Income Estimation: What a Noisy Number Does to a DTI Rule | HL Hunt
Income Estimation: What a Noisy Number Does to a DTI Rule
An applicant earns $4,000 and pays $1,700 a month in debt, a true DTI of 42.5% under a 45% ceiling. If their income is estimated with a typical 20% error, they have a 38.8% chance of being declined. Across a book, the same error wrongly declines 13.0% of qualifiers and wrongly approves 28.7% of applicants who exceed the ceiling. The rule is precise. The number feeding it is not.
What you'll learn
One applicant, one threshold, one estimate
An applicant earns $4,000 a month and carries $1,700 of monthly debt payments. Their true debt-to-income ratio is 42.5%, under a 45% policy ceiling. They qualify.
The lender does not see the true income. It sees an estimate, from a bureau income model, a stated figure, or a reading of bank deposits, and every estimate has error. To fail the ceiling, the estimate only needs to come in below $3,778, which is 5.6% under the truth.
| Typical estimation error | Chance this qualifying applicant is declined |
|---|---|
| 10% | 28.4% |
| 20% | 38.8% |
| 30% | 42.4% |
Even a good income estimate makes this applicant's decision close to a coin toss. The policy says 45%. The decision actually being made is whether a noisy number lands on one side of 45%, and for anyone near the line those are different questions.
What the error does across a book
Apply the same logic to a stylized applicant population in which 20.8% are truly above the ceiling. Treat the estimate as the truth and apply the rule:
| Estimation error | Qualifiers wrongly declined | Non-qualifiers wrongly approved |
|---|---|---|
| 10% | 6.4% | 18.5% |
| 20% | 13.0% | 28.7% |
| 30% | 18.2% | 34.8% |
Both errors are concentrated near the threshold, which is where the rule was supposed to do its work. An applicant at 20% DTI is safe under any plausible error. An applicant at 43% or 47% is decided mostly by noise. Per our boundary analysis, the population just either side of a cut-off is always the one where a decision rule is least reliable, and income estimation widens that zone.
The wrongful approvals are the error lenders talk about less, and on these figures they are the larger one. More than a quarter of applicants who truly exceed the ceiling pass it, which means the DTI rule is providing much less protection than the policy document implies.
Who bears the error
Estimation error is not the same for everyone, and that is where an accuracy problem becomes a fairness one.
- Salaried income with a single employer and regular deposits is estimated well.
- Hourly, variable, gig and multi-source income is estimated badly. Per our irregular income analysis, the same annual total can produce very different monthly readings.
- Cash income and recently changed income are estimated worst of all, because the data the model reads lags or misses them.
So the applicants with the noisiest estimates face the highest chance of being wrongly declined near the threshold. Per our proxy analysis, a model input whose error varies systematically by group can produce disparate outcomes without any variable in the model referring to the group, and income estimation error is a clear example. Per our legibility analysis, the households whose finances are hardest for institutions to read are the ones that pay for the difficulty.
A declined applicant is owed an accurate reason
An applicant declined on an estimated DTI is told the reason was debt relative to income. If the estimate was wrong, the reason is wrong too, and the applicant has no way to know that the figure used was not their income. Per our adverse action analysis, the reasons given must reflect the factors actually used, which raises the question of whether a decision on an unverified estimate near the threshold should have been made at all.
A precise rule on a noisy number inherits all of the noise
At a 20% estimation error, a 45% DTI ceiling wrongly declines 13.0% of qualifiers and wrongly approves 28.7% of non-qualifiers. HL Hunt AI Underwriting carries the uncertainty of every income estimate into the decision instead of treating the estimate as the truth.
Use the estimate to decide who to verify
The fix is not a better model alone. It is using the estimate for what it can do, which is sort applicants into clear cases and unclear ones.
With a 20% estimation error, treat any estimated DTI between 36.8% and 55.0% as uncertain and verify income for those applicants only. Decide the rest on the estimate.
| Approach | Applicants verified | Qualifiers wrongly declined | Non-qualifiers wrongly approved |
|---|---|---|---|
| Estimate treated as truth | 0% | 13.0% | 28.7% |
| Verify the uncertain band | 36.5% | 2.7% | 6.8% |
Verifying about a third of applicants removes roughly four-fifths of both errors. The other two-thirds are far enough from the threshold that the estimate is good enough, and per our verification analysis spending verification effort on them adds friction without changing decisions.
The friction is the honest cost. Per our stipulations analysis, every verification request loses some applicants, and the ones with alternatives leave first. Targeting verification at the uncertain band rather than at everyone is what keeps that cost proportionate, and per our review economics analysis it is the same principle as sending to review only the cases where a second look can change the decision.
Building it properly
- Measure the estimator's error against verified income on a sample, by income type, not just overall.
- Set the uncertainty band from the measured error, wider for the income types the model reads badly.
- Verify inside the band; decide outside it. Per our cash flow analysis, bank data is often the cheapest verification for the irregular earners who need it most.
- Never decline on an estimate alone near the threshold. Offer verification as a route to approval instead.
- Test outcomes by income type. Wrongful decline rates by group are the fairness check that matters here.
- Record which income figure each decision used, estimated, stated or verified, so reasons and later reviews are accurate.
- Build estimates from data as of the decision date. Per our point-in-time analysis, an estimator validated on later data will look more accurate than it is.
Item four is the policy change with the largest effect for applicants. A decline near the threshold on an unverified estimate is, on these figures, wrong more than a third of the time, and offering verification instead converts many of those wrongful declines into approvals at the cost of one request. Per our counteroffer analysis, the same logic applies to any structure the lender offers in place of the amount requested.
Item one is the measurement that most lenders have not made. An income estimator is usually validated by a vendor on an aggregate accuracy figure, and the figure that matters for a decision rule is the error for applicants near the threshold, broken down by how they earn. Per our small numbers analysis, that sample has to be large enough to say something about each income type, which is a reason to build it deliberately rather than wait for it to accumulate.
Verify the uncertain third and decide the rest
Verifying income for the 36.5% of applicants in the uncertainty band removes about four-fifths of wrongful declines and approvals. HL Hunt AI Underwriting sets the band from measured estimator error by income type and routes only those applicants to verification.
Frequently asked questions
Near the threshold it becomes close to a coin toss. A qualifying applicant at 42.5% DTI has a 38.8% chance of being declined at 20% estimation error.
At a 20% estimation error, about 13.0% of qualifiers are wrongly declined and 28.7% of non-qualifiers wrongly approved.
Irregular, gig, multi-source and cash earners, whose income is estimated worst. That makes estimation error a fairness issue, not only an accuracy one.
As triage: verify only the uncertain band around the threshold. Verifying 36.5% of applicants cut wrongful declines from 13.0% to 2.7%.
Key takeaways
- A qualifying applicant at 42.5% DTI fails a 45% ceiling if income is estimated just 5.6% low.
- At a 20% estimation error, that applicant has a 38.8% chance of a wrongful decline.
- Across a stylized book, a 20% error wrongly declines 13.0% of qualifiers and wrongly approves 28.7% of non-qualifiers.
- Both errors concentrate near the threshold, which is where the rule was supposed to do its work.
- Irregular, gig, multi-source and cash earners are estimated worst, so they bear most of the wrongful declines.
- A reason code of debt relative to income may rest on a figure that was not the applicant's income.
- Verifying only the uncertain band, about 36.5% of applicants, cut wrongful declines to 2.7% and approvals to 6.8%.
- Near the threshold, offer verification as a route to approval rather than declining on an estimate.
This guide is educational and does not constitute legal, compliance, or financial advice. All figures are stylized illustrations using an assumed applicant population, DTI distribution, policy ceiling, and lognormal estimation error, and do not describe any particular model, lender, or portfolio. Real estimation error is neither symmetric nor uniform and varies by data source and income type. Ability-to-repay requirements, fair lending obligations, adverse action reason requirements, and permissible uses of consumer and bank data are governed by federal and state law and supervisory guidance that vary by product and institution type; consult qualified counsel before relying on any approach described here.