Affordability: Whether They Can Pay, Not Whether They Will

A credit model predicts default from history — it answers whether someone has met obligations and is likely to keep doing so. An affordability assessment asks something different: whether this specific payment fits in the money that remains after everything else they're committed to. Those questions come apart constantly. An applicant with an excellent file can be fully committed, and one with a thin file can have substantial room. A model optimized for the first is frequently treated as answering the second, and it doesn't — which is a modelling error before it's a compliance one.

Two different questions

	Creditworthiness	Affordability
Asks	Have they repaid, and will they?	Does the money exist?
Predicted from	History and behaviour	Income and commitments
Data source	Credit file, mostly	Income and cash flow, mostly
Thin file means	Hard to assess	No particular difficulty
Excellent file means	Low predicted default	Nothing in particular

The bottom two rows are where the practical value sits.

An excellent credit file tells you nothing about capacity. Someone who has always paid everything may be paying everything and have nothing left — and per our liquidity analysis, they may have substantial net worth and no accessible cash, which the file also doesn't show.

A thin file is a creditworthiness problem and not an affordability problem. Per our thin-file analysis, these applicants are hard to assess on history and perfectly assessable on capacity — which means affordability assessment is one of the few tools that works better on thin files than on thick ones. That's the opposite of most underwriting inputs and it's worth exploiting rather than treating as an obligation.

Why the ratio is weak

Debt-to-income persists because it's cheap, and it's a screen rather than a measure.

Two problems, and the first is arithmetic:

The same ratio means different things at different incomes.

Monthly income	Obligations at 40%	Remaining
$2,400	$960	$1,440
$9,000	$3,600	$5,400

Identical ratios, and one household has nearly four times the absolute room. Living costs don't scale proportionally with income, so the ratio systematically overstates capacity at low incomes and understates it at high ones.

The second problem is what feeds it. Obligations taken from a credit file omit:

• Rent — frequently the largest single outgoing, and absent unless the tenant is in a rent reporting program per our rent reporting guide.
• Childcare, which can exceed a mortgage payment.
• Support payments.
• Utilities and insurance, per our reporting analysis.
• Obligations to family, which are real and invisible.
• Products outside conventional reporting, per our definitional analysis — a growing category.

For many households the omitted items exceed the captured ones, which makes a bureau-derived ratio a measure of the smaller half of their commitments.

Residual income

The better measure: income, less all committed outgoings, less the proposed payment. What's left.

Why it's stronger:

• It's an absolute amount, so it means the same thing at every income level.
• It can be compared to a realistic estimate of what a household of that size and location needs.
• It handles the new payment directly rather than through a ratio.
• It's interpretable. "$340 a month remaining for a family of four" is a statement anyone can evaluate; "38% DTI" isn't.

What it requires:

1. Reliable income, at the level discussed below.
2. Committed outgoings, including the items a file omits.
3. A benchmark for basic living costs by household size and location.
4. A threshold below which you won't lend.

The third is where operations struggle, and the honest position is that any benchmark is an estimate. Published cost-of-living data, statistical benchmarks, and applicant-declared expenses all have weaknesses, and using them requires documenting the choice per our validation analysis. An imperfect benchmark applied consistently is still substantially better than a ratio that ignores the question.

Establishing income

The input that most determines the answer, and the one most often taken at face value.

What to establish beyond the figure, per our verification guide:

• Reliability, not just amount. Per our irregular income analysis, a variable income assessed at its average overstates what can be relied on in any particular month — the right figure for affordability is closer to a low percentile than a mean.
• Composition — base, variable, and one-time components have different durability.
• Household versus individual, and whether a second income is genuinely available.
• Timing, since per our timing analysis an income that arrives after obligations fall due creates a gap regardless of the annual total.
• Duration — recent, expected to continue.

The variable income point deserves the most weight because it's where the largest errors happen. Someone earning between $2,100 and $4,300 monthly does not have $3,200 available every month, and an assessment using the average has approved a payment that fails in the bad months — which are the months when everything else is also difficult.

The obligations the file doesn't hold

Three ways to capture them, with a real trade-off between them:

Approach	Accuracy	Cost
Ask the applicant	Variable, and optimistic	High — abandonment
Statistical estimate	Population-average	Near zero
Connected transaction data	High	Low, with consent

Asking is the default and the worst option on both dimensions. Applicants under-report expenses — not usually dishonestly, but because nobody knows their own outgoings precisely — and every question costs completions.

Connected data is the strongest option and it's the mechanism our distribution analysis identifies as one of the few genuine access expansions: it observes actual outgoings, requires no assembly by the applicant, and can distinguish committed from discretionary spending.

What connected data shows that nothing else does:

• Actual recurring outgoings, including everything a bureau omits.
• Income timing and variability rather than a stated figure.
• Existing pressure signals — overdrafts, failed payments, balance troughs.
• Balance behaviour, which per our liquidity analysis predicts better than many bureau attributes.

And a caution: distinguishing committed from discretionary outgoings from transaction data requires judgment encoded in rules, and those rules are model inputs that need documenting and validating like any other, per our attribute analysis. A classification that treats a genuine commitment as discretionary overstates capacity systematically.

Testing against a bad month

The check that distinguishes a serious assessment from an arithmetic exercise.

A payment that fits only in a good month doesn't fit. What to test:

• Income at a low percentile rather than the average, for variable earners.
• A rate increase, on variable-rate products.
• An expense shock of a realistic size — per our savings analysis, most households face one periodically.
• The payment after any promotional period ends, which is the version they'll pay for most of the term.
• Whether residual income stays above the threshold in each case.

The fourth is the one most often skipped and most consequential. Assessing affordability against an introductory payment assesses a payment the borrower will make briefly, and per our prediction analysis the borrower is systematically optimistic about their position when the higher payment arrives. The lender has no such excuse, since the schedule is known at underwriting.

How strict to be is a policy decision with a genuine trade-off — a very conservative stress test excludes people who would have managed, which per our rationing analysis is an invisible cost. Testing against something realistic rather than something catastrophic is the defensible position, and the level should be documented rather than inherited.

The abandonment cost

The cost of assessment that doesn't appear in any report.

Every question asked loses applicants, and the ones who abandon leave no record of what they'd have been — the Family A structure from our measurement analysis. Which means:

• The cost of a longer application is invisible while the benefit of better assessment is measurable.
• So application requirements drift heavier over time, with nothing pushing back.
• Abandonment is not random. It falls hardest on people with less time, less document access, and more complicated finances — which correlates with the population affordability assessment is meant to protect.

What to do about it:

1. Measure abandonment by step, so you know what each question costs.
2. Use connected data instead of asking, wherever possible.
3. Assess progressively — light assessment for small amounts, fuller for larger, proportionate to what's at stake.
4. Ask only what changes a decision. A field that never moves an outcome is pure abandonment cost.
5. Randomize a relaxed requirement for a share of applicants and measure completion against realized performance — the holdout our measurement analysis recommends, and the only way to see this cost.

The fourth is worth auditing directly. Most applications contain fields nobody uses, added at some point for a reason nobody remembers, each costing completions every day.

Building the assessment

1. Define affordability separately from the credit decision, with its own criteria.
2. Establish reliable income, at a low percentile where variable.
3. Capture obligations including what the file omits, by connection where possible.
4. Compute residual income, not only a ratio.
5. Set and document a threshold, with the benchmark it rests on.
6. Stress test against realistic variation and the post-promotional payment.
7. Measure abandonment at every step.
8. Document the methodology, since it will be examined.
9. Validate it against outcomes — do affordability failures actually predict distress?
10. Monitor by segment, per our governance framework.

Item nine is the one almost nobody does and it's the only test of whether the assessment works. An affordability model that has never been validated against realized outcomes is an assumption with a procedure attached — and it can be tested the same way any other model is, by comparing predicted capacity against what actually happened.

Frequently asked questions

---

This guide is educational and does not constitute legal or compliance advice. Ability-to-repay and affordability requirements vary substantially by product type, jurisdiction, and institution, and specific rules apply to certain products that are not described here. Worked figures are stylized illustrations. Consult qualified counsel about the requirements applicable to your products, and your model risk function about validating any affordability methodology.