What Counts as Default: A Convention Mistaken for a Fact

Ask when a loan defaulted and you'll get a date. Ask why that date, and the answer is a threshold somebody chose — thirty days, ninety days, a charge-off decision, a legal event. Each produces a different count from the same accounts, and each was adopted for accounting or operational reasons rather than because it identifies anything about the borrower. That matters more than it sounds, because every published loss rate, every model's target variable, and every comparison between lenders rests on a convention that isn't standardized and isn't measuring what people think it measures. Default records when a lender's position changed. It does not record when a household got into trouble, and the two are months apart.

The conventions in use

Convention	Triggered by	Chosen because
30 days past due	A missed billing cycle	It's one cycle
60 / 90 days	Two or three cycles	Multiples of the same cycle
Charge-off	A lender's decision to remove the asset	Accounting and supervisory practice
Bankruptcy	A legal filing	An external event that's unambiguous
Contractual default	Any breach of terms	Frequently much broader than non-payment
Non-accrual	Ceasing to recognize interest	Accounting

Six definitions, all in simultaneous use, all called default in different contexts. They don't nest cleanly and they don't agree.

The billing-cycle basis of the first two is worth pausing on. A thirty-day threshold is thirty days because that's how long a statement cycle happens to be — it isn't a finding about when borrowers become unlikely to recover. Had billing evolved on a fortnightly or quarterly cycle, the industry's delinquency buckets would sit at different places and everyone would treat those as equally natural.

And the fifth row is the widest gap between the word and its use. Contractual default frequently includes events that have nothing to do with missing a payment — a covenant breach, a changed circumstance — which per our covenant analysis gives a lender decision rights while the borrower is paying perfectly well.

Thirty days is thirty days because a statement cycle is a month. It is not a finding about borrowers.

Where they came from

The conventions were adopted to serve three constituencies, none of which was trying to describe a borrower's situation:

• Accounting. An institution has to say what its assets are worth and when to stop recognizing income, and charge-off timing is largely a response to that requirement.
• Supervision. Regulators need comparable-looking figures across institutions, which pushes toward standardized buckets whether or not they mean anything.
• Operations. Collections work is organized by stage, and stages need boundaries — the ones in our cure analysis exist so that treatment can be assigned.

Each is a legitimate purpose and none of them is measurement of borrower distress. The conventions were then borrowed for that purpose by everyone downstream — modellers, researchers, journalists, policymakers — because they were the numbers that existed.

This is a familiar shape. Our attribute analysis found the same thing one layer down: values constructed for one purpose get consumed for another, and the construction decisions travel invisibly with them. Default definitions are that problem at the level of the outcome variable rather than the inputs.

Why lenders can't be compared

The practical consequence, and it undermines a great deal of published analysis.

Both parts of a rate are discretionary:

Choice	Effect on the reported rate
Threshold used	Direct, and large
Whether cured accounts count	Very large — see below
Treatment of arrangements	Accounts in a plan may be excluded entirely
Charge-off timing	Moves accounts out of the numerator
Denominator	Accounts, balances, or originations — all different
Vintage or point-in-time	A growing book looks better point-in-time

Two lenders with identical portfolios and identical borrower behaviour can report materially different rates through these choices alone, none of which is improper.

The last row deserves emphasis because it's the least understood and the most exploited. A point-in-time delinquency rate on a rapidly growing book is diluted by new accounts that haven't had time to go bad — so growth flatters the number mechanically. The vintage analysis in our benchmarking analysis is the correction, and published figures rarely use it.

Which yields a rule worth applying to any comparison: if you cannot state both parties' definitions, you are comparing conventions rather than performance. That rules out most cross-lender comparison as it's normally conducted.

What the definitions measure

The central claim: these conventions track the lender's position, and the borrower's situation only incidentally.

Consider the sequence a household actually goes through, against what gets recorded:

What happens	Recorded as
Income falls or an expense hits	Nothing
Savings depleted	Nothing
Borrowing to cover the gap	Utilization rising — visible but not distress
Prioritizing among obligations	Nothing, until one is missed
First missed payment	30 days past due, a cycle later
Sustained non-payment	60, then 90
Lender removes the asset	Charge-off

The first four rows are where the household's situation actually changed, and none of them is recorded. By the time a thirty-day marker appears, per our liquidity analysis and hierarchy analysis, the household has typically been managing a shortfall for months — depleting reserves, borrowing, and choosing which obligations to protect.

Which produces the finding this report would defend hardest. Default statistics are a lagging record of an accounting decision, not a measure of household distress, and the lag is long and varies by lender. Anyone using delinquency rates as an indicator of how households are doing is using a series that responds months late and is shaped by conventions unrelated to the question.

And it explains why our early warning analysis finds pre-delinquency signals valuable: they aren't early warnings of default, they're the first observations of a situation that started well before.

The cure problem

The single largest source of confusion, and it's arithmetic rather than interpretation.

Most accounts that reach an early delinquency stage recover. Per our cure analysis, natural cure rates at early stages are high — most people who miss a payment pay it.

So "30 days past due" is not a synonym for a bad account. It's a population that is mostly fine, containing a minority that isn't. Consequences:

• An early-stage delinquency rate overstates trouble by including a large majority who will recover.
• Later stages are more informative because the cure rate falls with age.
• Whether cures are netted out changes the number dramatically, and different reporters do it differently.
• A "roll rate" — the share moving from one stage to the next — is more informative than a level, and is published far less often.

The modelling consequence is the one with teeth. A model trained on thirty-day delinquency is trained on a target that is mostly noise, since most of the positive cases resolve. It will find whatever predicts a temporary missed payment — which per our cure analysis is substantially about forgetting and timing rather than about capacity — and that is a different thing from what predicts loss.

What models learn

Per our validation analysis, the outcome definition is the first question a validation should ask and usually the last one it does.

What the choice determines:

• What the model actually predicts. Ninety-day delinquency and charge-off are different events with different drivers.
• Whether the performance window was long enough. A charge-off target needs a much longer window than a delinquency target, and per our vendor analysis an immature window understates losses systematically.
• How cures are handled — an account that went delinquent and recovered is a positive case under one definition and a negative under another.
• Whether fraud is mixed in. Per our attribution analysis, a blended outcome teaches a blended and unstable relationship.
• What a purchased model means. A vendor's definition is unlikely to match yours, which produces calibration error attributed to the model.

The last is the practical trap. Two models both described as predicting default can be predicting genuinely different events, and comparing them on discrimination statistics compares their skill at different tasks. That's a validation finding that requires reading the definition rather than the performance table.

And it connects to what a lender actually needs. Most lenders care about loss — money not recovered — which is not any of the six conventions. Per our policy analysis, an approval decision is an economic decision about expected loss, and a model predicting a delinquency threshold is answering a proxy question whose relationship to the economic one is assumed rather than measured.

What consumers are told

The definitions reach households too, and the mismatch produces real confusion.

• "Charged off" sounds like the debt is gone. It isn't — per our debt sale analysis, charge-off is an accounting event after which the obligation continues and is frequently sold.
• "Default" on a report and default in a contract are different, and the second can occur without the first.
• A cured delinquency stays on the file, so recovering doesn't remove the record — per our timing analysis, the mark and its consequences persist independently of what happened afterwards.
• Different products default at different points, so experience with one is a poor guide to another.

The first is the one that causes the most harm. A household told an account was charged off frequently believes the matter is closed, and it isn't — collection continues, the balance stands, and per our old debt guide a payment made years later on the assumption that it was a formality can restart an enforcement period.

Which is a disclosure problem of the kind our disclosure analysis describes: a term of art borrowed from accounting is used in consumer communication where its ordinary meaning is misleading.

What a better measure would need

Two constructive directions, and being honest that neither is free.

For comparison purposes: publish the definition alongside the number. Not a new standard — a requirement that a stated delinquency or loss rate carry its threshold, its cure treatment, its denominator, and its vintage basis. This is cheap and it converts an incomparable figure into a comparable one, and it's the same functional-standard remedy our complexity analysis arrives at for prices.

For distress measurement: stop using default statistics for it. The series that would actually indicate household difficulty are earlier and mostly exist:

• Utilization trajectories, which move before payments are missed.
• Balance and overdraft behaviour, per our liquidity analysis.
• Hardship arrangement uptake, per our arrangements guide — though this is depressed by the framing our moral analysis describes, so it undercounts.
• Minimum-payment behaviour on revolving accounts.
• Roll rates rather than levels.

The honest limitation: these are held by lenders and mostly not published, and standardizing them would recreate the same convention problem one layer earlier. There is no definition-free measure available — the argument is for choosing conventions deliberately and stating them, not for the existence of an objective one.

Testable implications

1. Reported delinquency rates should vary across lenders more than underlying performance does, with the residual attributable to definitional choices — checkable where portfolio composition can be controlled.
2. Point-in-time rates should fall during rapid book growth independently of borrower behaviour, and rise when growth stops.
3. Models trained on early-stage delinquency should underperform models trained on loss when evaluated against loss, by a margin reflecting the cure rate.
4. The lag between household distress and recorded delinquency should be measurable and long, using cash-flow data as the earlier series.
5. Consumers should systematically misunderstand charge-off, testable directly and relevant to disclosure design.
6. Roll rates should predict subsequent loss better than delinquency levels, which would justify publishing them.

The fourth is the one that would matter most and is now feasible. Connected transaction data makes it possible to observe the depletion sequence directly and compare it to the date a delinquency was eventually recorded — which would put a number on how late the standard series is. Our affordability analysis describes the same data being used prospectively; this would use it to calibrate what the existing measures miss.

The conclusion we'd hold: default is a bookkeeping event that the industry, its regulators, and its observers have collectively agreed to treat as a fact about borrowers. It isn't one, the conventions don't agree, and almost every comparison built on them is comparing something other than what it claims.

Frequently asked questions

---

This report presents an analytical framework and the authors' interpretation; it is not accounting, legal, or regulatory advice. Charge-off timing, non-accrual treatment, delinquency classification, and disclosure requirements are governed by accounting standards and supervisory guidance that vary by institution type and jurisdiction and are described here only in general terms. The implications identified as testable are hypotheses.