mortgagetape^™

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

📌 Executive takeaways by role

• Rate-sheet engineers: Band-keyed overlay above the grid. Half the cohort (50.7%) lands in low — the grid already prices the loan correctly. The other half splits across baseline (37%), elevated (10%), and high (3%) — that’s where you add overlay. Rank order validates on OOT (AUC 0.67); raw bps under-predicts magnitudes by ~60% on rate-shock vintages and is labeled informational until v2.
• MSR strip owners: Per-loan decile (Q0-Q10) rank-orders within-cell credit-event risk. The 8.8× within-cell DTI variance documented in our LLPA-risk piece shows up cleanly in the model: predicted overlay rises monotonically from 0.07 bps at DTI<25 to 0.64 bps at DTI>50, an 8.7× spread that mirrors the empirical mod-rate finding. Use the decile to weight pool-level expected-loss decomposition.
• Secondary-marketing desks: Identify which loans within a similar grid cell warrant spec-pool placement. The model surfaces hidden risk within geographically diverse pools, and surfaces hidden quality within state-concentrated ones — same way the empirical paper said the residual flowed.

The premise (very brief)

Our May 2026 piece showed empirically that within a single LLPA cell — same FICO band, same LTV bucket, same base bps — modification rates vary 8.8× by DTI and 21× by state. The grid is the floor of risk-based pricing, not the ceiling, and the unpriced residual flows to MSR strip valuations, spec pool pay-ups, and lender overlays. That paper raised the diagnosis. This piece is the model.

What we built

A two-stage HistGradientBoosting regressor with isotonic calibration, trained on 11.99M GSE conventional conforming acquisitions (Fannie SFP + Freddie STACR), 2014-2020 origination vintages, with the 2021-2022 vintages reserved as the out-of-time validation cohort (1.25M loans). The architecture is the design decision that matters most:

Stage 1 trains on pass-through features that mirror the LLPA grid’s pass-through: FICO bucket, LTV bucket, loan purpose, occupancy, property type, units, product type, high-balance flag, subordinate-financing flag. Stage 1 learns whatever the grid’s implicit risk model would have predicted. This is the cell-mean baseline.

Stage 2 trains on the residual: target = realized 60-month cumulative loss minus Stage 1’s prediction. Stage 2’s feature set is the full universe — Stage 1’s features plus DTI, four state proxies (HPA volatility from FHFA HPI, unemployment volatility from BLS LAUS, FEMA National Risk Index, and BLS QCEW employment HHI), channel, origination rate, seller (top-30 + OTHER), first-time-homebuyer flag, and number of borrowers. By construction, Stage 2 cannot re-predict the grid. It only learns what the grid leaves unpriced — the within-cell residual.

After both stages train, isotonic regression calibrates the combined Stage 1 + Stage 2 output, fitted on a 500K state-stratified holdout never seen during training. A WAL factor (6 bps upfront per 100 bps cumulative loss, anchored on 7-year WAL conforming) translates the calibrated cumulative-loss prediction into an upfront flat bps charge. The overlay output is max(combined_calibrated × WAL − Stage 1 baseline × WAL, 0) — floored at zero because lenders don’t issue LLPA credits.

What the model produces

For each loan, the API returns an ordinal band and a decile:

Band	% of cohort	What it means operationally
low	50.7%	The grid prices this loan correctly. No overlay warranted.
baseline	36.9%	Grid is roughly correct. Modest overlay (or none) depending on appetite.
elevated	9.9%	Grid likely under-prices. Consider a 5-15 bps overlay or tighter UW controls.
high	2.5%	Grid materially under-prices. Recommend 15+ bps overlay and enhanced UW review.

The heatmap above is the load-bearing capital-markets view: average predicted overlay (bps) per FICO × LTV cell across the 2021-2022 OOT cohort. Green cells are where the LLPA base fee prices credit risk correctly; red cells are where the model says the grid materially under-prices. The mispricing concentrates in high-LTV low-FICO cells (top-right of the grid) — cells where the LLPA base fee already tilts against the borrower, but not enough. That’s the residual a lender’s overlay schedule needs to price, and it’s the composition signal a pool-level analyst reads first.

The second capital-markets view: average overlay by state-level risk composite (the z-sum of the model’s four state proxies — HPA volatility, unemployment volatility, FEMA disaster index, employment concentration). The LLPA grid does not price state-level macro risk directly — this quartile gradient is entirely captured by Stage 2 overlay. For spec-pool execution, the practical implication is that a Q4-heavy pool priced against a Q1-average TBA is mispriced by a few bps on this signal alone — which stacks with any FICO/LTV mispricing the pool inherits from the cell heatmap above.

The decile (Q0-Q10) gives finer granularity within the nonzero subset — useful for portfolio-level risk decomposition and within-cell rank-ordering of similar loans. Q0 = “no overlay needed”; Q10 = top 10% of the model’s residual-risk distribution.

A raw bps estimate (overlay_bps_raw) is also surfaced in the response, but flagged informational only — see Considerations below for why.

How to operationalize

For rate sheets: key off the band, not the raw bps. The band map is stable across the holdout and OOT; the raw bps under-predicts magnitudes on the OOT cohort by ~60%. A defensible v1 lender overlay schedule looks like {low: 0, baseline: 0-5 bps, elevated: 10-15 bps, high: 20-30 bps} — anchored on your shop’s appetite and absorption capacity, not on the model’s continuous output.

For MSR strip valuation: use the per-loan decile to weight pool-level credit-event expectations. Loans in Q8-Q10 carry roughly 4× the empirical modification rate of loans in Q1-Q3 (within the same LLPA cell), which translates cleanly into a DTI-and-state-adjusted multiple haircut.

For spec pool execution: the model surfaces hidden risk in pools that look clean by grid-cell composition, and hidden quality in pools that look risky. A state-concentrated pool with low average DTI and clean state proxies will skew toward low and baseline; a geographically-diverse pool with stretched DTIs will skew toward elevated and high. That’s the directionality the LLPA-risk paper predicted; the model now quantifies it per loan.

Considerations, guardrails, and the NARROW verdict

The most important thing to know about this model is what the Phase 4 out-of-time validation produced — and why we’re shipping anyway.

Rank order works. Magnitude doesn’t. On the 2021-2022 OOT cohort, overlay AUC is 0.67 and combined-prediction AUC is 0.74 — solidly above the “useful rank-order signal” threshold. But the calibration error is 62% (gate is 15%): the model under-predicts realized loss magnitudes by roughly that much across all 10 decile bins.

The decile chart above is the visual of the NARROW verdict. Predicted-loss bars (dark) rise monotonically across deciles, matching the realized-loss bars (red) in rank order — the model’s decile assignment correctly identifies higher-credit-event loans. But realized loss in the top decile is roughly double the calibrated prediction. This is why pricing decisions key off the band, not the raw bps.

The mechanism is clear. We stratified OOT against a matched-seasoning subset of the training cohort (loans resolved within 36 months of origination). In that matched window, 2021-2022 vintages had 5× the train cohort’s realized loss rate — the model didn’t undershoot because it overfit; it undershoot because 2021-2022 originations genuinely behaved worse than 2014-2020 originations after controlling for seasoning. Rate-shock, post-stimulus DTI drift, and the rate-and-term refi mix shift weren’t in the training distribution. v2, retrained when 2025-2026 reporting-year data adds 2-3 more years of observation to the 2021-2022 vintages, should narrow this gap substantially.

The honest framing: ship rank order, label magnitudes informational, retrain when the data matures. We call this the NARROW verdict in our internal validation framework.

Fair-lending audit. The Phase 5 disparate-impact audit (state-level, against Census ACS 2023 5-year estimates) returned ACCEPTABLE. AIR(Q4/Q1) = 0.946 — within the EEOC 4/5ths rule. The regression effect of state minority concentration on overlay bps, after controlling for FICO/LTV/DTI, is +0.067 bps (immaterial). One proxy — the FEMA disaster index — correlates with minority concentration at r = +0.63 across states; we document this as a known limitation and have a re-fit playbook ready if a future fair-lending exam pushes back on the disaster feature.

Assumptions baked in. Flat 30% LGD across credit events (anchored to the industry midpoint; per-event severity is too noisy to learn from the data we have). Right-censored loans excluded from training (only loans with observed outcomes by 2025 contribute to the target). High-balance flag punted to v2 (requires a county-conforming-limit lookup; modal value False is ~94% of the cohort, so the prior is near-neutral). The state proxies use 20-year vol windows for HPA + unemployment, 10-year window for employment HHI, and the December 2025 FEMA NRI vintage.

How to use it today

The model is live at /api/score_llpa_overlay. Required inputs: borrower_fico, original_ltv, loan_purpose, property_state. Everything else is optional with sensible cohort defaults. Response includes the band, decile, raw bps (informational), recommendation text, and a structured caveats array surfacing the NARROW + DI audit notes for downstream compliance documentation.

Full request/response shape lives in the model card. The companion piece for a less-technical audience — covering the same model from a workflow lens — is Beyond the grid.