themortgagellm™The DPA credit premium: real in FHA, bigger in VA, absent in USDA
·
Written by The Mortgage LLM Team—a group of industry analysts leveraging our proprietary mortgage-domain language models to synthesize and decode housing data.
Across the full universe of Ginnie Mae purchase-money originations — 17.7 million loans across FHA, VA, USDA-RD, and PIH-184, spanning every monthly snapshot the agency has published — down-payment-assistance-flagged borrowers default more than non-DPA borrowers. On the raw FHA book, DPA loans are 21% more likely to hit 90+ day delinquency than non-DPA. The gap survives full FICO × LTV × DTI matching at about 15%. But push the same test into VA and the premium runs 26%. Push it into USDA and the premium disappears entirely. DPA is a real credit stress marker in some agencies and no signal at all in others, and the mechanism that explains why is what makes the finding actionable.
📌 Executive takeaways by role
- Servicers & affordable-lending LOs: The FHA DPA credit premium is real, matched-adjusted, and has widened. Baseline is 15% matched D90 vs non-DPA — but 2023–2024 vintages run 44%, driven by a non-DPA quality improvement DPA loans didn’t share. VA runs a 26% matched premium across every LTV band. USDA DPA is credit-neutral (matched ratio 1.007). Early-intervention loss-mit should flag recent-vintage FHA DPA originations specifically, not the flat book. Jump to servicing takeaway.
- Capital markets & Ginnie Mae MBS investors: FHA pools with >80% DPA concentration run 7pp higher ever-D90 than the modal 5–15% DPA pool. Because serious delinquencies trigger servicer buyouts at par, that stress translates into buyout-driven principal returns behaving like accelerated prepayments. Concentrated-DPA pools carry an embedded buyout-speed premium — model it separately from national-average CPR, layered on top of the 2023–2024 vintage widening. Jump to capital-markets takeaway.
- Program stakeholders & policy analysts: The USDA null is the load-bearing finding. DPA acts as a credit-stress marker only when the underlying program requires a down payment the borrower cannot satisfy on their own. Overlays that penalize DPA programs across the board misdiagnose the problem — regulatory focus should pivot toward program design and addressable capital gaps, not blanket credit-score or pricing penalties on low-wealth first-time buyers using structured assistance. Jump to policy takeaway.
The raw numbers
Ginnie Mae publishes a down_payment_assistance flag at the loan level for every FHA, VA, USDA-RD, and PIH-184 mortgage in its loan-level disclosure (llmon2). Across the full available history through May 2026, the raw picture per agency looks like this:
| Agency | Purchase loans | avg FICO | avg LTV | avg DTI | non-DPA D90 | DPA D90 | Raw D90 ratio |
|---|---|---|---|---|---|---|---|
| FHA | 10.9 M | 682 / 677 | 96.08 / 96.14 | 42.82 / 43.15 | 18.60% | 22.56% | 1.213 |
| VA | 4.90 M | 717 / 703 | 97.57 / 95.98 | 41.17 / 42.34 | 8.41% | 12.17% | 1.447 |
| USDA-RD | 1.42 M | 697 / 696 | 99.68 / 99.14 | 35.29 / 35.71 | 19.04% | 18.76% | 0.985 |
| PIH-184 | 32 K | 700 / 696 | 95.93 / 93.80 | 33.07 / 33.41 | 11.58% | 14.94% | 1.290 (thin) |
Two things stand out before any statistical work. First, the raw DPA delinquency premium is not uniform across Ginnie agencies. FHA is the mid-tier at 21%; VA is the largest at 45%; USDA is essentially zero. Second, DPA volume is heavily concentrated in FHA (1.63 M DPA loans, 15% of the agency’s purchase book) with meaningful presence in VA (102 K, 2%) and USDA-RD (37 K, 3%). PIH-184 is a small tribal-housing program whose DPA cell is too thin for matched analysis and will not feature further.
The raw ratios are the noisy version of the answer. DPA borrowers have marginally lower FICO than non-DPA borrowers in every agency, and mildly higher DTI, and slightly different LTV composition. The obvious question is whether any of the raw premium survives once we hold FICO, LTV, and DTI fixed.
What survives matched-cell adjustment
For each agency we bin loans on the standard three underwriting axes — FICO in five bands (below 620, 620–659, 660–699, 700–739, 740+), LTV in six bands (80–90, 90–95, 95–96.5, 96.5–97, 97–100, above 100), and DTI in four bands (below 36, 36–42, 43–49, 50+) — for 120 possible cells per agency. We then compare DPA to non-DPA within each cell that has at least 200 DPA loans and 500 non-DPA loans, and weight the cell-level DPA-vs-non-DPA ratio by the non-DPA loan count. That gives a composition-adjusted delinquency rate you can read as “what DPA loans would experience if they had the same FICO × LTV × DTI mix as non-DPA loans.”
| Agency | Matched cells | Matched DPA n | Matched non-DPA n | Matched D30 ratio | Matched D90 ratio | Matched D180 ratio |
|---|---|---|---|---|---|---|
| FHA (F) | 108 | 1,356,752 | 7,775,318 | 1.128 | 1.154 | 1.150 |
| VA (V) | 80 | 82,356 | 4,221,395 | 1.208 | 1.264 | 1.132 |
| USDA (R) | 31 | 29,579 | 1,005,690 | 1.030 | 1.007 | 0.993 |
DPA is not an inherently risky borrower trait. It is an indicator of stretched financial capacity when — and only when — the underlying program requires a down payment the borrower cannot satisfy on their own.
FHA loses about a third of the raw premium to composition. The residual — a 15.4% matched D90 premium on 1.36 M DPA and 7.78 M non-DPA loans across 108 matched cells — is real, not a compositional artifact. VA loses less of its premium and lands at 26.4%. USDA collapses to essentially zero.
The persistence in FHA is worth staring at for a moment. Standard treatments of DPA versus non-DPA credit risk in the trade press cite raw premiums of 20-25%, which is roughly right for FHA nationally. When those premiums are naively attributed to LTV bunching — DPA borrowers push to the 96.5% ceiling and are therefore riskier — the implicit claim is that the premium collapses at matched LTV. In FHA specifically, it does not. The 15% residual is the load-bearing empirical for the “DPA carries a real credit premium” claim. The full-cell adjustment strips the LTV composition story out of the way and leaves 15pp of residual.
The LTV-band decomposition on FHA makes it easy to see:
| LTV band | DPA n | non-DPA n | DPA D90 | non-DPA D90 | ratio |
|---|---|---|---|---|---|
| 80–90 | 89,926 | 471,798 | 18.7% | 16.2% | 1.157 |
| 90–95 | 85,532 | 530,764 | 20.0% | 16.4% | 1.221 |
| 95–96 | 35,501 | 221,872 | 22.0% | 17.1% | 1.286 |
| 96–96.5 | 175,821 | 1,091,311 | 22.9% | 18.3% | 1.250 |
| 96.5 exact (modal) | 343,664 | 2,161,683 | 21.5% | 17.9% | 1.199 |
| 96.5–97 | 45,434 | 329,653 | 19.6% | 18.0% | 1.089 |
| 97–100 | 796,594 | 4,108,710 | 23.9% | 19.7% | 1.209 |
The FHA premium is persistent across every LTV band we can measure — 9% to 29%, depending on the slice, and 20% at the 2.5M-loan modal 96.5% exact ceiling. This is not a compositional artifact hiding at one end of the LTV curve. It is a real behavioral difference between borrowers who used DPA and borrowers who did not, at the same LTV.
The VA surprise
The story gets more interesting once we cross into VA. VA’s raw DPA D90 premium is 45%; the matched-cell adjustment leaves a 26% residual — nearly twice the FHA residual, on a program whose DPA universe is 1/16th the size of FHA’s. That is not a scale artifact — VA’s matched analysis draws from 80 matched cells across 82 K DPA loans and 4.22 M non-DPA loans, more than enough to move the number.
The mechanism is worth naming, because it flips a common assumption about DPA. VA loans are structurally 100% financing for most eligible veterans; the VA entitlement covers the borrower’s down payment gap at zero cost. A VA borrower who nonetheless uses DPA has, by definition, a gap that even the VA guarantee could not close. That is a very different selection story than DPA in FHA, where the DPA is filling the standard 3.5% minimum-down-payment gap most FHA borrowers do have. VA DPA is a marker of stretched capacity above and beyond what the program itself already provides — and the matched-cell premium reflects that.
By LTV band, the VA premium is more consistent than FHA’s:
| LTV band | DPA n | non-DPA n | DPA D90 | non-DPA D90 | ratio |
|---|---|---|---|---|---|
| 80–90 | 13,498 | 397,424 | 4.4% | 2.9% | 1.532 |
| 90–95 | 8,454 | 262,539 | 7.4% | 5.0% | 1.490 |
| 95–96 | 2,943 | 99,906 | 9.5% | 5.6% | 1.701 |
| 96–96.5 | 1,863 | 67,358 | 10.3% | 5.5% | 1.849 |
| 97–100 | 39,443 | 2,358,508 | 14.0% | 9.0% | 1.557 |
The VA DPA premium runs 49–85% in every LTV band we can measure, and doesn’t converge at any ceiling the way FHA’s does. If FHA DPA is a moderate credit stress marker, VA DPA is a strong one.
Hover any point for the DPA and non-DPA rates and cell counts. USDA sits at ratio ≈ 1.00 across every LTV band.
The USDA null
USDA-RD is the counterexample that clarifies the mechanism. USDA is a 100%-LTV program by design — the average LTV in our USDA cohort is 99.68% for non-DPA and 99.14% for DPA. The DPA and non-DPA borrower populations are indistinguishable on FICO (696 vs 697), DTI (35.3 vs 35.7), and the ever-D90 rates are within a fraction of a percent (19.04% vs 18.76%). The matched-cell D90 ratio is 1.007 — 0.7% premium, effectively zero. The 90+ DQ premium at every LTV band ranges from 0.91 to 1.04.
The interpretation is straightforward. In FHA and VA, DPA is a marker of a borrower whose down-payment gap exceeded what the program’s minimum-down-payment structure could absorb, and that gap is proxying for other unobserved stresses — thinner reserves, less financial cushion, higher marginal reliance on the DPA program to keep the deal together. In USDA, the program’s own 100% financing already absorbs that gap. Adding DPA on top is redundant liquidity assistance, not a marker of stretched capacity — because the borrower’s capacity constraint was never a down-payment gap in the first place. Program design absorbs the DPA-DQ link entirely.
That is the sharpest empirical point in the analysis: DPA is not universally a credit-stress marker. It’s a stress marker only when the program requires a down-payment the borrower cannot make on their own.
The vintage widening
The FHA DPA premium has not been static. Broken out by origination vintage, the matched-LTV D90 ratio (weighted across LTV bands using non-DPA composition) tells a two-phase story:
| Vintage | DPA n | non-DPA n | non-DPA D90 | DPA D90 | D90 ratio |
|---|---|---|---|---|---|
| pre-2013 | 213 K | 1.29 M | 19.0% | 21.9% | 1.149 |
| 2013–2017 | 464 K | 2.79 M | 19.1% | 23.5% | 1.226 |
| 2018–2021 | 530 K | 2.54 M | 22.5% | 27.1% | 1.205 |
| 2022 | 94 K | 534 K | 25.0% | 28.6% | 1.146 |
| 2023–2024 | 109 K | 1.03 M | 15.0% | 21.6% | 1.442 |
| 2025+ (early) | 167 K | 763 K | 3.3% | 4.2% | 1.256 |
For a decade of origination cohorts through 2022, the FHA DPA D90 premium sat in a tight band of 1.15–1.23. In 2023–2024 vintages, it jumped to 1.44. The mechanical driver is compositional in a specific way: the FHA non-DPA book saw a meaningful quality improvement post-2022 — matched-LTV non-DPA D90 dropped from 22.5% to 15.0% as post-COVID origination tightened underwriting overlays and refinance flushed out weaker legacy loans. The DPA book didn’t get the same benefit. The DPA D90 rate stayed at 21.6% in 2023–2024 vintages, so the gap between DPA and non-DPA widened even though the absolute DPA number is modest.
Hover any vintage for DPA / non-DPA weighted D90 rates and sample size; the diamond marker line shows the ratio on the right axis.
The 2025+ cohort is still too young for reliable D90 accumulation — only about 3–4% of loans have hit the 90-day threshold. But the ratio there sits at 1.256, so the widening trend that showed up in 2023–2024 is not, so far, reverting.
For servicers and MBS investors, that widening is the operational signal buried in the level. The current-vintage FHA DPA premium is not the historical 15–20%. It is 40%+, and it is coming from a widening gap between two subpopulations whose underlying stress paths have diverged.
Pool-level: what it means for MBS investors
Ginnie’s loan-level disclosure gives us pool_id on every FHA loan, so we can aggregate to the pool and correlate pool-level DPA share against pool-level 30+ DPD rate. Restricting to pools with at least 100 FHA purchase loans, and weighting the pool D90 rate by loan count:
| Pool DPA share | Pools | Loans | Pool D90 rate |
|---|---|---|---|
| 0% (small legacy pools) | 1,364 | 249 K | 20.5% |
| Under 5% | 552 | 298 K | 14.7% |
| 5–15% (modal) | 1,295 | 5.16 M | 19.2% |
| 15–30% | 1,059 | 2.16 M | 18.8% |
| 30–50% | 338 | 161 K | 23.3% |
| 50–80% (rare) | 11 | 1.3 K | 25.6% |
| Over 80% DPA (concentrated) | 257 | 35 K | 27.4% |
Pools with more than 80% DPA concentration run about 7 percentage points higher D90 than the modal 5–15% DPA pool. The 0% pools are a legacy small-pool artifact — they are older and smaller and are best set aside from the trend interpretation. Within the meaningful volume, pool DQ scales monotonically with DPA share above the 15% threshold.
That has a direct implication for Ginnie MBS investors and MSR desks. When a Ginnie pool’s active loan count declines through delinquency-triggered buyouts — the servicer’s obligation to repurchase seriously delinquent FHA loans at par — the return of principal looks identical to a fast prepayment. Concentrated-DPA pools carry an embedded buyout-speed premium that should be priced separately from any national-average CPR assumption. On the pricing side, the pool-level premium is 7pp higher D90 than modal pools, which converts materially to CPR-equivalent speed assumptions on a monthly buyout basis over the life of the pool.
VA and USDA pool-level signals are weaker because DPA-concentrated pools are rare in those programs — most VA and USDA pools sit below 5% DPA share. The pool-level story is essentially an FHA story.
What it implies
🏦 For servicers and affordable-lending LOs: target the vintage gap
DPA is not a flat credit penalty across the FHA book. It is a real credit signal, and its magnitude has moved. A DPA borrower in FHA is 15% more likely to hit 90+ DQ than a non-DPA borrower with the same FICO, LTV, and DTI — and 44% more likely in current 2023–2024 vintages. In VA, the matched-cell multiplier is 26%. Early-intervention loss-mit protocols should specifically flag recent FHA DPA originations rather than treating all vintages the same. Conversely, USDA-RD DPA can be treated as credit-neutral liquidity assistance, not a risk multiplier — the matched-cell ratio there is 1.007.
📊 For capital markets and Ginnie Mae MBS investors: price the buyout premium
Ginnie pools with over 80% DPA concentration run about 7 percentage points higher ever-D90 than the modal 5–15% DPA pool. Because serious delinquencies trigger servicer repurchases at par, that stress rate translates directly into buyout-driven principal returns that behave like accelerated prepayments. Secondary desks should model differential CPR assumptions on concentrated-DPA Ginnie pools rather than relying on national-average speeds. The 2023–2024 vintage widening means the current-book premium is running above the historical average — a level effect layered on top of the concentration signal.
🏛️ For program stakeholders and policy analysts: program design, not blanket overlays
The USDA null is the most important finding for anyone thinking about DPA regulation. It is direct evidence that the DPA-to-delinquency relationship is program-conditional, not universal — DPA acts as a credit-stress marker only when the underlying program requires a down payment the borrower cannot satisfy on their own. Overlays that penalize DPA programs across the board are misdiagnosing the problem. Regulatory focus should pivot toward program design and addressable capital gaps, not blanket credit-score or pricing penalties on low-wealth first-time buyers using structured assistance.
The DPA question is not “does DPA raise default risk?” It is “in which programs, at which severities, in which vintages?” This analysis is one answer to that reshaped question.
For quantitative pre-application screening on this dimension, we have released a GNMA DPA propensity scorer trained on the same universe covered in this analysis (OOT AUC 0.74 on 2024–2025 vintages, calibrated Brier 0.070), with a fair-lending AIR audit passing the 4/5ths rule across race, ethnicity, and sex. The scorer produces the probability that a Ginnie-eligible purchase application will carry DPA; this analysis quantifies what that probability implies for realized delinquency. The two are complementary.
Methodology. Universe: mortgage.gnma_loans_raw filtered to loan_purpose = '1' (purchase) across all Ginnie agencies (FHA, VA, USDA-RD, PIH-184), from the earliest available snapshot through May 2026. Loans deduplicated to one row at first appearance for static origination attributes (FICO, LTV, DTI, DPA flag, pool_id, state, origination year); delinquency observation aggregated as MAX(months_delinquent) over the full history of that loan across all monthly snapshots — the “ever-DQ” convention. Matched-cell adjustment bins on FICO (5 bands), LTV (6 bands), DTI (4 bands) and weights DPA rates by non-DPA composition within cells with ≥200 DPA loans and ≥500 non-DPA loans. Pool-level analysis restricted to pools with ≥100 active loans. GNMA does not expose Section-of-Act at the loan level in the loan-level disclosure, so the FHA 203(b) sub-restriction confound check that would normally be run in a similar analysis of the traditional-purchase-only universe was not available here; 203(b) is the dominant FHA program and contamination from Section 234 and 245 is expected to be small. Down-payment-assistance flag is Ginnie’s down_payment_assistance field, per-loan Y/N. Fair-lending audit for the referenced DPA propensity scorer used geographic-weighted HMDA state × demographic composition (loan_type IN (2, 3, 4) for government programs) — the industry-standard fallback for loan-level GNMA↔HMDA matching. Informational, not advice.