themortgagellm^™

Written by The Mortgage LLM Team—a group of industry analysts leveraging our proprietary mortgage-domain language models to synthesize and decode housing data.

📌 Executive takeaways by role

• Capital markets, MBS investors & MSR owners: The LLPA grid prices roughly a third of the risk function — the rest (state concentration, DTI distribution, prepay drivers) flows to your pricing layer. Pricing pools off the LLPA-implied risk number alone systematically undervalues geographically diverse pools and overpays for state-concentrated ones; MSR strip valuations should explicitly carry the DTI and employment predictors the grid omits. Jump to capital-markets takeaways.
• Originators & aggregators: Inside a single LLPA cell — same FICO, same LTV, same 75 bps fee — modification rates vary 8.8× by DTI and 21× by state. That gap is structural margin. The lender who can identify and retain the lower-risk borrowers within a cell captures the residual; the one who can’t gets adversely selected on the rest. Mid-size lender economics often live in that bid-ask. Jump to originator takeaways.
• Rate-sheet engineers & secondary marketing: LLPA is the floor of risk-based pricing, not the ceiling. Overlay models that meaningfully differentiate within a single LLPA cell — by DTI, by state, by channel — capture the residual. Models that flatten to the grid surrender it to a competitor’s overlay book. Jump to rate-sheet takeaways.

Risk-based pricing on conventional mortgages is, formally, a one-page table. Fannie Mae’s Loan-Level Price Adjustment (LLPA) Matrix and Freddie Mac’s Exhibit 19 set the upfront fee any GSE-eligible loan pays at acquisition, indexed almost entirely on two attributes: credit score and loan-to-value. A handful of attribute add-ons — investment property, second home, condo, manufactured housing, high-balance — round out the grid. The May 2023 redesign that consolidated Fannie and Freddie under FHFA’s “aligned pricing framework” preserved this two-dimensional spine.

We were curious how complete that spine actually is. Real-world loan performance varies continuously across attributes that aren’t in the grid. Within a single LLPA cell — one FICO band crossed with one LTV bucket — modification rates aren’t flat. They vary 8.8× across DTI sub-bands and 21× across states. That gap between the grid’s discrete bands and the underlying continuous risk function has to flow somewhere; in practice it migrates to mortgage servicing rights pricing, to spec pool pay-ups, to lender-level overlays, and finally to borrower-facing rate spread. The map of where the risk flows is useful for almost everyone whose pricing depends on understanding it.

This piece works through the empirical evidence and then walks the implications for each market participant.

The cells inside the cell

One scoping note before we work the cell: Fannie and Freddie each maintain separate LLPA matrices for purchase, rate/term refinance, and cash-out refinance, with materially different fee structures across the three. Everything that follows works through the purchase grid; the analytical framework — the gap between what discrete bands price and what continuous risk does — applies analogously to the rate/term and cash-out grids, but the cell-level math does not transfer.

Start with one of the densest cells in the purchase grid: Fannie purchase loans at 720-739 FICO and 75.01-80.00% LTV. Under the 2017 grid that was in force when 2018-2020 origination vintages closed, every loan in this cell paid a 75 bps base LLPA — about $2,250 on a $300,000 loan. The cell contained roughly 82,800 owner-occupied single-family Fannie purchase originations across those three vintages.

Within that cell, every loan was, by construction, identically priced for credit risk under the LLPA grid. The grid says these loans are the same.

The data disagrees.

Modification rate rises 8.8× as DTI moves from <25 to 45-50. The DTI <25 group modifies at 0.25%. The DTI 45-50 group modifies at 2.20%. The same FICO. The same LTV. The same 75 bps fee. The same vintage. By every dimension the grid prices, these are identical loans.

This is the variable the GSEs themselves almost priced. In the May 2023 redesign, a DTI > 40% LLPA was on the announced grid. Three weeks before its August 1, 2023 effective date, the FHFA withdrew it. The political pressure against pricing DTI explicitly was substantial enough to kill a fee the agencies had already published.

We tested whether the DTI gradient might just be a re-statement of “higher-risk borrowers get higher rates, and rates predict performance.” It isn’t. We binned the same cohort by the actual note rate borrowers received at origination and checked whether DTI still predicts mod rate within each rate band:

Rate band	DTI <30	30-34	35-39	40-44	45-50	DTI gradient
<3.50%	0.12%	0.26%	0.60%	0.35%	0.60%	5.1×
3.50-3.99%	0.25%	0.63%	0.92%	1.18%	1.72%	6.8×
4.00-4.49%	0.52%	0.73%	1.03%	1.95%	2.09%	4.0×
4.50-4.99%	1.14%	1.10%	1.72%	3.26%	4.88%	4.4×
≥5.00%	1.01%	2.73%	4.66%	2.79%	5.26%	5.2×

Within every rate band, DTI predicts modification 4-7×. The DTI gradient is independent of rate-based pricing. Both the rate environment and the borrower’s debt burden carry information; the grid prices neither, the rate channel captures the first, and the rest moves elsewhere.

State variance is the bigger gap

State-level performance variance is even larger than DTI.

Within the same LLPA cell, Nevada loans modify at 2.91%, while Missouri loans modify at 0.14% — a 21× spread. California (1.98%), New York (1.99%), Massachusetts (1.81%), and Washington (1.70%) cluster near the top. Missouri, Alabama (0.21%), Tennessee (0.24%), Idaho (0.30%), and Pennsylvania (0.30%) anchor the bottom. The pattern is vintage-robust: testing the same cell across 2014-17, 2018-20, and 2021-23 origination windows produces consistent state rankings. California is in the top three across every era; Missouri is in the bottom three.

This is geography doing real work. Volatile-HPA states — Sun Belt, West Coast — show consistently higher modification frequencies even at otherwise identical loan characteristics. The natural-disaster correlation matters too: hurricane exposure in Florida, wildfire exposure in California and Oregon, regional employment concentration in oil-dependent or tech-dependent metros. None of this is in the LLPA grid.

But state pricing is the political third rail of grid design. “Why does my Florida mortgage pay more than a Texas mortgage at the same FICO and LTV?” is not a question the FHFA wants to answer in front of Congress. So the variance lives in pool-level pricing instead, where bond markets quote different premiums for high-concentration NY pools versus FL pools without naming individual borrowers.

What about the band edges?

If the grid is doing its job, the edges of each band should price step changes in risk. They mostly don’t.

The 75% LTV cliff at the 720-739 FICO band carries a 25 bps step under the 2017 grid — from 50 bps below 75 to 75 bps above. Performance through the boundary:

Modification rate is essentially flat. The pricing step is a pure policy artifact — useful for the GSE’s administrative simplicity, not anchored in any observable performance discontinuity. And the borrower behavior around the boundary is striking: 67% of the cohort lands at LTV 74.5-75.0%, immediately under the boundary, paying the 50 bps fee. The bins just above (75.01-76.0%) are nearly empty. Borrowers and originators structure down payments to dodge the pricing step.

The 80% LTV cliff at the 680-699 FICO band is even more extreme:

Ninety-one percent of the cohort lands at LTV 79.5-80.0%. Zero loans in the 80.01-81.0% range. The boundary repulsion is total: once a borrower can’t quite hit 80%, they jump straight to 82%+ and pay PMI for years. The 80% cliff differs from 75% in being anchored to the PMI threshold rather than to grid policy alone — there is a structural risk reason for the price to step (above 80%, MI absorbs first-loss credit), and the LLPA on the high-LTV side actually drops 25 bps because the GSE is partially insulated by the MI. But the behavioral distortion the structure creates is enormous, and the underlying loan-level performance is roughly flat through it.

What’s priced and what isn’t

The cleanest way to see what the grid prices and what it leaves on the table is to lay it out as a heat map. Each row below is a single LLPA-priced cell — a fixed FICO/LTV combination at one purchase-grid fee — across five cells that span the credit spectrum, from the 50 bps tier to the 175 bps tier. Each column is a DTI band the grid is silent on. The color shift across any row is the residual the LLPA fee does not capture:

Read by row. Every cell in a single row paid the same LLPA at acquisition; every cell modified at a different rate. The DTI-driven within-cell gradient — roughly 4× to 8× across all five priced cells — is structurally present in every credit tier, not a peculiarity of the headline cell. The reverse read works too: scan down any column and the LLPA fee climbs from 50 bps to 175 bps while the within-DTI mod rate moves in the same direction the grid intended. The grid is doing real work; it is also leaving a comparable amount of work undone inside every cell.

Now compare that picture against the variance summary. The dimensions in the grid show modest within-band variance because the bands are reasonably calibrated. Point-FICO variance within a band is roughly 1.1× on modification rate. Point-LTV variance within a band is similar. The grid is doing fine at the spine. The dimensions not in the grid are an order of magnitude larger:

Channel (retail vs correspondent vs broker) is roughly 1.8× — broker and correspondent loans at the same FICO/LTV modify at ~1.6%, retail at 0.9%. DTI is 8.8×. State is 21×. None of these dimensions live in the grid.

The picture is consistent: the grid prices what it can defend politically and administratively, and the rest of the risk function — the dimensions where pricing would face fair-lending challenges, borrower-confusion problems, or grid-complexity backlash — gets squeezed into other layers of the market.

Where the unpriced risk actually flows

The unpriced variance doesn’t evaporate. It migrates, predictably, to four other pricing layers.

Mortgage servicing rights price the prepayment risk that DTI and employment patterns predict. An MSR portfolio with high concentration of DTI-44 borrowers carries different prepay assumptions — and a different strip valuation — than an otherwise-identical portfolio of DTI-28 borrowers. The MSR market priced this distinction long before the FHFA proposed the DTI LLPA, and continues to price it after the LLPA was withdrawn.

Specified pool pay-ups price the geographic concentration. A 100% Florida or 100% California pool trades at a different spec premium than a geographically diversified pool, even when the FICO/LTV strata are identical. The pool’s “story” — the narrative the trader uses to defend the pay-up — is exactly the state-level performance variance the LLPA leaves unpriced. Spec pool buyers are paying for the geographic risk the grid doesn’t capture.

Lender overlays price the channel and underwriting quality the grid doesn’t differentiate. Aggregators apply seller-specific stipulations to loans coming from particular originators with particular performance histories. Banks running portfolio lending overlay manual underwriting standards on top of DU/LP that effectively price DTI, asset reserves, employment stability, and other dimensions the grid passes through silently.

Borrower-facing rate spread is where this ultimately lands for the consumer. The 75 bps LLPA the grid charges on the 720-739/75-80 cell is the same for everyone. The rate they actually pay differs because the lender’s pricing model layers in everything the grid doesn’t — and the differences in rate spread across borrowers in that cell encode the DTI/state/channel variance the grid is silent about. The borrower sees the rate, not the layers. They effectively pay for state risk through a regional pricing concession, for DTI risk through a margin adjustment, and for channel quality through a markup decision made before the rate sheet ever quotes.

Implications for each participant

For borrowers, the practical takeaway is that LLPA isn’t the full pricing story. The actual rate they pay encodes their state, their DTI, their employment profile, and their lender’s overlay decisions on top of the grid LLPA. Shopping across lenders is shopping across these overlay policies as much as across origination fees.

For originators, especially mid-size lenders and aggregators, the gap between LLPA-priced risk and continuous-risk-priced loans is structural margin opportunity. Loans that perform better than the grid implies — the 720-739 FICO Missouri retail borrower with 28 DTI — pay the same LLPA as the otherwise-identical Nevada broker correspondent loan at DTI 47 that modifies 15-20× more often. The originator who can identify, retain, and selectively pass through the lower-risk loans captures the residual; the originator who can’t gets adversely selected on the higher-risk loans. The bid-ask between identifying the gap and pricing it (in rate sheets, in marketing, in retention) is where mid-size lender economics often live.

For MBS investors, the punch line is that the LLPA grid is not a substitute for understanding pool-level risk composition. Two pools at identical FICO/LTV strata can have meaningfully different expected loss and prepay profiles based on state concentration and DTI distribution. Pricing those pools off the LLPA-implied risk number alone systematically undervalues geographically-diverse pools and overpays for state-concentrated pools. The spec pool market has long known this; it’s why pay-ups exist.

For MSR owners, the DTI and employment patterns the grid omits are exactly the prepay predictors that flow into strip valuation. Servicers running large flow-buy MSR programs are effectively buying baskets of unpriced LLPA risk — the risk the grid abstracts away from the bond market and pushes onto the cashflow timing of the servicing strip.

For rate-sheet engineers at lenders and aggregators, the takeaway is that the LLPA grid is the floor of risk-based pricing, not the ceiling. Overlay models that meaningfully differentiate within a single LLPA cell — by DTI, by state, by channel — capture the residual. Models that flatten to the LLPA grid surrender it.

A descriptive map, not a critique

This piece isn’t an argument that the GSE grid should be richer. There are real structural reasons it isn’t — fair-lending exposure on geographic and employment-based pricing, simplicity constraints on borrower comprehension of the upfront-fee disclosure, political infeasibility on state-level differentiation. The May 2023 redesign that consolidated Fannie and Freddie under FHFA’s aligned framework was already a politically expensive exercise; expanding the grid to incorporate DTI, state, or income-type pricing would require a much larger fight than the agencies have appetite for.

But participants in the market need a map of where the risk flows. The grid handles roughly a third of the risk function explicitly; the other two-thirds run through MSR strips, spec pool pricing, lender overlays, and borrower-facing rate spread. Understanding which layer is doing the work on a given dimension is what separates accurate pricing from approximate pricing.

The grid prices what is politically tolerable. The market prices the rest.

---

Methodology. Loan-level performance data sourced from Fannie Mae Single-Family Performance Data; LLPA fee structure read from the published Fannie Mae purchase LLPA Matrix in force at each cohort’s origination window. The analysis throughout uses the purchase grid; Fannie and Freddie maintain distinct LLPA matrices for rate/term and cash-out refinance loans, and cell-level fees do not transfer across grids. All consolidated via The Mortgage LLM’s analytics instance. The headline cohort for DTI and state analysis is Fannie purchase originations from 2018–2020 vintages at a FICO of 720–739 and an LTV of 75.01–80.00%, owner-occupied one-to-four-family residences, observed through the 2025 reporting snapshot (~82,800 loans). Multi-vintage robustness was tested by re-running the same cell across the 2014–2017, 2018–2020, and 2021–2023 origination windows. Cell generalizability was tested across four additional LLPA cells (700–719 / 75–80, 720–739 / 70–75, 680–699 / 75–80, and 760–779 / 75–80) at the 2019 vintage. Modification rate is the share of loans in the cohort that have been reported as modified at any point through the latest disclosure. Independence of the DTI gradient from rate-based pricing was tested by stratifying the cohort by the borrower’s actual note rate at origination and re-running the gradient within each rate band. Survivorship bias affecting older origination cohorts is noted but not corrected — it does not affect the within-cell relative gradients the piece reports. Informational, not advice.