Self Attempt 02

Bank Branch Credit-Risk Lending-Committee Skill

API Workflow

Base URL: http://34.46.77.124:8011 (override any localhost/setup.sh references).

Discovery: Start with GET /api/manifest (lists endpoints, benchmark versions, counts). Then GET /api/health.

Key endpoints (see manifest for all):

Risk Rating Derivation (Dominant Factor Rule)

Rule: Final rating = worst (highest numeric) rating from DSCR, LTV/collateral, and delinquency factors. If a factor is missing (null), skip it.

DSCR → Rating

LTV → Rating (use collateral_value / outstanding_balance if ltv field is null)

Delinquency → Rating Floor

Material downgrade: ≥ 2 notches worsening from current_rating to final_rating.

NPA

NPA = loans where payment_status is "Nonaccrual" OR "90+ Days Past Due". NPA exposure = sum of outstanding_balance for NPA loans.

CDFI Factor Scoring (for watch-list / adverse-rated loans)

Score each factor by ranges. Null factor → skip (contribute 0).

Risk Class from total score:

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

CRE Weighted Score (for CRE applications)

Weights: capacity 0.45, capital 0.03, character 0.05, collateral_exposure 0.36, conditions 0.11.

Each dimension scored 1–5 (lower is better). Compute weighted sum:

weighted_score = Σ (dimension_score × weight)

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

Stress Tests

Watch-List Stress (+200bp parallel shock)

stressed_dscr = dscr / (1 + 0.18)

Apply only to loans with DSCR available. Breach if stressed_dscr < 1.0.

CRE Dual Stress

stressed_dscr = dscr × 0.85 / (1 + 0.18)

Used for CRE competing-decision analysis. Coverage breach threshold = 1.0.

Concentration Rules

• Lending capacity: branches.lending_capacity_q1 (hard cap for gross approvals).
• Sector ceiling: branches.sector_ceiling_pct is the default. The sector_exposures table provides per-sector limit_pct overrides.
• CRE policy limit: branches.cre_policy_limit_pct — total CRE exposure / total assets must not exceed this.
• Grandfathering: Existing over-ceiling exposure may be grandfathered, but new approvals may not worsen that sector's overage without mitigation.
• Allowed mitigations: participation_required, reduced_amount, board_exception.
• Sector exposure pct = sector current_exposure / branch total_assets. Post-approval adds the approved amount.
• CRE concentration = sum of all CRE-type loan outstanding_balances / total_assets. Post-approval adds approved CRE amounts.

Application Decisioning

Decision Enums

approve | conditional_approve | decline | defer | participation_required

Conditions

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

Watch-List / Workout Actions

monitor | watchlist | special_assets | workout | partial_chargeoff_review | legal_referral

Credit Union Segment Posture

Posture Enums

continue_approving | continue_with_tighter_conditions | temporarily_pause

NCUA Benchmark Metrics

Compare state_code row vs US row and vs peer_states median for: delinquency_bps, loan_to_share_pct, roaa_bps, positive_net_income_pct.

Controls

Required checklist gates: board_authorization | equipment_invoice | fleet_replacement_plan | payer_contract_summary | public_contract_or_tax_support | proof_of_insurance | ucc_or_title_lien

Added operating controls: pre_close_insurance_binder_verification | lien_perfection_prior_to_funding | senior_underwriter_second_review | quarterly_state_benchmark_monitoring | monthly_segment_delinquency_watch | committee_exception_for_capacity_overrun

Escalation Triggers

Condition choices: segment_recent_delinquency_ge_90_bps | missing_insurance_or_lien_exception | quarterly_capacity_exceeded_or_exception_requested | state_delinquency_gap_widens_25_bps

Owner choices: credit_risk_manager | operations_control_manager | lending_committee_chair

Benchmark Comparison

FDIC Benchmarks

fdic_q4_2024 fields:

• total_loans_noncurrent_pct (0.0098)
• total_real_estate_noncurrent_pct (0.0121)
• construction_development_noncurrent_pct (0.0076)
• total_real_estate_30_89_pct (0.0051)
• construction_development_30_89_pct (0.0042)

Computation

branch_npa_ratio = branch_npa_exposure / branch_total_loans
variance_ratio = branch_npa_ratio - fdic_benchmark_ratio
variance_bps = variance_ratio × 10000

Branch delinquency ratio for 30-89 comparison: delinquency_30_plus_pct from metrics. Convert to ratio (divide pct by 100) if needed.

Output Field Conventions

Precision

• Currency / exposure (USD): 2 decimal places
• Ratios / percentages (as decimals): 4 decimal places
• BPS (basis points): 2 decimal places (except NCUA metrics which are integer BPS as-reported)
• Weighted CDFI score: 1 decimal place
• Integers: loan counts, ratings, notches, factor scores, NCUA BPS

Sorting

• Lists sorted ascending by the key field specified in the answer template (loan_id, application_id, sector, final_rating, current_rating, action, trigger_id, etc.)
• For multi-key sorts: primary key first as specified (e.g., descending exposure, then ascending loan_id for workout_queue)
• reason_codes arrays: ascending alphabetically
• loan_ids arrays: ascending loan_id
• List items sorted as specified in template; when template says "ascending by X", sort by X ascending.

Null/Missing Handling

• When computing derived ratings, skip factors where the underlying field is null
• When DSCR is null, skip DSCR-based rating and stress computation
• When LTV field is null but collateral_value and outstanding_balance exist, compute LTV = outstanding_balance / collateral_value
• When both LTV field and collateral_value are null, skip LTV-based rating
• For CDFI scoring, null factors contribute 0 to the total score
• fico may be null for business loans; skip FICO scoring when null

Common Pitfalls

1. Rating re-derivation scope: Only re-derive ratings for loans meeting the stated criteria (e.g., "currently rated 3 or worse" or "rated 6 or worse"). Do NOT re-derive for all loans unless instructed.

2. Dominant factor = worst: The final rating is the maximum (worst) of DSCR, LTV, and delinquency ratings — not an average or weighted combination.

3. Delinquency is a floor: The delinquency rating sets a minimum. If DSCR and LTV suggest 3 but the loan is 60 DPD, the final rating is at least 5.

4. Nonaccrual always maps to 8: Overrides all other factors per the delinquency minimum table.

5. CRE vs C&I vs other loan types: CRE concentration uses only loan_type == "CRE" loans. Sector concentration uses all loans regardless of type.

6. Sector ceiling vs CRE limit: These are separate constraints. The sector ceiling applies per-sector across all loan types. The CRE policy limit applies only to CRE-type loans relative to total assets.

7. Grandfathered exposure: Sector entries with grandfathered: 1 are already over-ceiling. New applications in that sector cannot be approved without mitigation, even if the overall capacity exists.

8. Capacity computation: remaining_capacity = lending_capacity_q1 - committed_capacity_amount where committed includes all approved/conditional amounts.

9. Stress DSCR for watch-list: Uses the formula dscr / (1 + 0.18), NOT the CRE dual-stress formula. The CRE dual-stress (dscr * 0.85 / (1 + 0.18)) is only for CRE competing decisions.

10. Benchmark metric selection: Match the benchmark metric to the context. Use total_loans_noncurrent_pct for general NPA comparisons. Use total_real_estate_noncurrent_pct or construction_development_noncurrent_pct when the analysis focuses on real estate / construction segments.

11. Percentage format: When computing concentration ratios, divide by total_assets or total_loans (whichever is appropriate) and express as a decimal ratio (e.g., 0.1523 not 15.23%).

12. Peer median for NCUA: Compute the median across peer_states values for each metric. For even numbers of peers, take the mean of the two middle values (standard median).

13. Application priority ranking: Only approved and conditionally approved applications appear in the priority_ranking list, ordered highest priority first.

14. Documentation gap: Applications with documentation_complete: 0 get a documentation_gap decline/defer reason code.