themortgagellm^™

What it does

Gradient-boosting classifier rating the probability that a conventional conforming purchase-originated loan will be flagged as Home Possible — Freddie Mac's 3% down, 80% AMI affordable program (equivalent scope to Fannie's HomeReady). Trained on Freddie's full SLLD tape, 2018-2022 originations (4.55M purchase loans, primary residence only), calibrated on 2023-2024, tested on 2025 (584K loans): OOT AUC 0.941. Fair-lending AIR audit ACCEPTABLE across race (0.806), ethnicity (0.810), and sex (0.974) — passes the 4/5ths rule on all dimensions.

Why it matters. Home Possible is Freddie's counterpart to Fannie HomeReady: 3% down, reduced MI, income-capped at 80% AMI, supports community seconds and DPA. Same operational value as the HomeReady model — LO routing, MSR pool composition, capital-markets pipeline sizing. Notably: the Freddie model was retrained on the full Freddie SLLD tape (not just the STACR CRT-referenced subset), so it reflects Freddie's actual origination distribution instead of a curated CRT slice.

What it is NOT. Same disclosure posture as HomeReady: propensity signal only, NOT underwriting, NOT product steering. Home Possible eligibility is verified by Loan Product Advisor (LPA) — income limits, first-time buyer homebuyer-education requirement (in most cases), occupancy checks, and program-specific overlays live inside Freddie's AUS. Using this score to steer borrowers AWAY from Home Possible is an ECOA violation.

Universe note. Only scores conventional conforming purchase originations against Freddie's SLLD tape (~46M loan history 1999-2025). Refinance loans (use Refi Possible-specific analytics if available), Fannie loans (use the HomeReady model), FHA/VA/USDA, jumbo, and investor loans are out of scope. Same top features as HomeReady: mi_pct dominates, followed by upb, cltv, dti. Model version v1.2, retrained 2026-07-24 on the SLLD-expanded cohort.

› Try it on the home page (Loan-level model scoring → Home Possible Propensity)

API connector

Programmatic access. Calibrated probability + risk band + operating recommendation in the response.

POST /api/score_homepossible_propensity
Content-Type: application/json

{
  "fico": 720,                    // required (credit_score at origination)
  "dti": 38,
  "ltv": 97,                      // Home Possible goes to 97% LTV
  "cltv": 105,
  "upb": 320000,
  "mi_pct": 25,
  "fthb": "Y",
  "state": "IL",
  "property_type": "SF",          // SF / PU / MH / CO / CP
  "units": 1,
  "channel": "R",
  "num_borrowers": 2,
  "loan_term": 360
}

Response includes home_possible_probability, quantile band (P99 / P95 / P90 / P75 / P50 / P25 / P10), and a disclosure that this is a pre-LPA heuristic. Not for steering.

Schema reference (request / response shape): GET /api/score_homepossible_propensity/schema

Model metadata (training cohort, AUC, calibration): GET /api/score_homepossible_propensity/info