{"model":"llpa_overlay","model_version":"llpa_overlay_v1","trained_at":"2026-07-01T00:22:41Z","target":"cumulative_loss_bps","target_definition":"is_credit_event * 3000 + IF(modified='Y', 500, 0). Flat 30% LGD anchor + 500 bps modification servicing-cost proxy. Source-specific credit events: Fannie {03,06,09,15}, Freddie {03,09}. Right-censored loans + Freddie admin-removal codes (02,06,96,97,98,99) excluded.","universe":"GSE conventional conforming acquisitions (FNM_SFP + FRE), purchase + refi, owner-occupied, originated 2014-2022. Per-source loan_purpose + occupancy code sets harmonized; FICO clamped to [300,850]; LTV clamped to [1,105] (drops sentinels + HARP refi).","vintages_trained_on":[2014,2015,2016,2017,2018,2019,2020],"vintages_oot":[2021,2022],"n_train":11990779,"n_oot":1245730,"n_holdout":500000,"stage1_n_iter":143,"stage2_n_iter":500,"calibrator_source":"holdout-fit (Phase 3)","wal_factor":0.06,"bps_translation":"upfront_bps = max(combined_calibrated_bps × WAL_FACTOR - stage1_bps × WAL_FACTOR, 0). WAL_FACTOR = 0.06 anchors a 7-year WAL conforming purchase at ~6 bps upfront per 100 bps cumulative 60-month loss. Stage 1 baseline component is subtracted so the overlay represents only the residual the LLPA grid leaves unpriced; floored at 0 because lenders don't issue LLPA credits.","phase4_verdict":"NARROW","phase4_rationale":"Magnitude calibration off by 61% (gate is 15%), but rank-order AUC 0.683 > 0.60 says the model successfully orders within-cell credit-event risk. Ship as ordinal risk-rank model — surface decile labels (Q1-Q10) instead of raw bps until OOT selection bias can be properly addressed.","phase4_auc_overall_overlay":0.6832766480828701,"phase4_auc_overall_combined":0.7602278911509266,"phase4_decile_gate_pass_rate":"0/10","phase5_di_verdict":"ACCEPTABLE","phase5_di_rationale":"AIR is clean; Q1-Q4 effect after controls is +0.0678 bps (below the 0.10 bps materiality threshold). 1 of 4 state proxies show |r|>0.40 with minority concentration, suggesting partial encoding — but the overlay magnitude is too small for this to translate to actionable disparate impact.","phase5_air_q4_over_q1":0.9413822375069978,"phase5_proxy_correlations":{"hpa_volatility_20yr":{"r":0.289109095592023,"p":0.039625178426364795},"unemp_volatility_20yr":{"r":0.32793333687310455,"p":0.01881245018874168},"state_disaster_risk_index":{"r":0.6292053134002613,"p":7.600460338556616e-07},"state_emp_hhi_latest":{"r":0.07632186842197507,"p":0.5945097382650423}},"band_thresholds":{"baseline_max":0.5110222683654208,"elevated_max":1.4682242011785034},"decile_cuts_nonzero":[0.024323982228460946,0.05057362457473868,0.09710639815911488,0.14756354073969444,0.20661863238168043,0.302750882889162,0.44221062787269505,0.6240834430540038,1.0334443820809025],"expected_inputs":["borrower_fico","original_ltv","loan_purpose","property_state","gse","original_cltv","original_upb","dti","occupancy","property_type","number_of_units","product_type","channel","original_interest_rate","canonical_seller_name","first_time_homebuyer","number_of_borrowers"],"caveats":["Magnitude calibration is approximate (Phase 4 NARROW verdict): the model rank-orders within-cell credit-event risk reliably (AUC 0.67-0.74) but undershoots realized loss magnitudes on the 2021-2022 OOT cohort by ~60%. Use `overlay_band` and `overlay_decile` (ordinal outputs) for operating decisions; `overlay_bps_raw` is informational only.","Fair-lending DI audit (Phase 5) verdict: ACCEPTABLE. AIR(Q4/Q1) = 0.946 (4/5ths rule clean). One state-proxy feature (FEMA disaster index) correlates with minority concentration r = +0.63 across states; documented as a known limitation, with a re-fit playbook ready if a future fair-lending exam pushes back on the disaster feature.","Within-cell residual model: predicts what the LLPA grid leaves unpriced, after the grid's own FICO × LTV charge. A 'low' band means the grid already prices this loan correctly — not that the loan is low credit risk."]}