Self Attempt 03

Bank Branch Credit-Risk Lending-Committee Skill

Environment & API Entry

Base URL: http://34.46.77.124:8011 (from environment_access.md; overrides any localhost refs in task prompts).

Start every session with GET /api/manifest to confirm benchmark versions and available endpoints.

Core Endpoints

Risk Rating Re-Derivation (Dominant Factor Rule)

Rule: Final re-derived rating = worst (highest numeric) rating from all available factors among DSCR, LTV, and delinquency.

DSCR → Rating

LTV → Rating

Delinquency Floor (minimum rating from payment status)

Null/missing factors: If DSCR or LTV is null, skip that factor entirely (it contributes nothing). Delinquency floor always applies regardless of other factors.

Material downgrade: A re-derived final_rating that is ≥ 2 notches worse than current_rating (final_rating − current_rating ≥ 2).

CDFI Risk-Class Scoring

Assign each loan a risk class by summing independent factor scores. Score each factor where data is available; skip nulls.

Risk class from total score:

• 0–5: Prime
• 6–9: Desirable
• 10–13: Satisfactory
• 14–18: Watch
• ≥ 19: Doubtful
• ≥ 19 AND LTV > 1.0: Projected Loss

CRE Weighted Credit Scoring

Used for competing CRE application decisions. Weights: capacity 0.45, capital 0.03, character 0.05, collateral_exposure 0.36, conditions 0.11.

Score classes (lower is better):

• ≤ 2.0: approve_quality
• ≤ 3.0: conditional

• 3.0: weak

Stress-Test Formulas

Watch-list +200bp DSCR Stress

stressed_dscr = base_dscr / 1.18

Applied to watch-list loans where DSCR is available. Breach threshold = 1.00.

CRE Dual Stress

stressed_dscr = dscr × 0.85 / 1.18

Coverage breach threshold = 1.00.

Concentration Rules

Sector Concentration

• Each branch has a default sector_ceiling_pct on the branch object.
• The /sector-exposures endpoint provides per-sector limit_pct which may differ from the default.
• grandfathered: 1 = existing over-limit exposure is tolerated but new approvals must not worsen that sector without mitigation.
• Allowed mitigations: participation_required, reduced_amount, board_exception.
• Concentration ratio = sector_exposure / total_loans_outstanding (use most recent quarter, Q1).

CRE Concentration

• Branch cre_policy_limit_pct caps total CRE exposure as a fraction of total loans.
• Post-approval CRE = (existing CRE exposure + selected app amount) / total_loans_outstanding.
• FDIC benchmark: total_real_estate_30_89_pct for delinquency comparison.

Capacity

• lending_capacity_q1 from branch is the Q1 lending budget.
• Bank capacity used = requested_amount (or approved amount).
• Committed capacity = sum of approved amounts.

NPA Benchmark Calculation

Metric: total_loans_noncurrent_pct from FDIC Q4 2024 (0.0098).

branch_npa_exposure = nonperforming_loans (from 2025Q1 metrics)
branch_total_loans  = total_loans_outstanding (from 2025Q1 metrics)
branch_npa_ratio    = branch_npa_exposure / branch_total_loans
variance_ratio      = branch_npa_ratio - fdic_benchmark_ratio
variance_bps        = variance_ratio × 10000

Credit Union Segment Posture

From segment endpoint: risk_tolerance, recent_delinquency_bps, state_code, peer_states, quarterly_capacity, current_outstanding, minimum_checklist, internal control_issue, staffing_constraint, notes.

State Metrics

Read from NCUA benchmark row matching the segment's state_code. Values are integers exactly as reported.

Peer Comparison

Compare NC state metrics to US national row (state_code: "US") and peer-state median. Direction: higher, lower, or equal for each of delinquency_bps, loan_to_share_pct, roaa_bps, positive_net_income_pct.

Posture Decision

• continue_approving — capacity available, external risk stronger/neutral, risk_tolerance expansive or moderate with strong metrics.
• continue_with_tighter_conditions — capacity available but external risk mixed or requires added controls.
• temporarily_pause — severe external risk weakness or capacity exhausted.

Controls & Triggers

• required_checklist_gates: from segment minimum_checklist plus any additional required gates.
• added_operating_controls: derived from internal context (control issues, staffing constraints) and posture.
• escalation_triggers: conditions like segment_recent_delinquency_ge_90_bps, missing_insurance_or_lien_exception, quarterly_capacity_exceeded_or_exception_requested, state_delinquency_gap_widens_25_bps. Owners: credit_risk_manager, operations_control_manager, lending_committee_chair.

Allocation & Decision Logic

Priority Ranking

Rank applications by credit quality signals: higher DSCR, higher FICO, lower LTV, longer relationship, lower risk sector concentration impact. Applications with documentation_complete: 0 or red flags (recent bankruptcy, very low FICO, startup with no mitigants) rank lower.

Decision Codes

approve, conditional_approve, decline, defer, participation_required

Condition Codes

participation_required, reduced_amount, board_exception, sba_guaranty_required, startup_monitoring, none

Decline Reason Codes

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

Recommended Action Codes

monitor, watchlist, special_assets, workout, partial_chargeoff_review, legal_referral

Output Conventions (MUST FOLLOW)

Numeric Precision

Sorting

• Lists of loans/ratings: ascending by the key field (final_rating, loan_id, application_id, sector, trigger_id, payment_status).
• Exception: workout_queue sorts descending exposure, then ascending loan_id.
• material_downgrades: ascending loan_id.
• decline_reasons: reason codes sorted alphabetically within each application.
• conditions (Harbor): ascending alphabetically.

Enum Values — Exact Strings Required

Payment status: "Current", "30 Days Past Due", "60 Days Past Due", "90+ Days Past Due", "Nonaccrual" Risk class: "Prime", "Desirable", "Satisfactory", "Watch", "Doubtful", "Projected Loss" Score class: "approve_quality", "conditional", "weak" Monitoring cadence: "monthly", "quarterly", "semiannual" Posture: "continue_approving", "continue_with_tighter_conditions", "temporarily_pause"

Common Pitfalls

1. Nonaccrual dominates: A loan with payment_status "Nonaccrual" gets rating 8 regardless of DSCR/LTV — the delinquency floor is a hard minimum, and 8 is the worst possible.
2. Missing DSCR/LTV: When dscr or ltv is null, skip that factor in dominant-factor rating. Do NOT assign a rating from it.
3. Material downgrade threshold is 2, not 1: Only report downgrades where final_rating − current_rating ≥ 2.
4. FDIC benchmark is one object, not an array: GET /api/benchmarks/fdic/q4-2024 returns {} with named fields, not [].
5. Sorting migration_from_current_rating_3: Sort by final_rating ascending, not by loan_id.
6. variance_bps = variance_ratio × 10000: 1% = 100 bps, so ratio 0.01 = 100 bps. Multiply ratio by 10000 to get bps.
7. NCUA US row: The national row has state_code: "US" — use it for the national comparison.
8. Peer median, not mean: For NCUA peer comparison, compute the median of peer-state values, not the mean.
9. Watch-list stress only for loans with DSCR: Loans without DSCR are excluded from stress_results; only include loans where dscr is not null.
10. Grandfathering is per-sector: Check grandfathered field in sector-exposures. Sectors with grandfathered=1 may show exposure above limit_pct.
11. Top problem credit: Select the loan with the worst (highest) final_rating. If tied, pick the one with the highest exposure. Include recommended_action from the action-code enum.
12. watch_list_action_coverage: Count loans from the regrade population that received a watch-list action. Group by action type with loan_ids sorted ascending.
13. CRE dual-stress formula is dscr * 0.85 / 1.18: Apply the 0.85 revenue haircut AND the 1.18 rate shock multiplicatively, not additively.
14. approved_amount vs requested_amount: For conditional approvals, approved_amount may be less than requested_amount (e.g., reduced_amount condition).
15. Post-approval concentration: Recompute sector exposure after adding approved amounts to the relevant sectors.