything, and not a substitute for what follows.
Randomizing the line
The only design that answers the question, and it's cheaper than it sounds.
For a share of approved customers, assign a line randomly within a band around the table amount.
| Design element | Approach |
|---|---|
| Share | A few percent of approvals |
| Band | Meaningful variation — a narrow band answers nothing |
| Direction | Both, since the question is which way the table is wrong |
| Tagging | Permanent, or the experiment produces cost and no information |
| Duration | Long enough for defaults to mature |
What to measure across the variation:
- Utilization distribution and how it shifts.
- Balances carried and revenue.
- Losses and exposure at default.
- Retention and repeat usage.
- Whether the customer took credit elsewhere, which is the demand the line failed to serve.
- Net contribution per customer, which is the number that decides it.
Cost the experiment honestly and it's usually small. The higher-line arm carries some additional exposure on a few percent of approvals; the lower-line arm carries foregone revenue. Both are bounded, and the answer applies to every future assignment.
The governance framing that gets it approved: this is a research cost, budgeted separately from credit performance. Booked in the same line as ordinary losses, the program gets cancelled the first quarter the higher-line arm shows a loss — which is the expected outcome of a working experiment, not a failure of one.
Utilization distribution, not average
The monitoring change that costs nothing and reveals most of the problem.
Average utilization is close to meaningless. A portfolio averaging 42% might be customers mostly around 40%, or a mix of customers who never use the line and customers permanently at the limit. Those are entirely different books and the average is identical.
What the distribution shows:
- A cluster near zero — customers who don't use the product. Either the line is irrelevant to them or something else is wrong, and either way they generate nothing.
- A cluster near the limit — the strongest available signal that lines are too small, and it's directly countable.
- The shape's movement over time, which is a portfolio signal in its own right.
The near-limit cluster deserves acting on. Customers persistently at their limit are demonstrating unmet demand from creditworthy borrowers you already approved — the cheapest growth available, and the population where a targeted increase test would show its effect fastest.
This connects to the segmentation point in our segmentation analysis: distributions differ across populations, so this should be tracked by segment rather than portfolio-wide, where two opposite problems can cancel out.
Increases
The decision with the best available evidence, because behavioural data has accumulated.
Per our renewal analysis, behaviour on the account dominates application data once history exists. An increase decision made primarily on a refreshed bureau score is ignoring the better predictor sitting in your own system.
What predicts well:
- Payment behaviour on your account.
- Utilization pattern — persistently high with good payment is a different signal from erratic.
- Whether they've requested an increase, which is information about their demand.
- Tenure.
- Bureau data as a supplement, not the primary input.
Design points:
- Proactive increases generally beat reactive ones — a customer who had to ask has already experienced the constraint, and may have solved it elsewhere.
- Randomize a share of increase decisions, for the same reason as initial assignment.
- Watch what happens to utilization after an increase. A customer who returns to the same utilization at a higher line is telling you the original line was binding; one whose balance stays flat wasn't constrained.
That last diagnostic is unusually informative and costs nothing to compute.
Decreases
The decision requiring the most care, because the consequences run outward.
A decrease raises the customer's reported utilization immediately, lowering their score and potentially triggering reductions by other lenders — which lowers their available credit further. The reduction can contribute to the deterioration it was meant to avoid, and where it's driven by portfolio management rather than the customer's own behaviour, it affects someone who has done nothing wrong.
What follows:
- Distinguish behaviour-driven from portfolio-driven decreases, and hold the second to a higher bar.
- Consider whether reducing unused capacity is worth its cost. An unused line carries exposure and provides the headroom that prevents distress — and cutting it removes the protection without recovering anything already drawn.
- Meet notification obligations, per our notices guide, with reasons that trace to real facts as our explainability analysis requires.
- Monitor for disparity, since portfolio-driven decreases can concentrate in ways that raise fair lending questions under our governance framework.
- Measure what decreases actually save against what they cost in retention and induced deterioration — a comparison almost nobody makes.
Governance
The gap this whole report describes, stated plainly.
| Applied to approval | Applied to line assignment |
|---|---|
| A developed model | Frequently a table |
| Validation | Rarely |
| Ongoing monitoring | Sometimes |
| Documentation | Thin |
| Fair lending analysis | Frequently not |
| Governance review | Rarely |
The fair lending row is the one that should concern people most. Line assignment allocates a benefit among approved applicants, so disparity in assigned amounts is a disparity in treatment — and an operation testing approval rates by group while never testing assigned lines is examining half of its decision.
The minimum: document the rationale, monitor assigned amounts by segment and by group, test the strategy with randomization, and review it as you'd review a model. None of that is expensive, and the first line increase test typically pays for all of it.
The line is a decision, so treat it like one
HL Hunt AI Underwriting supports line assignment strategy alongside approval, with randomized assignment bands, utilization distribution monitoring by segment, and the same reason-code traceability and fair lending reporting applied to both decisions.
Frequently asked questions
The line determines utilization, which affects the customer's score, how other lenders treat them, and whether they have headroom. The assignment partly creates the risk it appears to reflect.
Mostly invisible things — unusable product, lost revenue attributed to demand, customers who don't return, and worse performance attributed to the customer.
Randomize the amount within a band for a share of approvals. The line is correlated with everything used to set it, so no observational analysis can separate rule from effect.
That it raises utilization immediately, can trigger reductions elsewhere, and may contribute to the deterioration it was meant to avoid — while removing the headroom that prevents distress.
Key takeaways
- The line partly creates the risk it appears to reflect, by driving utilization and how every other lender treats the customer.
- Every cost of a too-small line is unrecorded or attributed elsewhere, so assignment drifts conservative in operations that have never tested it.
- No observational analysis works — the line correlates with everything used to set it, so randomization within a band is the only answer.
- Track the utilization distribution, not the average; a cluster at the limit is directly countable unmet demand from approved customers.
- Behaviour on your own account beats a refreshed bureau score for increase decisions, and proactive beats reactive.
- Line assignment allocates a benefit among approved applicants, which makes it a fair lending surface most operations never test.
Find out what the line is actually doing
Get started with HL Hunt AI Underwriting for approval and line assignment as governed decisions — randomized test arms, distribution monitoring, and outcome reporting that shows what the line changed rather than what it correlated with.
This guide is educational and does not constitute legal or compliance advice. Line decreases and other adverse actions carry notification obligations, and line assignment is subject to fair lending requirements. Randomized assignment programs raise governance and compliance considerations that should be addressed before implementation — consult qualified counsel and your model risk function.