Self Attempt 03

Credit Risk Committee JSON Skill

Overview

Produce committee-ready JSON answers from a shared credit-office public API. Tasks include:

• Branch rating-migration review (train_001, REDWOOD)
• Lending-committee allocation package (train_002, LAKEVIEW)
• Credit-union segment posture page (train_003, CIVIC_NC_FIRE_EMS)
• Watch-list stress & workout review (train_004, SUMMIT)
• Competing CRE decision (train_005, HARBOR)

Environment

• Base URL is fixed in environment_access.md (e.g. http://34.46.77.124:8011).
• Do not run env/setup.sh or use localhost.

API Workflow (transferable across tasks)

1. Bootstrap

   • GET /api/manifest → list of valid endpoints and record counts.
   • GET /api/policies → credit policy, CDFI scoring tables, stress formulas, risk-rating rules.
   • GET /api/branches → all branches with branch_id, cre_policy_limit_pct, sector_ceiling_pct, lending_capacity_q1, total_assets, fdic_benchmark_set, institution_type.

2. Branch-level data

   • GET /api/branches/{branch_id} → branch details.
   • GET /api/branches/{branch_id}/metrics → quarterly metrics (delinquency_30_plus_pct, total_loans_outstanding, nonperforming_loans, etc.).
   • GET /api/branches/{branch_id}/loans → full loan book with current_rating, payment_status, days_past_due, dscr, ltv, debt_to_asset, liquidity_months, outstanding_balance, sector.
   • GET /api/branches/{branch_id}/sector-exposures → sector limits, current exposure, grandfathered flags.
   • GET /api/branches/{branch_id}/applications → pending applications with loan_type, sector, requested_amount, dscr, ltv, documentation_complete, etc.

3. Benchmarks

   • GET /api/benchmarks/fdic/q4-2024 → FDIC ratios (total_real_estate_30_89_pct, total_loans_noncurrent_pct, etc.).
   • GET /api/benchmarks/ncua/q1-2025 → NCUA state-level metrics (delinquency_bps, loan_to_share_pct, roaa_bps, positive_net_income_pct).

4. Credit-union segments

   • GET /api/credit-union-segments/{segment_id} → segment profile, capacity, controls, internal context, peer states.

Core Business Rules

Rating Re-derivation (applies to migration & watch-list tasks)

• Use the dominant-factor rule: final rating = worst numeric rating from DSCR, LTV/collateral, and delinquency factors.
• DSCR thresholds: ≥1.5→3, ≥1.25→4, ≥1.05→5, ≥1.0→6, <1.0→7.
• LTV thresholds: ≤0.65→3, ≤0.75→4, ≤0.85→5, ≤1.0→6, >1.0→7.
• Delinquency minimums: Current→null, 30 DPD→4, 60 DPD→5, 90+ DPD→7, Nonaccrual→8.
• Material downgrade = ≥2 notches. Track only loans whose final rating ≥ target threshold.

CDFI Factor Scoring (applies to watch-list & CRE tasks)

• Sum scores for: debt_to_asset, fico, liquidity_months, ltv (each from policy tables).
• Classes: Prime 0-5, Desirable 6-9, Satisfactory 10-13, Watch 14-18, Doubtful ≥19, Projected Loss ≥19 and ltv>1.0.

CRE Weighted Score (train_005)

• Weights: capacity 0.45, capital 0.03, character 0.05, collateral_exposure 0.36, conditions 0.11.
• Classes: approve_quality ≤2.0, conditional ≤3.0, weak >3.0.

Stress Formulas (from policies)

• CRE dual-stress: stressed_dscr = dscr * 0.85 / (1 + 0.18)
• Watch-list stress: stressed_dscr = dscr / (1 + 0.18) (the +200bp parallel shock)
• Breach threshold = 1.0 for both.

Concentration & Capacity (train_002)

• existing_cre_exposure = sum of outstanding balances where loan_type == "CRE".
• existing_cre_concentration = existing_cre_exposure / branch.total_assets.
• Post-approval concentration = (existing_cre_exposure + approved_amount) / total_assets.
• Policy variance bps = (post_approval_concentration - cre_policy_limit_pct) * 10000.
• Sector post-approval pct = (current_exposure + approved_amount) / total_assets.
• If post-approval pct > limit_pct, flag and apply mitigation (participation_required, reduced_amount, board_exception).
• Grandfathered exposure may not be increased without mitigation.

NPA / Delinquency Benchmarking

• branch_npa_ratio = nonperforming_loans / total_loans_outstanding (from latest metrics quarter).
• fdic_benchmark_ratio = choose the matching metric from FDIC data (e.g. total_loans_noncurrent_pct for NPA, total_real_estate_30_89_pct for delinquency).
• variance_ratio = branch_ratio - benchmark_ratio.
• variance_bps = variance_ratio * 10000, rounded to 2 decimals.

Output Conventions

• Always write pure JSON matching the task's answer_template.json. No markdown, no commentary outside the JSON.
• Sort orders (strictly obey template instructions):
  • Lists keyed by application_id or loan_id: ascending alphanumeric.
  • conditions, reason_codes: ascending alphabetical.
  • workout_queue: descending exposure, then ascending loan_id.
  • severe_bucket_counts: ascending current_rating, then payment_status.
  • concentration_flags: ascending sector, then application_id.
  • post_approval_concentrations: ascending sector.
• Precision:
  • Currency (USD): 2 decimals.
  • Percentages as ratios: 4 decimals (e.g. 0.2853).
  • Bps / variance: 2 decimals.
  • DSCR / stressed DSCR: 2 decimals.
  • Weighted CDFI score: 1 decimal.

Common Pitfalls

1. Using current_rating instead of re-derived final_rating for migration and watch-list tasks. Always re-derive per the dominant-factor rule.
2. Ignoring null DSCR/LTV — loans without DSCR cannot be stressed; omit from stress results or handle gracefully (do not crash).
3. Wrong benchmark metric — match the metric name exactly to the task context (NPA → total_loans_noncurrent_pct, CRE delinquency → total_real_estate_30_89_pct).
4. Forgetting grandfathered flags in sector exposures — grandfathered sectors have a higher limit; new approvals in those sectors still need mitigation if they would increase exposure.
5. Miscomputing CRE concentration — use total_assets from branch details, not lending_capacity_q1.
6. Wrong stress formula — watch-list uses dscr / 1.18; CRE dual-stress uses dscr * 0.85 / 1.18.
7. Including non-CRE loans in CRE exposure — filter by loan_type == "CRE".
8. Missing documentation_complete == 0 as a decline/conditional trigger (reason code documentation_gap).
9. Credit-union tasks: use NCUA benchmarks, not FDIC. Match segment's state_code to the NCUA row. Peer comparison uses the segment's peer_states list.
10. Posture page: required_checklist_gates is the intersection of the segment's minimum_checklist and the allowed enum; added_operating_controls are chosen based on identified risks (e.g. delinquency watch, lien perfection, quarterly monitoring).

Quick Reference: Decision Enum

• approve, conditional_approve, decline, defer, participation_required

Quick Reference: Action Enum (workout / watch-list)

• monitor, watchlist, special_assets, workout, partial_chargeoff_review, legal_referral

Quick Reference: Reason Codes (decline / conditional)

• capacity_limit, sector_breach, weak_dscr, high_ltv, low_fico, recent_bankruptcy, startup_risk, underwater_collateral, policy_floor_missing, documentation_gap, fdic_adverse_variance, ncua_peer_weakness