Credit Validation:
Delinquency and Default

1How the credit model works

The credit model holds a book as a distribution of loan mass over payment states and advances it one month at a time. It does not simulate individual loans. What moves mass between two states is a monthly hazard: the probability that a loan in the first is in the second a month later. Every figure the model reports follows from those hazards applied repeatedly.

Each hazard is a product: a base monthly rate, a smooth curve in each of its covariates, a month-of-year factor, and a dated period factor. Being multiplicative, a covariate scales a hazard rather than shifting it — a curve value of 1.5 raises that month's flow by half, whatever the base rate. Stocks lag flows: a hazard change moves this month's flow at once and the state's share only over the months that follow.

The model projects when loans default, not what a default costs: it carries no severity, no loss given default and no timeline to recovery.

The states

The Deep block holds the exact month delinquent rather than one 120-or-more bucket, because the monthly probability of foreclosure completion is not flat across it: one rate over a pool whose mix of durations moves would liquidate young arrivals too fast after a shock and long-duration survivors too slowly.

Two borrower types

A book is carried not as one population but as a mixture of two types: a small share of loans that enter delinquency at a high multiple of the rest, and the remainder. Neither type is observed. What is observed is a loan's own delinquency record, which is informative about which type it carries — and that is what the repeat ladder and the history clocks are for. A loan that has been delinquent before is likelier to carry the higher-risk type, so the model returns it to Current with a clock running and puts its next episode on the repeat ladder: the clock carries the elevation that decays, the type the part that does not.

Four numbers describe the mixture on each book. The conventional book carries a higher-risk share of 0.12 entering delinquency at 8.56 times the rate of the rest, a persistence of 0.45, and 1.052 for the entry level among loans with no delinquency history. The last is solved so that the model reproduces the measured share of the book 30 or more days delinquent; the other three are ranked over a grid (Section 3).

The estimated transitions

Ten transitions are estimated: entry from Current to D30; a cure and a roll at each of D30, D60 and D90; and three competing exits from the Deep state — cure, voluntary payoff and liquidation.

Their covariates differ. Entry reads credit score, loan-to-value and debt-to-income, and on the conventional book original balance as well (a small loan enters delinquency faster than a large one at the same credit score, loan-to-value, age and debt-to-income); the shallow cures and rolls read credit score, loan-to-value and loan age; the Deep-state exits read loan-to-value and the exact month delinquent, and on the conventional book alone Deep cure and liquidation also read original balance (Section 3). Entry carries a fitted loan-age curve the deployed conventional engine does not read: with the mixture in force, composition produces the age profile instead. Seven dated period windows run from the pre-2007 years through the 2020 shock and the 2022 payment restart, so no projection averages a foreclosure moratorium with an ordinary year.

What the covariates do

Each figure is the multiple of that hazard at that value of the covariate, read from the deployed conventional curves.

Entry falls steeply as credit score rises — a 620 score enters at about 11.8 times the rate of a 780 score, other things equal — rises throughout with loan-to-value, and rises with debt-to-income above 30 before flattening. Out of the Deep state, liquidation rises steeply with the month delinquent and cure falls, while voluntary payoff peaks in the second year. Loan-to-value on Deep liquidation rises throughout its fitted range, 15 to 95, and holds its end value above 95 — the buckets above 100 carry too little 2024–2025 deep exposure to fit. A curve holds its end value outside its fitted range, so the credit-score curve is flat below 600 and above 800.

One hazard, worked

Entry into delinquency in June 2026, for a conventional loan at credit score 740, loan-to-value 75, debt-to-income 38:

That product, 0.644 percent per month, is the fitted rate for the book as a whole; the mixture then splits it. A loan with no delinquency history enters at 0.321 percent per month on the lower-risk type and 2.745 percent on the higher-risk one, an average of 0.612 percent — below the fitted rate, which also contains the re-entries of loans that have been delinquent before. Every other hazard compounds the same way.

What reaches the model from the economy

One channel does. A projection may carry a path for the change in the unemployment rate, which multiplies entry into delinquency by 1.05 raised to the power of the rise in unemployment, in percentage points, twelve months earlier. It is lagged and one-sided by construction: this month's entry answers to the change a year before, and a falling path leaves a projection unchanged to the last digit. No other hazard reads a macroeconomic variable or a mortgage rate, so a projection responds to interest rates only through its loan-to-value path. Every figure on this page is computed with no unemployment path supplied, so the channel is inert in all of it.

Composition is measured, not fitted

The ten hazards set how fast mass leaves a state; a second set of rules sets where it lands: whether a cure returns a loan to Current or only to a less delinquent state, how a repeat episode differs from a first, how foreclosure procedure changes liquidation timing, and how the delinquency counter advances — a borrower who makes a single payment holds it flat, and about three quarters of observed Deep months advance it.

They are measured on the loan panels because fitting them would not identify them. A faster entry rate with more cures returning to Current produces nearly the same aggregate delinquency share as a slower entry rate with fewer, so an estimation free in both would trade one against the other and settle wherever it began. Measuring the destinations leaves the estimated hazards as the only thing that can account for the observed flows, and leaves a projection with nothing to tune (Section 8).

Estimation, and the government tail

Only the exit event differs between the two bases: the fitted liquidation hazard on the whole-loan basis, where a modification is a cure, and removal at par on the security basis. The conventional entry and shallow transitions are estimated on the Freddie Mac panel and scored afterwards on the Fannie Mae panel (Section 4); the conventional Deep-state hazards on both panels pooled over the recent window. The government books are estimated on the Ginnie Mae loan-level disclosure, which stops counting at six months delinquent, so a loan six months delinquent and one thirty months delinquent are one observation and the profile past six months has to be constructed.

Cure and voluntary payoff are taken to be borrower behavior rather than program procedure, so the conventional shape is carried over for them at the government book's own level: only the shape is borrowed. That borrowing rests on the judgement, not on the evidence — scored against the record below, the borrowed shapes are rejected on every government book, by the widest margin on FHA cure (Section 3). For liquidation, which is procedure and not behavior, nothing is borrowed: what the disclosure does record is how long a loan sits at the cap and what fraction leaves each month, and that recovers the shape beyond it, since a shape that liquidates early empties the cap quickly and one that liquidates late holds loans in it.

One system with the prepayment projection

The prepayment projection is computed as though every loan were current, the credit model owns the delinquency state, and each month the projected state shares recompose the voluntary channels: the current share at the performing speed, the delinquent share at its own measured multipliers. Running the two independently and adding them would double-count, a loan the credit model holds in foreclosure also refinancing on the prepayment curve.

Those multipliers differ between the two kinds of book. On the government books turnover among delinquent borrowers runs at 2.3 to 2.9 times the performing rate — distressed and relocation sales — while refinance, cash-out and curtailment run at 0.05 of it, a delinquent borrower being unable to refinance, qualify or curtail; a rising delinquent share therefore slows a government pool in a refinance market and speeds it up in a turnover market. On the conventional book every delinquent channel sits at or below the performing rate: turnover 0.869, refinance 0.204, cash-out and curtailment undamped at 1.000. A rising conventional delinquent share therefore never speeds a pool up, slowing it sharply in a refinance market and slightly in a turnover one.

What the evidence on this page establishes

Every comparison that follows is made within the window the hazards were measured on: the model against the 2024–2025 book, the deployed hazards against a second agency's loans over the same years, constructions against what they were built to reproduce, and measured rates against measured rates. Two things are therefore established: that the estimated hazards reproduce the book they were measured on, and that seven of the ten — entry and the six shallow cures and rolls — hold on an agency panel not used to estimate them. The other three, the Deep-state exits, are estimated on both panels pooled, so that panel is in sample for them, liquidation included.

No forecast record is established. Nothing here runs the model forward from a date and compares the projection with what happened, and nothing reports behavior in a credit downturn. A calendar-time comparison on the Fannie Mae panel is the planned extension.

2Delinquency composition against the 2024–2025 book

The model is run on the 2024-01 – 2025-12 book cell by cell — a cell being a credit score, loan-to-value, loan age and debt-to-income combination at its own observed values — over the calendar months its loans actually lived, so the dated period factors apply as in production. The result is averaged over the window and set against the measured share in each state.

Figures are in percent of book exposure, loan-count basis, with delinquency-history typing on. The ratio divides the model by the column it is compared against, except on the final row, where it divides by the measured share.

What the Deep state is compared against

The shallow states are compared against the measured in-pool share itself; the Deep state is not. Loans removed from pools at par — agency dispositions and issuer buyouts — stay delinquent on the whole-loan basis but leave pool data, so the measured in-pool share understates the whole-loan Deep share and is scaled up by an adjustment factor first. That makes the Deep row the least precise comparison on the page. The factors are Conventional (FNMA/FHLMC) 1.039; FHA 1.060; VA 1.073; USDA 1.080.

On the conventional book the factor is derived, not assumed: the ratio of the stationary deeply delinquent stock without the agencies' disposal exits to the stock with them, at the exit rates and counter dynamics measured on this window — 1.0387 on the loan-count basis, against 1.0638 under a deterministic counter and the 1.07 previously documented. That raw share already excludes loans held as real estate owned, which is like-for-like with the liquidation event. The government factors are documented in each program's coefficient file and carried here, not re-derived.

On FHA, VA and USDA the deep-tail levels were solved against these adjusted numbers, so those Deep rows read one by construction.

Conventional (FNMA/FHLMC)

FHA

VA

USDA

What delinquency-history typing carries

Switching the typing off leaves the structural baseline, the same ten hazards with no memory of a loan's own record. Its Deep-state share reads Conventional (FNMA/FHLMC) 0.286 against 0.364; FHA 1.762 against 2.446; VA 1.049 against 1.755; USDA 2.064 against 2.837 — short on every book by at least 21 percent, because most deeply delinquent loans are on a repeat episode. That gap is the deep stock the model attributes to borrowers typed as higher-risk from their own history.

3Conditional Default Rate

The Conditional Default Rate is the annualized rate of involuntary liquidation on the whole-loan basis. Figures are in annualized percent on the 2024-01 – 2025-12 book.

On the conventional book the measured column is the true default rate, involuntary liquidations over pool exposure, and the model reads 0.836 of it. The liquidation event is foreclosure completion: FNMA's liquidation as coded and FHLMC's acquisition of the property as real estate owned. On the government books it is the flow of in-pool foreclosures. Post-buyout outcomes are invisible in pool data, so that is a lower bound on the whole-loan default rate rather than a measurement, and the anchor the model carries is reconciled to it below.

Ratios on this page are computed from the unrounded rates while the columns are rounded for display, so dividing two printed figures need not reproduce the ratio quoted beside them.

With the typing switched off (Section 2) the same model's default rate reads Conventional (FNMA/FHLMC) 0.024 against 0.030; FHA 0.152 against 0.208; VA 0.098 against 0.164; USDA 0.302 against 0.416. Turning it on moves the conventional book nearer its measured 0.036 and the government books further above a column that is a floor rather than a target.

The conventional ratio rests on the two-type mixture of Section 1, three of whose four coefficients are ranked over a grid. Its four numbers are not this grid's lowest score. The grid scored a higher-risk share of 0.30 entering delinquency at 18.92 times the rest lowest; the model carries 0.12 at 8.56 times, which stands fourth of the 20 points ranked, 9.47 percent above the lowest score. The choice follows a stated rule rather than a fit: among the points within ten percent of the lowest score, four of them here, the model carries the least extreme higher-risk type. The ranking is flat across the higher-risk type's strength: four points, from 8.6 to 18.9 times the rest, sit within ten percent of the lowest score and the data cannot separate them; the model carries the least extreme of them, 0.12 of loans at 8.6 times, the size at which its two age strengths were solved, rather than the lowest-scoring 0.30 at 18.9 times, whose score is 9.5 percent lower. And the model cannot hold the measured delinquency share and the measured default ratio at once: they are one coefficient apart, and the share is what it holds, reading 1.624 against the 1.621 it is solved to. The ratio's most recent move, on 2026-09-09, is the loan-to-value axis. The previous axis carried seven buckets, one of them every loan below 60, and its curves reached no lower than 55; once a loan's loan-to-value fell through 55 the projection stopped responding to it, and the base conventional loan's 30-plus share levelled off in its ninth year and crept up for the rest of its life, against data in which it keeps falling. The axis was refit on eleven buckets, five of them below 60, and the curves now reach 15. That loan's projected 30-plus share at the end of its life reads 0.86 percent where it read 2.00. The ratio moved from 0.95 to 0.86: occupancy 0.964, composition 1.013, within-cell 0.883. Restoring the previous Deep-state hazards would read 0.88, so the move is not theirs; it is the refit entry and shallow transitions changing what flows into the deep pool, and it was accepted as the reading of the corrected axis rather than held at the previous figure by a level the data do not identify. Its latest move, on 2026-09-10, is the entry hazard reading original balance. Small loans enter delinquency faster than large ones at the same credit score, loan-to-value, age and debt-to-income — a loan under $100,000 at 1.22 times the rate of one between $200,000 and $300,000, so the book's deep pool holds more small loans than the current pool does, and small loans, once deep, liquidate fastest. Deep liquidation already read original balance; entry did not, so the model's deep pool was short of small loans and liquidated too slowly at matched cells. The model's deep pool now holds 11.2 percent loans under $100,000 where it held 9.3 and the book holds 14.5. The within-cell term moved from 0.883 to 0.909 and the ratio from 0.86 to 0.88, toward one, with the delinquency share held. Its latest move, on 2026-09-11, is the entry hazard's credit-score and debt-to-income multipliers decaying with loan age. Both are measured at origination and describe a loan less well the longer it has been paying; the hazard had applied one gradient of each at every age, so a seasoned low-score loan was projected to enter delinquency about twice as often as the book shows and a seasoned high-score loan about two thirds as often, and the two risks compounded in the low-score, high-debt corner. The estimation now measures the decay of each gradient with loan age on the full history, within origination cohorts, and the engine scales the multipliers by it, with the gradient fitted free at every age rather than shrunk by one factor: past ten years a score under 620 carries 3.82 times the reference entry rate against 10.20 at four years, while a score of 780 or more holds at 0.59 against 0.54; the departure of each age's gradient from the four-year one is applied at 1.40 times its fitted size, solved at the model level with the survivor composition in place. The delinquency ladder moved with it: the population's age pattern in cure and roll rates is the share of repeat delinquents rising with loan age, and the model had applied that pattern and a pooled first-versus-repeat multiplier together. The first-versus-repeat multipliers now carry the age axis, the first-episode 30-day cure running 1.10 times the blended rate under a year and 1.65 past ten, so young delinquents no longer cure too readily and old ones too little. On the 2024–2025 book the model's entry rate for loans with a credit score under 620 and a debt-to-income ratio over 43 ran 1.73 times the book's, relative to the overall ratio; it now reads 1.12. By loan age the model's entry rate ran 1.55 times the book's for loans under a year old and 0.65 for loans past ten years; it now reads 1.13 and 1.06. A borrower whose only missed payments were a forbearance or another program pause is no longer carried as a repeat delinquent for life. Such a record is its own class, with its own history clock measured on the disclosures' assistance and deferral fields: it raises the entry rate far less than a missed-payment delinquency of the same age does, and it leaves a smaller permanent mark. The model's replay of the pandemic years counts 71.08 percent of the 2020 delinquencies and 59.15 percent of the 2021 ones as pauses, the shares the disclosures record, so the population with a delinquency record that it builds is the one the book has. The entry hazard also carries 0.80 of the estimation's age effect at the reference covariates in log terms, the share the survivor composition does not already carry, solved on the 2013–2019 and 2024–2025 books together. The ratio moved from 0.88 to 0.84 with the delinquency share held.

Loan balance

At matched months delinquent and loan-to-value, small-balance loans liquidate at a higher rate and cure at a lower one. The multipliers below are that fitted margin against the 200-300K bucket, on the conventional book alone.

Balance is the original balance; a loan submitted without one is projected at 300,000. A realized ÷ model column is not shown for either hazard: it scores a fitted one-way margin on the same pooled cells it was estimated from, so all twelve of its cells sit at one within one percent by construction. The model's deep pool holds fewer loans under 200,000 than the measured deep exposure, the transitions feeding the Deep state carrying no balance term — an open composition item. The gradient shows in the aggregate: the balance-weighted CDR is 0.816 of the loan-count rate in the model (0.021 against 0.026) and 0.824 in the data (0.029 against 0.036). Those four figures are computed on the earlier evaluation, in which every cell starts in 2023, and are not comparable with the headline table above.

From the in-pool floor to the whole-loan rate

Government pool data record a buyout, not the outcome that follows, so the table reconciles the floor to the whole-loan anchor. Its columns are percentages of entrants in a 66-month follow-forward of loans entering serious delinquency at age six months or more between 2018-09 and 2020-12. That liquidation share is assumed, not measured: it is set at 0.200 against a published range of 0.17 to 0.30 for Ginnie Mae early buyouts (Bandyopadhyay, Kim and Smith, Journal of Financial and Quantitative Analysis), over which the implied eventual foreclosure spans FHA 10.722 to 15.687; VA 13.171 to 15.940; USDA 14.451 to 17.732. The last two columns agree by construction — each anchor is the implied value at the assumed share — so the table states how the anchor was set rather than testing it (Section 9).

Deep-state rendering on the government books

Under the six-month cap (Section 1) these transitions are estimated at months four and five and as one six-and-beyond bucket, and the curves are built to reproduce all three, the six-plus one averaged over the conventional 2024–2025 exposure. Those 27 ratios are one by construction and are not shown. The first column re-takes that average over the model's own deep composition, which is not constrained; the second describes the tail's shape.

The shape of the deep tail beyond the disclosure cap

Past six months the one remaining observation is how long loans sit at the cap and what fraction leaves each month; the record covers 2024-25 and carries 3,751,876 loan-months at the cap across the three books. Each candidate is pushed through it at its best single level and scored by Poisson deviance, the unconstrained fit being the lower bound any shape can reach. Degrees of freedom count the cells carrying exposure, less one; deviance ranks shapes against one another and is not an absolute measure of fit.

On each row one column is evidence and two are the same fit written twice. The liquidation rows carry a shape measured from this record, so their model and unconstrained columns are one number and agree by identity; the informative column there is the conventional profile, which the record rejects on all three books. The cure and payoff rows carry the conventional profile, so those two columns are one number and the informative column is the unconstrained fit — far lower on every one. The record therefore chooses the liquidation shape and is overruled on the other two, cure and payoff being taken as borrower behavior (Section 1).

The measured liquidation shape is dated to the window above and is refreshed with the data, so the government deep tail is constructed rather than observed. Its level is then solved jointly against the deep stock, the aggregate delinquency level and the eventual-foreclosure anchor.

4Transition accuracy by calendar year

4.1Conventional — FNMA panel (not used in estimating the entry and shallow transitions)

The conventional entry and shallow hazards are estimated on the FHLMC panel and scored here on the FNMA panel by calendar year; the Deep-state hazards on the pooled 2024–2025 window of both panels, so those years are in sample for them. Each cell is realized ÷ model for that transition in that year, shaded by distance from one; the final column pools 2020 onward (2023 onward for the Deep-state rows, marked ‡).

Nothing is re-estimated for this table. Panel: FNMA acquisitions 2000Q1 to 2026Q1, performance 2000-01 to 2026-02. A year is scored only above 50,000 loan-months of FNMA exposure in the originating state; blank cells fall below that threshold. § The Deep-state hazards are estimated on the 2024–2025 window pooled across the FNMA and FHLMC panels and carry no historical period factors, so they apply from 2023 onward and their pooled column (‡) starts there. For these rows 2024–2025 are in sample. Their earlier ratios measure the distance between the current-regime hazard and that year's rates, not a forecast for those years. 2026 covers 2026-01 to 2026-02 only. The three rows carrying the loan's original balance are scored on the loan-level FNMA cells available from 2000, the aggregated tables not recording it, so their earlier cells are blank. Entry is the one row on which the two agencies differ materially: the FNMA entry rate runs about ten percent above the FHLMC-estimated rate pooled 2020 onward. The model carries no agency factor, and Section 9 records the difference as unsettled. Two further hazards in the estimation file are descriptive and not projected, and so not scored here.

4.2Government, in sample by calendar year

The government hazards are estimated on the Ginnie Mae loan-level disclosure, so every year here is in sample. Each figure is the same realized ÷ model ratio: the 2025 column is that year's, and the range is the lowest and highest annual ratio the transition takes.

The range runs over 2013–2026. 2021 and 2022 are excluded from it. Each carries its own single-year period factor, so every one of the 60 ratios in those two years is one by construction and carries no information. 2013 covers October to December only; 2026 covers the months observed to date. The Deep-state buyout transition is a pool-removal event, not a borrower event, and is estimated for the security-basis overlay.

5Security-basis removals (CBR)

The comparison below sets the model's month-one removal flow against the measured involuntary removal CPR over the trailing twelve months of disclosure, weighted by unpaid principal balance, the model started from each book's own observed delinquency composition rather than its long-run one.

Each row states its own window. The conventional row is a book-level run: the 2024–2025 book's own cells, each started at its own age from the measured delinquency composition and projected 24 months on the security basis against the same book's balance-weighted implied pool CDR. Its removal flow is the fitted liquidation, the GSE repurchase at 24 months delinquent, and the measured share of cures through a first modification; the model reads 0.305 by loan count.

Basis difference

Both columns are the model's mean involuntary removal rate for a book-representative loan at sixty months of age over a 24-month horizon, as of 2026-09. The whole-loan runs start from the model's own long-run delinquency composition at that age; the security-basis runs start from each book's measured 2024–2025 composition at that age. The difference is the CBR contribution: loans bought out that then cure, are modified or prepay never appear as whole-loan defaults.

6Recent out-of-sample observations

6.1Conventional

Measured monthly transition rates out of the two deepest delinquency states on the FHLMC panel, for the 2024–2025 estimation window and for the most recent six observed months, in percent per month of the loan-months in the state. “REO / in liquidation” holds loans in real-estate-owned or other terminal status.

This is measured against measured — a stability check of the transition mechanics over time, not a model comparison. Liquidation out of the Deep state runs above the 2024–2025 level in the recent window (0.616 to 0.688 percent per month, an increase of about 12 percent); cure to current from the Deep state is little changed. “Any improvement” counts every move to a less delinquent state, cures to current included.

6.2Government — June 2026 observed against the 2024–2025 window

Monthly removal rates from each delinquency depth — buyout, foreclosure and other involuntary removal from the pool — observed in June 2026, against the 2024–2025 window on which the removal hazards were estimated, in percent per month.

• On the VA book removal runs above the window at every delinquency depth. The excess at the deep states is consistent with the volume of loans purchased out of pools under the VA Servicing Purchase (VASP) program.
• On the FHA book the removal profile has changed shape rather than level: removal at D90 is above the window while removal from the deep states is below it, consistent with the 2026 changes to FHA loss-mitigation options, which move resolution earlier in the delinquency sequence.

7Realized pool-basis default removals, 1999–2025

Realized annual rate at which seriously delinquent loans were removed from FHLMC pools (1999–2025) and FNMA pools (2012–2025), reconstructed from loan-level delinquency records and the documented buyout policy calendar. The vertical axis is logarithmic.

The 2010 spike is one month: the implied FHLMC removal rate ran 0.896 percent a year in 2010-01, 27.497 in 2010-02 as 209,266 loans were bought out under the change in buyout policy, and 2.888 in 2010-03.

This is realized history, not model output, and it is why the model separates the whole-loan default rate (CDR) from security-basis removals (CBR): the removal rate a bondholder experiences is set by buyout policy as much as by borrower default, and it moved by more than an order of magnitude when that policy changed. 2025 is a partial year: the FHLMC series ends 2025-09 (nine months) and FNMA runs through 2025-12 with terminal-month attenuation. The FNMA line begins in 2012 because the series predates that panel's back-extension and has not been recomputed, so the 2010 window is FHLMC only.

8Structural properties

Each of the following holds for every projection and is enforced by an automated test in the model's build.

• Identical inputs produce identical outputs: the engine carries no randomness and applies no perturbation.
• Every projected hazard and occupancy is finite and lies in [0, 1] over a 120-month horizon.
• Modifications are cures, never defaults: with the liquidation hazard removed the CDR is exactly zero while the delinquency pipeline remains populated — no other channel can generate CDR.
• The preceding zero-CDR property is not vacuous: the deployed coefficients do produce a positive CDR.
• CDR and 30+ day delinquency are monotone increasing in LTV.
• Covariate factor curves are continuous: crossing a former bucket edge (LTV 60) moves outputs by less than one percent, and the curves pass through the fitted bucket multipliers.
• The engine constructor exposes only scenario inputs (coefficient path, unemployment-shock amplitude, entry scale) and the delinquency-history typing switch; there are no free tuning parameters.
• The security-basis removal component is inert on the whole-loan basis: an engine carrying it produces outputs identical to the deployed engine when run on the whole-loan basis.
• On the security (pool) basis, government pool removal supersedes deep liquidation: the removal flow stands even when the fitted liquidation hazard is removed, because the two flows are different processes.
• The raw monthly credit hazard is withheld from the public tier of the application; the rounded display fields the composition views read travel to every signed-in tier unaltered (opened to all signed-in users on 2026-09-09).

One is narrower than it reads. The loan-to-value property is checked at two values, 55 and 115, on a ten-year projection of one representative loan: it establishes that the projected default rate and the 30-plus share are higher at the second, not that they rise at every point between. Every loan-to-value curve the projection reads rises throughout its fitted range (Section 1); the Deep-state curves hold their end value above 95, the highest loan-to-value the 2024–2025 deep pool carries in weight.

9Reading the evidence

What the composition shows

The model reproduces the 2024–2025 book within five percent on D30, D60 and D90, for every book. The Deep state is the largest composition error, seven percent under on the conventional book. On FHA, VA and USDA the Deep row sits on its comparison figure by construction and carries no information about fit.

What the default rate shows

The conventional ratio of 0.84 — 16 percent under the measured default rate — is the product of three terms that pull in opposite directions, so it is not evidence that the parts agree. Occupancy, 0.962: the model's Deep share sits below the measured share. Composition, 0.965: its deep pool leans less heavily on the long-duration delinquencies that liquidate fastest. Within-cell, 0.901: at matched cells its liquidation rate per deeply delinquent loan is below the book's. By loan age the agreement is uneven and the youngest loans carry the over-statement: model Deep share ÷ measured reads 1.57 at <12 months, 0.79 at 12-35 months, 1.18 at 36-59 months, 0.81 at 60-119 months, 0.98 at 120+ months. The conventional model has no seasoning ramp for newly originated loans, its entry hazard reading no loan age. The government default rates cannot be measured directly in pool data, so they are reconciled to a longitudinal anchor that follows seriously delinquent entrants forward through buyout to their eventual outcome. That reconciliation is a construction, not a check: each anchor was set equal to the in-pool foreclosure share plus the bought-out share times an assumed liquidation share of 0.200, which is why the last two columns of the table in Section 3 agree, and the deep-tail levels were then solved to reproduce it. What could falsify the anchor lies outside this page: a measurement of that liquidation share away from 0.200 would move every government default level here.

What is in sample and what is not

On the FNMA panel — not used in estimating the entry and shallow transitions — nine of ten conventional transitions sit within five percent of the model pooled 2020 onward (Deep-state transitions from 2023, the first year they apply, with 2024–2025 in sample for them). Seven of the ten are out of sample there; the other three, the Deep-state exits, are estimated on the two panels pooled, so the transition that produces the default rate is scored where it was estimated. The exception to the five percent is entry into delinquency (current → 30) at 1.10 — the one transition both out of sample and outside it. Why the agencies differ on it is an open item: the difference is documented as a property of the panels rather than of the borrowers, and nothing here evidences that reading. The calendar-year period factors absorb the 2020–2022 pandemic rather than explain it, so the in-sample ratios for those years are not evidence of forecasting skill.

What the model does not carry

Liquidation out of the Deep state has run above the fitted level in the most recent conventional window, and the model does not carry that drift. Government pool-removal mechanics moved in 2026, so the security-basis removal rates are refreshed monthly on a trailing-twelve-month window while the whole-loan hazards remain on 2024–2025 (Section 6). The conventional deep pool holds fewer small-balance loans than the measured deep exposure, the transitions feeding it carrying no balance term (Section 3).

Method notes

• What the model is compared against. Loan-count shares of the 2024–2025 pool, re-measured on the current data vintage from the same loan-level tables the model is run on — the book as it stood, not an aged steady state.
• CDR. Liquidation events over pool exposure in the window, converted from a monthly to an annual rate.
• Data. Conventional: the entry and shallow hazards deployed here were estimated on the Freddie Mac loan-level panel that its coefficient file records as 1999Q1 – 2026Q1; the Deep-state hazards on the pooled Fannie Mae and Freddie Mac 2024–2025 window. The Freddie Mac tables the estimation reads cover the same acquisition quarters today, with performance observed through 2026-02. Government: the Ginnie Mae loan-level disclosure from 2013-10 onward.
• Model versions. Conventional credit_conv_v1.13 · FHA credit_fha_v1.10 · VA credit_va_v1.10 · USDA credit_usda_v1.10, the deployed builds, from which these figures were regenerated.
• Elsewhere. Whitepaper · prepayment validation report.