SystemResidential Mortgage & Home Equity Origination

Mortgage QC Automation · accuracy and the human baseline

Fannie Mae target defect rate: what your sample can and can’t see

Fannie Mae does not publish a target defect rate: each lender must set its own for its most severe defects, measured on its post-closing random sample every quarter and reset every year. This guide shows how to set one, how much a 10% sample can actually tell you, and what it cannot see, with Mortgage QC Automation's sample data as the worked example.

Updated · 14 min read

See one case, screen by screen →
Watch with transcript →

The Fannie Mae target defect rate is a number each lender sets for itself: the share of loans in its post-closing random QC sample that may carry a defect at its highest severity level, the level reserved for loans not eligible as delivered. Selling Guide D1-1-01 requires the target to be "as reasonably low as possible", measured at least quarterly, reported to management, and evaluated and reset at least annually, with the rationale written down. Fannie Mae does not publish an acceptable rate. It reserves the right to look at the one you chose and "provide input on a more appropriate target". Mortgage QC Automation keeps that random sample as the yardstick.

The hard part is not choosing the number but measuring against it. A 10% random sample gives a lender an estimate of its defect rate with an error band that, at the defect rates lenders actually see, can be wider than the rate itself, and it finds the defects only in the loans it happens to draw. Botminds' Mortgage QC Automation keeps the random sample as the lender's measure and adds a review of every other closed loan, so the two numbers can be read side by side. The method first.

The Fannie Mae target defect rate, in four requirements

Top

severity level must carry a target; others may

The level for loans ineligible as delivered

10%

random post-closing sample, full-file, is the basis (or a statistical sample)

D1-3-01

≥4×

a year: performance measured at least quarterly, reported to management

1×

a year at least: target evaluated and reset if necessary, rationale documented

Source: Fannie Mae Selling Guide D1-1-01 (04/01/2026)

What is a target defect rate in mortgage QC?

It is the lender's own ceiling for the share of loans with serious defects, against which QC performance is reported. Fannie Mae requires a "methodology for identifying, categorizing, and measuring defects and trends against an established target defect rate"; different targets may be set for different severity levels, but "at a minimum a target defect rate must be established for the lender's highest level of severity" (D1-1-01). The target is "based on the lender's post-closing random QC sample", which D1-3-01 calls "a critical measure of loan quality" and "the basis for calculating the lender's defect rate" (D1-3-01).

The monthly report carries it. Post-closing reports must include "data for the highest severity classification defect rate, benchmarked against its defined target rate, at least quarterly", alongside three months of defect trending. A target with no rationale fails the rule as surely as a missing one: the lender "must document the rationale for establishing the target rate(s)".

What is Fannie Mae's acceptable defect rate?

There is none in the Selling Guide. Fannie Mae asks for a target "as reasonably low as possible" and says it "may assess how the lender's chosen target defect rate affects Fannie Mae's risk and may provide input on a more appropriate target". The absence is deliberate: a lender's rate depends on its severity definitions, its product and channel mix and the depth of its reviews, so one industry number would reward lenient grading. Industry trackers published by QC vendors are not Fannie Mae standards and define their categories in their own ways; they do not belong in a QC plan's rationale.

The government side is different, because FHA publishes its own review results. In fiscal 2022 FHA's Loan Review System completed 27,405 underwriting reviews and rated 11.26% net material, after lenders' responses. Random selections ran at 3.40%; early payment defaults at 9.12% (HUD, FHA Quality Assurance Update, January 2023). Those are FHA's findings on FHA loans, graded on FHA's scale, and not a conventional benchmark; they are the best public evidence of how much the way loans are chosen changes the rate you see.

How a loan is chosen changes the defect rate you measure: FHA reviews, fiscal 2022

Lender self-reports (3,167 reviews)
41.27%
FHA manual selections (565)
22.30%
Lender monitoring (2,461)
11.74%
Early payment defaults (9,667)
9.12%
Risk algorithm (3,340)
5.78%
Test cases (1,068)
3.93%
Random selections (5,761)
3.40%
Share of FHA underwriting reviews rated net material, by selection reason, for Loan Review System reviews completed from 1 October 2021 to 30 September 2022. Only the random row estimates the whole book; every other row is a targeted sample.Source: HUD OLAPC, FHA Quality Assurance Update, January 2023

What is a critical defect rate?

"Critical defect rate" is the industry's name for the share of loans with a defect at the top of the severity scale, the one that makes a loan ineligible as delivered. In Fannie Mae's own language the closest term is "significant defect": a defect that "necessitates a change to the price on which the loan was acquired or result[s] in the loan being unacceptable for purchase" had the facts been known (Selling Guide D2-1-03). Your plan can call the level critical, significant or severity 1; what matters is that it lines up with that definition, because that is the level the target, the self-report and any repurchase discussion all turn on.

How do you calculate a mortgage defect rate?

Divide the loans in the post-closing random sample with at least one defect at the top severity level by the loans in the random sample reviewed, for the quarter. Three rules keep the number honest. Count loans, not findings, since a loan with two significant defects is one ineligible loan. Leave discretionary selections out of the denominator and the numerator, because they were chosen for risk; report them separately. And count the grade after rebuttal but before cure: a defect that existed at delivery was a defect, whatever was fixed later.

Take Halden Ridge Home Loans, the fictional lender whose September runs through Botminds' Mortgage QC Automation in this series. It closed 1,186 loans; its 10% random sample was 119. Because the solution also graded every loan, the true September count is known: 14 loans with a significant defect, or 1.18%. The random sample happened to contain 2 of them, a measured rate of 1.68%.

How precise is a 10% random sample?

Less than most QC reports imply. Two significant defects in 119 loans gives a 95% confidence interval, by the Wilson method, of roughly 0.5% to 5.9%: the sample cannot tell a lender running at 1% from one running at 5%. A full quarter of samples narrows it but not much: four in 357 loans gives about 0.4% to 2.9%. Those intervals are our arithmetic on the sample lender's figures, using the standard methods for a proportion (NIST/SEMATECH e-Handbook of Statistical Methods, 7.2.4).

That is the same arithmetic behind the statistical alternative to 10%. Fannie Mae allows a sample built on "a 95% confidence level with a 2% precision rate and a statistical statement of six months"; FHA, for lenders above 3,500 FHA loans a year, a stratified sample at 95% confidence and an interval of no more than 2% "based on the defect rates for FHA-insured Mortgages recently reviewed" (4000.1, V.A.3.a.iii). The required size depends on the defect rate you assume, which is why that assumption is audited.

How big a statistical QC sample must be, by the defect rate you assume

Assumed defect rateSample, large bookSix months of 1,186 a monthPer month10% sample per month
50% (no assumption)2,4011,795299119
10%864771128119
5%45642972119
3.4%31530250119
2%18818331119
1%959416119
Our arithmetic, not an agency table: n = 1.96² × p(1 − p) ÷ 0.02², with a finite-population correction for a six-month book of 7,116 loans (the sample lender's September volume times six). At low defect rates a ±2% precision band is wider than the rate itself, so the small samples it permits measure little.Source: NIST/SEMATECH e-Handbook, sample sizes for proportions

Getting the assumption wrong is an audit finding. HUD's inspector general found that loanDepot "used a flat 5 percent defect rate stated in its QC plan rather than defect rates on loans recently reviewed as HUD required", did not stratify by product type and origination source, and so performed 2,376 post-closing reviews over two years where it estimated at least 3,566 were due, about 1,190 short, on loans totaling about $329 million (HUD-OIG 2025-NY-1002). A statistical sample saves work only if its inputs are defensible, and at the defect rates in the table the saving buys a less precise number.

A ±2% precision band on a 1% defect rate does not measure the rate; it confirms the rate is small. Set the target knowing which of the two your sample can do.

How to set a target defect rate you can defend

Start from your own history on the random sample, not an industry number: the rate at the top severity level over the last four to eight quarters, with its confidence interval. Set the target below the recent rate if the trend or the rule demands improvement, and never above what Fannie Mae would see if it re-reviewed your loans. Write down why: product and channel mix, a change in severity definitions, a known problem being fixed. Then decide in advance what happens when a quarter breaches the target. A target with no consequence is a reporting line, not a control.

Expect the definitions to move the number more than the loans do. A lender that moves a defect type from moderate to significant will see its rate jump with no change in production quality. That is why D1-1-03 requires "consistent methodology and terminology, including defect categories and severity levels", and why a change to the scale needs its own note in the report. FHA's grading adds a further wrinkle: its fiscal 2022 statistics separate gross ratings at first review from net ratings after lenders mitigate, and the target belongs on one basis, stated in the plan.

What does a random QC sample miss?

Nine loans in ten, by design. The random sample estimates the rate; it finds defects only in the loans drawn. At the sample lender, a 10% sample would be expected to contain about 1.4 of September's 14 significant loans, and the other twelve or so would go to their pools without review. That is not a flaw in the rule: Fannie Mae's sample is a measurement, and it adds a discretionary sample to look for trouble. The flaw is treating a measurement as a control.

Same September, two views of the defect rate

10% random sample aloneRandom sample plus every loan reviewed before sale
Loans reviewed

10% random sample alone: 119 random (plus discretionary)

Random sample plus every loan reviewed before sale: 1,186, with the 119 random loans still re-verified in full by auditors

Significant defects seen

10% random sample alone: 2 found in the sample; the rate is estimated at 1.68%, interval about 0.5%–5.9%

Random sample plus every loan reviewed before sale: 14 found, 1.18% of the month, each with its loan and evidence

When they are seen

10% random sample alone: As the sampled reviews finish, inside a cycle of up to 90 days after the month

Random sample plus every loan reviewed before sale: As each closing package is complete, before its pool cut-off

Loans sold with a significant defect

10% random sample alone: Unknown; at least the defects outside the sample

Random sample plus every loan reviewed before sale: 0.25% after cures in the sample data, against 1.52% in May, the last month on the sample alone

What the random sample measures

10% random sample alone: The lender’s production

Random sample plus every loan reviewed before sale: The lender’s production and the agents’ accuracy on the same loans

Fictional sample data from Mortgage QC Automation, not a customer's results. The defect rate reported against the Fannie Mae target is still calculated on the random sample, as D1-3-01 requires.

The second column changes what the target is for. The random-sample rate still goes against the target, as the Guide requires. But the lender also knows the number the target was standing in for: how many loans it actually sold with a significant defect. In the sample data that fell from 1.52% in May, when only the sample was reviewed and it was reviewed after sale, to 0.25% in September after pre-sale cures. Why reviewing before the sale matters is the subject of QC before the sale.

In Mortgage QC Automation, the random sample measures the agents too

The re-verifications screen: 157 loans in the September sample, 119 random and 38 discretionary; 445 of 469 requests returned, 24 outstanding, 11 exceptions found of which 3 became significant findings. A panel reads: 97.8% of agent findings confirmed by auditors on the random sample, July to September; auditors found 2 the agents missed, both minor.
Fictional sample data. The 119 random loans are re-verified at the source and reviewed in full by an auditor, independent of the agents; comparing the two gives the agents' accuracy alongside the lender's defect rate.

The random sample as the human baseline

Any automated QC review needs a yardstick, and the rulebooks already provide one. The random sample is drawn without regard to risk, reviewed in full and re-verified at the source by people independent of production, which makes it the cleanest available judgment of what each sampled loan's defects really were. Run the agents over the same loans and compare: findings both found, findings only the auditor found, and findings only the agents raised that the auditor rejected. On the sample lender's data from July to September, auditors confirmed 97.8% of agent findings on the random sample and found two the agents missed, both minor.

That comparison is what a model risk team will ask for, and it should be reported the way the defect rate is: by quarter, with counts and not just percentages. A lender considering this can test it on its own loans before relying on it; one month side by side with a QC vendor sets out the measures. And because early payment defaults run far above random loans in FHA's data, they are the first place to look for what a sample misses; early payment defaults first explains why.

What a 10% sample reads, against every loan Mortgage QC Automation graded

The September grid of loans by grade on the QC test script screen: significant 13, moderate 33, minor only 118, clean 1,021, running now 1. A note: a 10% sample of 119 loans, reviewed by about day 75, after the pools settle, holds 2 of the 14 significant loans; the rest would have been sold.
Fictional sample data. Each square is a loan, coloured by grade; the 10% sample view dims every loan a sample would never have read.

Questions people ask

Does Fannie Mae publish a target defect rate?

No. The Selling Guide requires each lender to set its own target for its highest severity level, as reasonably low as possible, based on its post-closing random sample. Fannie Mae may assess the target and suggest a more appropriate one.

Should the discretionary sample count in the defect rate?

Not in the rate measured against the target. Discretionary selections target higher-risk loans, so including them overstates the book's rate. Report their results separately, as D1-1-03 requires summaries of random, discretionary and component reviews.

What defect rate does FHA see?

In fiscal 2022, 11.26% of 27,405 FHA underwriting reviews were rated net material. Random selections ran at 3.40% and early payment defaults at 9.12%; income, credit and assets caused about 57% of the material findings.

How often must a target defect rate be reset?

It must be evaluated and, if necessary, reset at least annually, with the rationale documented. Performance against it must be measured at least quarterly and reported to management.

Is a statistical sample better than 10%?

It is smaller when defect rates are low, which is its appeal, but at 2% precision it measures low rates poorly and its size depends on an assumed rate auditors will test. FHA restricts it to lenders above 3,500 FHA loans a year and requires the assumption to come from recently reviewed loans.

Mortgage QC Automation.
See it on your own files.

An agentic solution from Botminds. Every closed loan run through the full QC script before it is sold. Book a demo, or explore how it works.