mortgagetape^™

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

This article was written in collaboration with Optimal Blue, an Official AI Partner of Mortgage Tape, using Optimal Blue’s Competitive Data License Plus (CDL+), which covers approximately 35% of locked-and-hedged U.S. residential production.

📌 Quick takeaways by role

• Capital markets & secondary: The right way to size the refi pool is with the LLPA already charged, not against a purchase-blended survey rate. Doing that collapses the actionable pool from $291B (gross) to about $69B (net, +40 bps refi premium) — a 76% haircut. Prepay models that anchor to headline in-the-money numbers are overstating speeds. Jump to capital-markets takeaways.
• Loan officers & originators: Ignore the 2007–2021 book — 0.0% of it is in the money at any FICO. Your call list lives in 2023–2024 vintages plus the low-FICO tail of 2025+. Within 2023–2024 the whole FICO surface is live (23–31% ITM). 2025+ is bifurcated — 620–699 borrowers still clear 26–32% ITM, but 740+ borrowers priced too close to today’s market to work (15–16% ITM). LLPAs made lower-FICO orig rates high, so today’s par leaves more room. Jump to originator takeaways.
• Servicers & MSR desks: $2.5T of pre-2022 GSE UPB shows exactly 0.0% in the money — the deepest lock-in ever measured in this dataset. Whatever your S-curve says at the current rate node, this book will not move. Jump to MSR takeaways.

Three months ago we sized the refi pool at ~$289 billion — the loans sitting at least 75 bps in the money on a gross basis, measured against Freddie’s headline survey rate. That number was fair, but it was measured on the sticker side of the transaction — the rate borrowers would see on a rate sheet before an LLPA is added.

The pool the borrower actually clears is smaller. Once you charge the LLPA a refi would pay in the current pricing regime, the number falls to roughly $69 billion — a 76% haircut. The direction has been visible for a while; what’s new here is that Fannie SFP + Freddie SLLD tapes let us measure it loan by loan, matched to a live retail rate anchor from Optimal Blue’s CDL+ lock database, with home-price appreciation applied at the MSA level rather than the state.

This is that measurement.

What the pool actually looks like on paper

Run against the full active book (20.8M loans / $5.15T UPB, from Fannie SFP Dec 2025 + Freddie SLLD March 2026), benchmarked to Optimal Blue’s CDL+ retail lock median cut by FICO × current LTV × occupancy, the sticker side looks like this:

Threshold	Loans	UPB ($B)	Share of book
≥50 bps gross in the money	941K	$291	4.6%
≥75 bps gross	471K	$158	2.3%
≥100 bps gross	209K	$61	1.0%

That’s the biggest number an LO can reasonably brief. It also treats the market rate as if the borrower could take it as-is, which they can’t.

Why the number moves so hard on LLPA

Every conforming refi carries an LLPA — a rate/price add priced off the GSE grid at the borrower’s refinance profile. Fannie’s current grid puts base rate-term refi about 36 bps above base purchase (123 bps average vs 87 bps across the FICO × LTV surface). Our retail anchor is purchase-heavy, so the median understates a true refi rate by roughly that same 36 bps. Charge the missing LLPA and the pool moves fast:

Refi LLPA premium added	pct ≥50bp ITM	UPB ≥50bp ($B)
0 bp (gross, as-published)	4.57%	$291
25 bp (light refi lift)	2.29%	$142
40 bp (realistic base case)	1.15%	$69
75 bp (pessimistic — cash-out weight)	0.33%	$18

40 bps isn’t pessimistic — it’s roughly the base-refi vs base-purchase delta on the published grid, before layering cash-out surcharges (50-100 bps depending on cell) or investment-property adjusters (100-400 bps). The reason it compresses so hard is that in-the-money incentive at current rates is shallow. Most of the pool sits in the 50-100 bp band; a 40 bp headwind eats most of it. The 100+ bp threshold — the traditional “call the borrower” cutoff — shrinks from $61B to $7B once LLPAs are honestly charged. That is a rounding error against a $5.15T book. Any S-curve anchored to a purchase-median headline rate is overstating incentive at the current rate node.

Concentration: 2023–2024 and the low-FICO incentive inversion

Cut the same eligible book by origination vintage × FICO band, showing the share of loans in each cell that clear 50 bps gross:

Vintage	620-659	660-699	700-739	740-779	780+
2007-2013	0.0%	0.0%	0.0%	0.0%	0.0%
2014-2018	0.0%	0.0%	0.0%	0.0%	0.0%
2019	0.0%	0.0%	0.0%	0.0%	0.0%
2020	0.0%	0.0%	0.0%	0.0%	0.0%
2021	0.0%	0.0%	0.0%	0.0%	0.0%
2022	1.0%	1.2%	1.3%	1.2%	1.3%
2023	28.1%	28.0%	25.6%	23.1%	24.0%
2024	30.8%	30.6%	28.9%	24.7%	25.1%
2025+	31.6%	26.0%	20.9%	15.6%	15.4%

The pre-2022 book is dead. $2.5T of UPB across 15 years shows 0.0% in the money at every FICO band. Loans that priced at sub-5% coupons will not refinance at 6.5-7% — no borrower profile in the panel makes the math work. Most of the book sits on the wrong side of a 250-350 bp gap from par.

2023-2024 is uniformly hot. Every FICO band clears 50 bps ITM at 23-31% — roughly 1.6M loans, ~$1.5T UPB, the bulk of the actionable pool. Borrowers who locked at 6.5-8% between rate peaks, now above a market drifted to the low-to-mid 6s.

2025+ is bifurcated, and the FICO incentive inverts. 620-659 in 2025+ clears 31.6% ITM; 740+ collapses to 15-16%, because high-FICO 2025 borrowers priced too close to today’s market to leave room. Same inversion is visible inside 2023-2024 (28-31% at 620-699 vs 23-25% at 780+) — LLPAs made the lower-FICO orig rate higher, so the same drop in par translates to more bps of incentive. For servicer retention: your lower-FICO recent-vintage book is your churn risk, not your high-FICO book. The opposite of the usual mental model.

The lock-in wall

The 2007-2021 book isn’t just not refinancing at current rates — it’s structurally frozen. Those 15M+ loans locked at 3-5% need rates to fall 100-150 bps before the math works for any of them, and even then only for the most recent slice. Two second-order consequences:

1. Prepay speeds stay slow. Aggregate CPR is a UPB-weighted average across the whole book. Even if the in-the-money pool prepays at the fastest S-curve speed you can defend, the 2007-2021 wall drags aggregate speeds to low single digits. Model bottom-up, not as a rate-response curve on the whole book.

2. The actionable pool doesn’t replenish organically. 2023-2024 will fully burn out inside a year or two of any sustained refi wave. The 2025+ vintage is coming in at a much smaller in-the-money share because those borrowers priced closer to today’s market. Growth is a rate rally, not a pipeline.

Operational takeaways: pricing, retention, and where the pool actually sits

📊 For capital markets & secondary desks: charge the LLPA before you size the pool

The right pool for pricing and modeling is $69B at a realistic 40 bp LLPA charge, $18B at 75 bp (cash-out-weighted) — not $291B. Prepay S-curves anchored to a purchase-blended headline rate systematically over-attribute incentive at the current rate node. TBA vs spec pricing should reflect the vintage split: a 2024-heavy pool has real refi optionality, a 2020–2021-heavy pool does not, and pricing that averages the two miscalibrates both.

⚙️ For loan officers & originators: your call list is 2023–2024, and the incentive inverts by FICO

Lowest-credit-tier borrowers in 2023–2024 carry the biggest refi incentive — 28–31% ITM at 620–699 vs 23–25% at 780+ — because LLPAs made their orig rate higher. Retention math on a 660-FICO 2024 borrower is a different conversation than on a 780. Many lower-FICO 2023–2024 borrowers have also seen credit-score recovery that qualifies them for tighter LLPAs on the refi than on the original — a call worth making before their credit union or a D2C lender does.

💼 For servicers & MSR desks: two books, priced separately

The pre-2022 book (15M+ loans, $2.5T UPB) is a locked strip — 0.0% ITM at every FICO band, 250–350 bps from par. A 2020-vintage MSR strip and a 2024-vintage MSR strip should not price on the same S-curve or expected life: one is extension-risk with no meaningful prepay optionality, the other is short-duration with real convexity. Mark to a two-book model, not a single curve.

Post-script: what changes if AI doubles refi take-up

Morgan Stanley strategists, cited in Bloomberg last week, argued that AI-driven origination — the two-minute Better.com refi, the 30-minute Rocket app-to-lock — could roughly double the share of eligible borrowers who actually refinance when in the money, from ~30% historically to perhaps ~60%. The framing: a 30-year mortgage that behaves “closer to a floating-rate instrument that only floats down.”

Our LLPA-adjusted pool suggests two adjustments to that read.

The take-up rate is being applied to the wrong denominator. 60% take-up on the $291B gross pool implies ~$175B of forward throughput. 60% on the $69B LLPA-adjusted pool that actually clears is ~$41B. AI dissolves behavioral friction (paperwork, shopping, inattention). It does nothing to a 40 bp LLPA. The “eligible but inattentive” borrower gets unlocked; the “eligible in the raw calc but priced out at the cell level” borrower is untouched. Recomputing take-up on $69B is the honest exercise.

AI’s structural effect is the collapse of the cross-sectional FICO gradient, which is worse for MBS than a uniform 2x. The historic S-curve averages a FICO distribution where high-credit borrowers shop hard and low-credit borrowers refi lazily. Our vintage × FICO grid (above) shows the incentive runs the other way inside 2023–2024 — 28–31% ITM at 620–699 versus 23–25% at 780+, because LLPAs made lower-FICO orig rates higher. AI’s operational lift lands hardest on the cohort that historically didn’t act on that incentive. If AI flattens the FICO-take-up gradient, pools heavy in low-FICO 2023–2024 collateral could see speeds accelerate well beyond 2x, while pre-2022 pools remain locked — AI can’t manufacture incentive that isn’t there.

Practical read for prepay models: the AI adjustment isn’t a single global multiplier. It’s a FICO-dependent lift — largest at the low-FICO end of recent vintages, near-zero on the pre-2022 book. Spec pay-ups should widen on 2023–2024 collateral where the low-FICO share is elevated, and stay put on 2020–2021 collateral where AI has nothing to accelerate.

What this doesn’t measure

Not net-of-cost. A refi carries closing costs (2-3% of loan balance for most borrowers) plus opportunity cost against the remaining term. The $69B pool is the rate-eligible pool, not the transaction-worthwhile pool. Applying even a plain 24-month break-even recoup test on top of the LLPA-adjusted incentive would collapse it further; a chunk of the 50-99 bp cohort simply won’t clear break-even at typical closing costs regardless of the LLPA math. Call the true targetable pool somewhere meaningfully below $69B once you’re honest about both.

Not a call list. The pool is measured on the loan’s rate, LTV, and FICO. That’s what a public GSE disclosure carries. It doesn’t know the borrower’s DTI headroom, employment continuity, or intent to move. Every servicer’s retention model should be tighter than this, because they have those extra columns.

Methodology

Loan-level performance and origination data sourced from Fannie Mae Single-Family Performance Data (December 2025 complete-panel snapshot) and Freddie Mac Single-Family Loan-Level Dataset (March 2026 snapshot), consolidated via Mortgage Tape’s analytics instance. Active-book filters: no terminal-status flag (no zero-balance code, current on payments, no modification history), positive unpaid principal balance, borrower FICO present. Current LTV was computed by growing origination property value forward using FHFA MSA-level House Price Index (410 MSAs through Q2 2026), with a state-level HPI fallback for loans in non-MSA counties. Market-rate anchor: median 30-year fixed conforming lock rate from Optimal Blue’s Competitive Data License Plus (CDL+) service, retail-channel locks only, trailing 30 days from the most recent lock date, segmented by borrower FICO band × current LTV band × occupancy; market cells with fewer than 30 locks were excluded to avoid noisy corner cases. The LLPA sensitivity analysis applies a flat basis-point premium to the market rate across all cells; the 40 bp base case reflects the current Fannie Mae LLPA matrix’s average differential between rate-term refinance and purchase pricing. The pool figures reported are rate-eligible pools, not net of closing costs or break-even recoup. Informational, not advice.

Try it yourself. Ask Mortgage Tape: “what share of the 2024 Freddie book at 700–739 FICO is 50 bps in the money right now, MSA-adjusted?” — the underlying loan-level data and HPI grid are queryable directly in the platform.