After Approval: The Selection Hidden in Who Accepts | HL Hunt
After Approval: The Selection Hidden in Who Accepts
Underwriting decides who gets an offer. The applicant decides whether to take it — and that second decision isn't random. Applicants with strong files usually have other offers and walk away more often; applicants with fewer options accept more often. So the booked portfolio can be riskier than the approved one even when the model is right about every individual, and raising prices makes the effect worse. Most lenders never measure it, because the people who walked away leave no outcomes behind.
What you'll learn
Two decisions, not one
| Stage | Who decides | What the lender sees |
|---|---|---|
| Approval | The lender | Everything |
| Acceptance | The applicant | Only the outcome — yes or no |
| Performance | Circumstances | Only for those who accepted |
Per our reject inference analysis, lenders know they can't see how declined applicants would have performed. Acceptance creates a second blind spot of the same kind: approved applicants who walked away also leave no outcomes — and they aren't a random slice of the approved group.
Per our search analysis, applicants who shop compare offers; those with the strongest files have the most offers to compare, and they take the best one. Any single lender sees its own offer turned down by exactly the applicants it most wanted.
The gap, worked through
A stylized lender approves 1,000 applicants in two segments at one price:
| Segment A — strong alternatives | Segment B — few alternatives | Total | |
|---|---|---|---|
| Approved | 500 | 500 | 1,000 |
| Predicted default | 3% | 6% | 4.5% |
| Acceptance rate | 40% | 80% | 60% |
| Booked | 200 | 400 | 600 |
| Booked default rate | 3% | 6% | 5.0% |
The model is right about both segments, and the book still defaults at 5.0% instead of the 4.5% the approvals implied. Nothing went wrong in underwriting — the mix shifted after it, because the lower-risk segment turned the offer down more often.
A lender pricing on the 4.5% it approved at is underpricing what it actually booked. The gap is invisible unless someone compares the approved population with the booked one.
What a price rise does
The feedback loop that makes this a pricing problem.
Raise the price on the same approvals. Applicants with alternatives are the most price-sensitive, so segment A's acceptance falls further than segment B's:
| Before | After a price rise | |
|---|---|---|
| Segment A acceptance | 40% — 200 booked | 25% — 125 booked |
| Segment B acceptance | 80% — 400 booked | 75% — 375 booked |
| Booked default rate | 5.0% | 5.25% |
| Booked volume | 600 | 500 |
The price went up and so did the risk. Part of the extra revenue per loan is consumed by the higher loss rate, on lower volume. Per our policy analysis, that's why performance should be monitored by offered price, not only by risk band.
Per our rationing analysis, this is the classic reason lenders ration rather than raise prices indefinitely: beyond some point, a higher price selects a worse pool fast enough that it earns less, not more. Acceptance selection is how that happens in practice.
Why models miss it
Performance models are trained on booked accounts, because only they have outcomes.
- The model learns risk among acceptors. Within each segment that may be accurate.
- It can't learn who didn't accept. Non-acceptors vanish from the training data.
- If acceptance changes, the training population no longer matches the population the model is applied to.
What changes acceptance:
- Your own price changes.
- A competitor's price changes — which you don't control and may not see.
- A new channel, per our channel analysis, where comparison is easy and acceptance is more selective.
- Market conditions — per our overlays analysis, when competitors tighten, applicants you'd normally lose start accepting.
The last case cuts the other way, and it's worth noticing. When the market tightens, your acceptance rate among strong applicants can rise, because their other offers disappeared — so your booked mix improves in a downturn for reasons that have nothing to do with your underwriting.
Measuring it
It's measurable with data most lenders already produce and don't keep.
- Record every approved offer — terms, price, segment, predicted risk — and whether it was taken. Per our audit trail analysis, this is the record most commonly discarded.
- Compare predicted risk for acceptors and non-acceptors within each segment. A gap is the selection, measured.
- Track take-up by segment, channel, and offered price over time.
- Where permitted, look at non-acceptors' later bureau performance — how the people who walked away actually fared elsewhere.
- Randomize price within a band for a sample, per our testing analysis, and compare both take-up and booked performance.
Item five is the only clean measure of price effects. Per our measurement analysis, comparing prices across segments confuses price with everything else that differs between them. Randomization within a band is the only way to see what price alone does to who signs.
The offer data nobody keeps
Many systems store booked accounts in full and approved-but-not-taken offers barely at all. Without the offers that were declined, acceptance selection can't be seen, and pricing decisions are made on the approved population's risk rather than the booked one's.
Counter-offers
A specific case where selection is sharpest.
An applicant who asked for $10,000 and is offered $6,000 — or offered a higher rate than they expected — faces a choice the full-offer applicant doesn't.
- Applicants who accept a reduced offer may disproportionately be those with no other option.
- The shortfall between what they needed and what they got may be filled elsewhere, per our exposure analysis — adding obligations the file won't show for months.
- Counter-offer acceptors should be tracked as their own segment, since their performance can differ from both full approvals and what the model predicted.
Per our line analysis, a counter-offer is often the right decision — but the decision is only right if you know how the people who accept it perform.
Managing it
- Price for the booked population, using expected loss adjusted for who actually accepts, not for who you approved.
- Compete on the segments you want — a slightly lower price for strong applicants may improve the booked mix more than it costs in margin.
- Watch take-up as a risk indicator. A sudden rise in acceptance among a segment can mean competitors withdrew — or that your offer became the only one for people nobody else wants.
- Refresh models with acceptance in mind, per our validation analysis: check whether the booked population the model was trained on still looks like the one it's scoring.
- Separate "approved" and "booked" in reporting, so the gap is on the dashboard.
The second point is counterintuitive and frequently right. Per our pricing analysis, fixed costs dominate small-loan pricing — a lender that wins more low-risk applicants spreads those costs over a better book, which can matter more than the margin given up.
The fairness dimension
Acceptance selection has a distributional side worth stating.
- Applicants with fewer alternatives accept worse terms — per our monoculture analysis, often because every lender reads their file the same way.
- Their higher acceptance makes them a larger share of any high-priced book, and the pricing that follows falls on them.
- Price testing and counter-offer policies can affect groups differently, and per our fair lending analysis, pricing outcomes by group should be monitored alongside approval outcomes.
Measuring acceptance selection isn't only a profitability exercise. It shows who ends up paying the highest prices and why, which is information a lender should have regardless of what it decides to do with it.
See the book you actually signed
HL Hunt AI Underwriting records every approved offer alongside booked accounts, reports take-up by segment, channel, and price, compares predicted risk for acceptors and non-acceptors, and supports randomized price testing within bands.
Frequently asked questions
The change in portfolio risk caused by which approved applicants accept. Those with better alternatives decline more often, so the booked book can be riskier than the approved one.
Applicants with alternatives are most price-sensitive and walk away first, leaving a higher share of those without options.
They're trained on booked accounts. Non-acceptors have no outcomes, so when acceptance shifts, the model's training population no longer matches.
Record every approved offer, compare acceptors with non-acceptors by segment and price, and randomize price within a band for a clean measure.
Key takeaways
- The applicant's decision to accept is a second selection step, and it isn't random.
- On the worked example, a correct model still produces a 5.0% book from 4.5% approvals.
- A price rise pushes out low-risk applicants first, so risk rises with price and volume falls.
- Keep every approved offer; without non-acceptors the effect is invisible.
- Randomizing price within a band is the only clean way to see what price does to who signs.
- Report approved and booked separately, so the gap is visible and priced.
Price for who signs, not who you approved
Get started with HL Hunt AI Underwriting for offer-level records, take-up analytics, and pricing tests that show the effect of price on the book you actually hold.
This guide is educational and does not constitute legal, compliance, or model risk advice. Worked figures are stylized illustrations. Rules governing risk-based pricing, price testing, counter-offers, adverse action notices, the use of bureau data on applicants who did not book, and fair lending monitoring vary by product and jurisdiction. Consult your compliance and legal functions before testing or changing pricing.