themortgagellm^™

What it does

For a given loan that's eligible for both GSEs, predicts which channel produces lower expected loss. S-learner architecture: scores the loan with source='FNM_SFP' and source='FRE' against the existing repurchase v4 + EPD 12-mo / 24-mo / 36-mo models, then computes per-channel repurchase + delinquency probabilities, expected dollar losses, the differential, and a recommended channel.

No single-classifier AUC. Channel Choice is a composition, not a separately-trained classifier — its accuracy inherits from the underlying Repurchase (0.72 OOT) and EPD 12/24/36-mo models (0.83 / 0.78 / 0.75 OOT). The counterfactual ("what would this loan have done at Fannie if we routed it to Freddie?") isn't observable in production data, so there's no ground truth for a per-loan routing-accuracy AUC.

Why it matters. A 5-10 bps expected-loss differential on a $300K loan is $15-30 per loan — modest single-loan, but for a mid-size shop running 100K originations a year that's $1.5-3M annually captured by routing optimally rather than ad-hoc. The model surfaces a dollar-quantified preferred channel per loan, so routing becomes data-driven instead of pricing-grid-of-the-week.

› Try it on the home page (Loan-level model scoring → Fannie vs Freddie channel choice)

API connector

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

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

{
  "borrower_fico": 720,
  "dti": 38,
  "original_ltv": 80,
  ...
  "loss_per_repurchase_usd": 200000,   // optional, default 200000
  "loss_per_epd_usd":         10000    // optional, default 10000
}

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

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