themortgagellm^™

Natural-language access to U.S. mortgage origination and performance data.

Enter to run · Shift+Enter for new line · new conversation

Explore additional capabilities

Explore lead generation angles

Check Fannie + Freddie eligibility

Ask in a language other than English

Ask about GSE guidelines

Conduct market research

Benchmark a lender

Score loans before funding

Track loan performance post-funding

Assess and action rep & warranty exposure

Loan-level model scoring

Trained gradient-boosting models scoring the same loan-feature payload:

Repurchase risk (v4) — probability of rep-and-warranty repurchase (zbc 06/96). For post-funding QC + R&W reserve setting. Trained on 2013-2023 GSE cohorts, tested on 2024-2025: AUC 0.72 (in-cohort random-split: 0.81).
Prepayment 12/24/36-mo — probability of prepayment (zbc 01) within first 12 / 24 / 36 months. For MSR valuation, pipeline lock-desk risk, fast-pay vs slow-pay pool composition. Trained on 2013-2023 GSE cohorts.
EPD 12-mo (v2) — probability of 60+ DQ within first 12 months. For pre-funding pricing / LLPA tier. Trained on 2013-2023 GSE cohorts, tested on 2024: AUC 0.83 (in-cohort random-split: 0.88).
EPD 24-mo — same target extended to 24 months. For mid-life risk pricing + reserve setting. Trained on 2013-2023, tested on 2024: AUC 0.78 (in-cohort random-split: 0.87).
EPD 36-mo — same target extended to 36 months. Captures full-cycle DQ risk including the refi-window peak. Trained on 2013-2022, tested on 2023: AUC 0.75 (in-cohort random-split: 0.86).
GNMA EPD — government-insured EPD model (FHA / VA / USDA-RD / PIH). Different feature schema (use agency, credit_score, ltv, ...). Trained on 2018-2023, tested on 2024: AUC 0.76 (in-cohort random-split: 0.81).
Fannie vs Freddie channel choice — S-learner over the repurchase + 12/24/36-mo EPD models; predicts which GSE produces lower expected loss for this loan, with full 12/24/36-mo delinquency curves.
Higher-Priced loan (HPML) — probability the loan prices into a Reg Z Higher-Priced Mortgage Loan (first-lien rate_spread ≥ 1.5pp or sub-lien ≥ 3.5pp or HOEPA). HMDA-style input. Trained on 2018-2023 HMDA originations, tested on 2024-2025: AUC 0.87 (in-cohort random-split: 0.93), ECE 0.0003.
Pull-through — probability an application closes as an originated loan (action_taken=1). HMDA-style input. For pipeline-hedge sizing, lead pricing, capacity planning, lender benchmarking. Baseline pull-through ~62%. Trained on 2018-2023 HMDA applications, tested on 2024-2025: AUC 0.92 (in-cohort random-split: 0.93), ECE 0.0004.
Credit Denial Probability — probability an application is denied for credit (action_taken=3). HMDA-style input. For pre-credit-pull screening, fair-lending self-assessment, counter-offer routing. Baseline denial rate ~16% in train (19% in 2024-25). Trained on 2018-2023 HMDA applications with class_weight=balanced, tested on 2024-2025: AUC 0.91 (in-cohort random-split: 0.91), ECE 0.0004.
Credit Approval Probability — positive-framing sibling of Denial; probability the application receives a lender approval decision (action_taken IN (1,2)). Same HMDA universe, same input schema. For LO triage + lead prioritization + consumer-facing positive framing. Baseline ~65% (train) / 62% (OOT). Trained on 2018-2023, tested on 2024-2025: AUC 0.94, ECE 0.0003.
Appraisal Waiver Probability (PIW / Value Acceptance / ACE) — probability a conventional conforming loan is granted an appraisal waiver instead of a full appraisal. GSE-input schema. Pre-DU/LPA heuristic for LO intake messaging, pricing strategy, pipeline planning. Trained on 2018-2023 GSE acquisitions (22.4M loans), tested on 2024-2025 (2.6M loans): AUC 0.85 (in-cohort random-split: 0.93), ECE 0.0008.

Same GSE input fields for repurchase + EPD models + channel choice. GNMA EPD and Higher-Priced loan models use their own input schemas — "load example" auto-fills the right shape for the selected model. Missing fields are handled by the model's native NaN routing.

Model: