Reflect 3 Attempt 02

Credit Risk Lending Committee Skill

API Workflow

Base URL from environment (not localhost). Start with:

• GET /api/manifest — lists endpoints, benchmark versions, record counts
• GET /api/health — confirms service status
• GET /api/policies — risk-rating thresholds, CDFI factor tables, stress formulas, CRE weights, concentration rules

Key data endpoints:

• GET /api/branches / GET /api/branches/{id} — branch metadata, lending capacity, sector ceiling, CRE limit, state, institution type
• GET /api/branches/{id}/metrics — quarterly metrics (NPA, total loans, delinquency, deposits)
• GET /api/branches/{id}/loans — loan portfolio with ratings, DSCR, LTV, payment status, FICO, debt-to-asset, liquidity
• GET /api/branches/{id}/sector-exposures — per-sector current exposure, limit_pct, grandfathered flag
• GET /api/branches/{id}/applications — pending applications with DSCR, LTV, FICO, DTI, guarantor, relationship, SBA, bankruptcy
• GET /api/benchmarks/fdic/q4-2024 — FDIC noncurrent/delinquency benchmarks
• GET /api/benchmarks/ncua/q1-2025 — NCUA state-level metrics (delinquency_bps, loan_to_share_pct, roaa_bps, positive_net_income_pct)
• GET /api/credit-union-segments/{id} — segment details, peer states, minimum checklist, internal context, quarterly capacity

Risk Rating Re-derivation (Dominant Factor Rule)

Final rating = worst (highest numeric) of available DSCR, LTV, and delinquency factors:

DSCR thresholds:

LTV thresholds:

Delinquency minimums:

Material downgrade = ≥ 2 notches (per policy material_downgrade_notches).

When a loan has no DSCR, LTV, or delinquency-based factors available, retain its current rating.

CDFI Factor Scoring

Factor score = FICO score + LTV score + Debt-to-Asset score + Liquidity score. Null/missing factors score 0.

Risk classes: Prime (0-5), Desirable (6-9), Satisfactory (10-13), Watch (14-18), Doubtful (≥19), Projected Loss (≥19 and LTV > 1.0).

For Nonaccrual loans with underwater collateral (LTV > 1.0), "Projected Loss" classification may apply even when the numeric score is below 19, reflecting the credit's actual loss exposure.

Stress-Test Formulas

Watch-list parallel shock (+200bp):

stressed_dscr = dscr / (1 + 0.18)
breach_threshold = 1.0

CRE dual-stress formula:

stressed_dscr = dscr × 0.85 / (1 + 0.18)
breach_threshold = 1.0

CRE Weighted Scoring (Competing Credit)

Score each application 1-5 (1=best) on five dimensions, then compute weighted sum. Lower total is better.

Score classes: approve_quality (≤ 2.0), conditional (≤ 3.0), weak (> 3.0).

Concentration Rules

Sector concentration limit = limit_pct × total_loans_outstanding. Per-sector overrides live in sector-exposures (e.g., Healthcare may have a tighter limit_pct than the branch default sector_ceiling_pct).

Existing over-ceiling exposure is grandfathered, but new approvals may not worsen that sector without mitigation: participation_required, reduced_amount, or board_exception.

CRE exposure = sum of outstanding balances for loans with loan_type = "CRE". CRE concentration = CRE exposure / total_loans_outstanding. Compare against cre_policy_limit_pct.

For SBA-guaranteed loans, only the unguaranteed portion (approved_amount × (1 - sba_guaranty_pct)) counts against bank_capacity_used, but the full approved amount counts toward sector/CRE concentration.

Benchmark Conventions

NPA benchmark: Use total_loans_noncurrent_pct from FDIC Q4 2024. Branch NPA = nonperforming_loans from branch metrics. Ratios computed against total_loans_outstanding.

FDIC delinquency benchmark: Use total_real_estate_30_89_pct for CRE-related reviews. Branch ratio = branch-level delinquency metric (from branch metrics or computed from loan-level data).

NCUA benchmarks: Use exact integer values from the NCUA Q1 2025 table. Peer median = median of listed peer states for each metric. Direction comparisons: "higher", "lower", or "equal" relative to US national or peer median.

Output Field Conventions

• Exposure/balance fields: rounded to 2 decimal places (USD)
• Ratios (percentages as decimals): rounded to 4 decimal places
• Basis points (bps): rounded to 2 decimal places; bps = ratio × 10000
• Weighted CDFI CRE score: rounded to 1 decimal place
• NCUA metric integers: exact values as reported (no rounding)
• factor_score: integer
• Sorting: loan_ids ascending within lists; final_rating ascending in exposure totals; sector then application_id for concentration flags; descending exposure then ascending loan_id for workout queues
• Lists with no items: use empty array [], not omitted
• Enum values: use exact string values from templates; do not invent new ones

Common Pitfalls

1. Sector limit computation: Always use limit_pct × total_loans_outstanding (not total_assets or lending_capacity). Verify by summing sector_exposures — they should equal total_loans_outstanding.

2. Re-derivation population: Only loans meeting the rating threshold (e.g., current_rating ≥ 3) are re-derived. Loans below threshold keep their current rating and are excluded from target counts but included in portfolio-wide exposure totals.

3. Final rating exposure totals: Include only the re-derived (target) population, not all branch loans. Each loan's final re-derived rating is its assigned bucket.

4. Watch-list action coverage: All target-population loans receive an action assignment. Action tiers: monitor (ratings 3-4), watchlist (5-6), special_assets (7, current), workout (7, past due), partial_chargeoff_review or legal_referral (8).

5. Delinquency factor is a minimum, not additive: If a loan is 30 DPD, the rating cannot be better than 4 regardless of DSCR/LTV.

6. SBA capacity treatment: bank_capacity_used = approved_amount × (1 - sba_guaranty_pct). The SBA-guaranteed portion reduces the bank's capital commitment. But full amount counts toward gross and sector exposure.

7. Post-approval concentrations: Only include sectors with approved applications or existing exposure changes. Unaffected sectors may be omitted from the summary.

8. Monitoring cadence for mixed adverse portfolios: Use quarterly as default when the population spans multiple risk classes; use monthly only when most credits are Watch or worse.

9. DSCR-stressed rounding: Compute stressed DSCR from raw base DSCR first, then round the result to 2 decimals.

10. NCUA peer median: Sort peer state values, take middle. For 3 peers, it's the 2nd value. Compare NC against this median for each of the 4 metrics independently.