How Long a Failure Counts: The Memory Problem in Credit Records

What a missed payment says about someone fades as the years pass and clean behaviour accumulates. The record of it doesn't fade at all. It sits on the file in the same form for a fixed period — for most negative items in the US, in the region of seven years — and then disappears. Relevance declines like a slope; the record behaves like a step. Scoring models already smooth over that mismatch by weighting recent information more heavily. So the cliff matters least where recency is modelled, and most in the landlord's screen, the manual review, and the blunt policy rule — the uses least equipped to judge how old the item is.

Slope and step

Years since the delinquency	What it plausibly tells you	How the record presents it
0–1	A lot — the difficulty may be ongoing	Present
2–3	Less, if clean behaviour followed	Present
4–5	Considerably less	Present
6–7	Little, after years of clean history	Present, in the same form
After the period	Little	Gone entirely

The middle column slopes; the right column steps. For six and a half years the record says the same thing, and then on one day it says nothing.

Per our reporting analysis, the retention period is set in rules rather than derived from evidence about when information stops being useful, and the exact periods vary by item type and have changed over time. This report doesn't turn on the precise length. It turns on the shape — full weight, then zero — and on who is affected by that shape.

Relevance fades like a slope. The record behaves like a step. The mismatch lands on whoever isn't modelling the difference.

Why relevance fades

Three reasons an old delinquency is weaker evidence than a recent one.

• Circumstances change. A missed payment frequently marks an event — a job loss, an illness, a separation — rather than a trait. Per our default analysis, the record captures the lender's accounting position, not the household's situation, and the situation is what moves on.
• Subsequent behaviour is better evidence. Five years of on-time payments after a delinquency is a longer and more recent observation than the delinquency itself.
• Survivorship. Someone whose old delinquency is followed by a long clean history has already demonstrated recovery — which the old item, read alone, doesn't show.

None of this means old information is worthless. A history of repeated failures carries more weight than a single one, and some patterns stay informative for a long time. The claim is narrower: a single old item, followed by years of clean behaviour, says much less than its unchanged presence on the file implies.

What models already do

The part that complicates the usual criticism of retention periods.

Credit scoring models generally weight recent information more heavily than old information. Recency of the most recent delinquency is a standard input, and the effect of a delinquency on a score typically diminishes as it ages — per our attribute analysis, "months since most recent delinquency" is exactly the kind of attribute built to capture this.

Which means that inside a well-built model, the slope is approximated even though the record is a step. A six-year-old item contributes much less than a six-month-old one, and its disappearance produces a modest change rather than a transformation.

That's why the common argument — "the seven-year period punishes people too long" — is only half right. For modelled decisions, the model has already done much of the forgetting. The question is what happens to the file everywhere else.

Where the step lands

The central finding: much of the use of credit files happens outside models, and those uses tend to read items as present or absent.

Use	How an old item is typically read
Scoring model	Down-weighted by age
Blunt policy rules — "no charge-offs on file"	Present or absent
Tenant screening	Frequently present or absent
Manual underwriting review	Depends on the reviewer
Employment screening, where permitted	Frequently present or absent
Specialty reports	Varies by provider

Row two is the one inside lenders themselves. Per our policy analysis, knock-out rules sit alongside models, and a rule written as "decline any applicant with a charge-off" treats a charge-off from six years ago identically to one from last month — undoing, at the policy layer, the recency weighting the model just applied.

Row three is where the consequences are largest. Per our specialty report analysis and our renting guide, housing decisions are frequently made by people or simple systems reading a report without a model — and a six-year-old collection can cost someone a flat in a way it would never cost them a loan.

Which gives the finding: the retention cliff is least consequential where recency is already handled, and most consequential in exactly the uses least able to handle it. Arguments about the right length of the period are mostly arguments about the wrong layer.

Same period, different failures

The second property of the step: it's largely the same length regardless of what happened.

• A small collection and a large default persist on broadly similar timelines.
• A single late payment and a pattern of them age off on each item's own clock, but a one-off event sits for as long as a symptom of something chronic.
• An account disputed and never resolved can sit alongside one that was simply unpaid.

There are exceptions and they've been growing. Per our medical debt analysis, particular categories have had their treatment changed — which is itself evidence that uniform retention was never derived from what the information means, since carving out categories one at a time is what happens when a uniform rule meets cases it fits badly.

Per our framing analysis, a record that treats a $90 collection and a $40,000 default as equivalent for the same number of years is making a statement about failure as a category rather than about risk, and that's a moral position embedded in a technical rule.

The missing information: what happened next

The most useful single improvement, and it's about data that largely exists.

A negative item records the failure. It doesn't foreground what happened afterwards — whether the debt was resolved, whether behaviour recovered, how quickly.

What a reader of the file would want to know about an old delinquency:

1. How long ago — present, but not always prominent.
2. Whether it was resolved — settled, paid, or left.
3. What the record looks like since — the length of the clean run.
4. Whether it was an isolated event or part of a pattern.
5. Whether it was disputed, per our errors analysis.

Items two and three are the recovery evidence, and they're the information that makes an old failure interpretable. Per our coverage analysis, the system is built to record failure in detail and recovery by implication — which is the same asymmetry that report identified in what gets recorded at all, reappearing in how records age.

Per our second-chance analysis, the same question arises for criminal records, where the debate over whether and when old records should stop counting has been more explicit. Credit records are having the same debate with less vocabulary for it.

What fading properly would look like

The constructive side. The principle: the record's weight should decline at roughly the rate its meaning does.

• Present age and subsequent history together — an old item displayed with the clean run since it, so no reader sees one without the other.
• Replace blunt presence rules with recency-weighted ones in lender policy — "no charge-off in the last 24 months" rather than "no charge-off on file."
• Recognize resolution explicitly, so a resolved item reads differently from an abandoned one.
• Guidance for non-model users — landlords especially — on how to read an item's age.
• Consider severity-scaled retention, which is a policy question with real trade-offs rather than a technical one.

The second is available to any lender today and costs nothing. A knock-out rule with a recency window aligns the policy layer with the model it sits beside — and per our override analysis, if reviewers are regularly overriding a blunt rule for old items, that's the evidence it's misspecified.

And per our monoculture analysis, a lender that reads old items correctly while competitors apply blunt rules is identifying applicants the market is excluding for stale reasons — which is an underwriting advantage as well as a fairness improvement.

The strongest objections

"Old information still predicts something." Agreed, and the report doesn't claim otherwise. The claim is that its predictive value declines and the record's weight doesn't — so the error is in the shape, not in retaining information at all. A slope that reaches near-zero is different from a step that reaches zero, and both differ from a flat line that suddenly drops.

"Shorter periods would reduce the information available to lenders." Partly, which is why this report doesn't recommend a shorter period. It recommends that the information be read with its age, which preserves what lenders need while removing the full-weight treatment of stale items by users who aren't modelling age at all.

"You haven't shown how much old items actually matter outside models." Correct. The report argues from how screens and rules are typically structured, not from measured outcomes, and the size of the effect in housing specifically is the biggest open question here. The discontinuity test below would answer it.

Testable implications

1. People whose only negative item has just aged off should perform almost identically to people whose item is a month from aging off — a discontinuity test of whether the cliff corresponds to any change in actual risk.
2. Approval outcomes in non-model uses should jump at the aging-off date much more than model scores do.
3. An old delinquency followed by a long clean run should predict little beyond the clean run itself, controlling for current behaviour.
4. Lender policy rules without recency windows should generate overrides concentrated on old items.
5. Resolved and unresolved old items should predict differently, which would justify presenting them differently.
6. Severity should affect the rate at which predictive value decays — large defaults staying informative longer than small collections.

The first is the decisive one and the design is clean. Compare people whose single negative item disappeared last month with people whose item disappears next month. They're nearly identical in every respect except one day on a calendar. If their subsequent performance is the same, the step marks no change in risk — and every decision that changed on that date was responding to the rule rather than to the person.

The conclusion we'd hold: credit records remember failure at full strength for a fixed term and then forget it completely, while its meaning fades steadily. Models compensate; many other users don't. The fix is not a shorter memory but a more honest one — records that show how old a failure is and what happened after it, read by rules that know the difference.

Frequently asked questions

---

This report presents an analytical framework and the authors' interpretation; it is not legal advice. Retention periods for credit information vary by item type and jurisdiction and have changed over time, and the rules governing use of credit reports in housing and employment decisions differ by jurisdiction. The effects described are argued from structure rather than measured; the implications identified as testable are hypotheses.