themortgagellm^™

What it does

Gradient-boosting classifier rating the probability that a Ginnie-eligible purchase origination (FHA / VA / USDA-RD / PIH-184) will layer Down Payment Assistance. Trained on 6.2M purchase originations (2018-2022 vintages), isotonic-calibrated on 2023 (~900K loans), tested on 2024-2025 out-of-time (1.9M loans): OOT AUC 0.741. Fair-lending AIR audit ACCEPTABLE across race (0.939), ethnicity (0.980), and sex (0.973) — well within the 4/5ths rule. Uses borrower + product + geography signal only; lender identity is deliberately excluded because HFA program-participation churn was driving a 20-point OOT AUC drift in the ablation run.

Why it matters. DPA is the operational lever that lets LMI borrowers meet FHA's 3.5% minimum down (or bridge the gap on VA/USDA with cash-to-close constraints). The universe of DPA programs is fragmented — state HFAs, county programs, city/employer programs — and each has its own income cap, purchase-price cap, homebuyer education requirement, and DPA-source acceptability rules (per HUD's Mortgagee Letter constraints on gov-entity vs. seller-funded DPA on FHA). A pre-intake propensity score lets you route the likely-DPA borrower to an HFA/DPA specialist LO immediately instead of losing days learning it mid-file. Capital markets: for MSR pricing on GNMA pools with material DPA content, HFA-funded seconds prepay differently from cash-out refinances, so pool composition matters.

What it is NOT. This is a propensity signal, NOT a DPA-eligibility determination and NOT an underwriting call. Program eligibility (income, occupancy, homebuyer education, property location, DPA source) is state/program- specific and must be verified against the actual HFA program guidelines. Using this score to steer borrowers AWAY from DPA to protect origination margin would be an ECOA violation.

Universe note. Only scores GNMA-eligible purchase originations (FHA/VA/USDA-RD/PIH-184). Conventional (Fannie/Freddie), refinance, and jumbo are out of scope. Fannie/Freddie loan-level tapes don't publish a DPA flag — the closest conventional proxies are pct_homeready (Fannie MBS) and HFA Preferred/Advantage program pool shares. Top features by importance: agency (FHA carries the bulk of DPA volume), cltv (DPA layers push CLTV above 100%), tpo_type (retail vs. TPO channels reach HFA networks differently), ltv, upfront_mip (FHA-specific), and state (state HFA program availability).

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

API connector

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

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

{
  "fico": 680,                  // credit_score at origination
  "dti": 42,
  "ltv": 96.5,                  // FHA max is 96.5%
  "cltv": 105,                  // DPA layered as a second → CLTV over 100%
  "upb": 285000,
  "rate": 6.75,
  "loan_term": 360,
  "upfront_mip": 1.75,
  "annual_mip": 0.55,
  "num_borrowers": 1,
  "fthb": "Y",
  "property_type": "SF",        // SF / CO / MF / MH etc.
  "agency": "F",                // F=FHA, V=VA, R=USDA-RD, N=PIH-184
  "refinance_type": "N",        // N for purchase
  "tpo_type": "R",              // R=retail, B=broker, C=correspondent
  "state": "GA",
  "msa": 12060                  // optional; MSA code
}

Response includes dpa_probability, quantile band (P99 / P95 / P90 / P75 / P50 / P25 / P10), and a disclosure that this is a pre-intake heuristic, not a DPA-eligibility determination or product-steering tool.

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

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