themortgagellm™Did 2018-2020 Fannie purchase LLPAs exceed realized credit losses?
“Did 2018-2020 Fannie purchase LLPAs exceed realized credit losses on the 740+ FICO / 70-85% LTV cohort?”
Response
On the cohort you specified — Fannie Mae purchase originations, 2018-2020 vintages, FICO ≥ 740, original LTV 70-85% — the LLPA-vs-realized-loss comparison:
| Vintage | Loans | Loan-weighted avg LLPA | Realized credit loss (% of orig UPB) | Spread (bps) |
|---|---|---|---|---|
| 2018 | ~1.04M | 44.2 bps | 0.030% (3.0 bps) | +41.2 |
| 2019 | ~0.96M | 43.8 bps | 0.022% (2.2 bps) | +41.6 |
| 2020 | ~1.41M | 40.1 bps | 0.013% (1.3 bps) | +38.8 |
| Pooled (2018-2020) | ~3.41M | 42.0 bps | 0.021% (2.1 bps) | +39.9 |
Bottom line: The LLPA Fannie charged on this cohort exceeded realized credit losses by ~40 basis points — about 20× the realized loss rate. By any reasonable definition, the LLPA earned its risk premium on this segment in this window.
Why the spread is so wide. Three forces compressed credit losses on the 740+ FICO 2018-2020 vintages:
- Benign credit cycle. The cohorts avoided a serious unemployment shock — COVID-19 was met with rapid forbearance interventions that kept default rates suppressed.
- The refi wave (2020-2021) prepaid out the weaker tail. By 2025 only ~12-49% of the original cohort is still active, and the surviving loans skew lower-risk.
- The lock-in wall (2022-present) kept move-up sellers in place, holding home prices up. Even loans that became distressed had positive home equity to absorb most of the loss.
The capital-cost overlay. The 40-bps spread is gross of capital cost. A conforming GSE loan ties up roughly 4-5% of GSE capital at a ~10% required return — about 40 bps/year of opportunity cost. Over an effective WAL of ~6-8 years for this cohort, the capital cost works out to ~250-320 bps cumulative. So the LLPA collected on this cohort, while it exceeded realized credit losses by a wide margin, was still below the cumulative capital cost. This is consistent with GSE pricing being a thin commercial margin over the combined credit-loss + capital-cost wedge.
Three honest caveats:
- Severe survivorship bias for the 2018 cohort. Only ~12% of originated 2018 loans were still active in the 2025 reporting snapshot. The avg LLPA above is computed on surviving loans — which slightly overstates the original cohort’s LLPA distribution (since refiers tended to be lower-LLPA loans that had refi incentive).
- The realized-loss field captures only credit-event losses on loans that hit a 60+ DPD modification or foreclosure. Recently delinquent loans still on book are not captured here. For 740+ FICO this gap is small (< 5 bps); for sub-680 FICO it would be larger.
- One macro cycle is not the long run. The 2018-2020 vintage avoided a serious downturn. A 2007 cohort at the same FICO/LTV profile would have shown realized losses 30-60× higher. The LLPA is priced for an averaged-across-cycles loss distribution; the 2018-2020 result tells us this cohort happened to fall on the favorable side of that distribution.
Reference: Fannie LLPA Matrix in effect July 2017 through March 2022, base grid for FICO 740-759, 760-779, 780+ at LTV 70-75% / 75-80% / 80-85% cells. Realized losses from mortgage.fnm_sfp_raw.cumulative_credit_event_net_gain_or_loss at the latest reporting period per loan (2025-12 or earlier zero-balance event). Capital-cost overlay assumes 40 bps/year × 7-year weighted average life.