# MODEL GOVERNANCE DOCUMENTATION
## Sentinel Predictive Intelligence Engine – Credit Rating Model
### SR 11-7 Compliance Framework

**Document Version:** 2.0  
**Effective Date:** 2026-Q2  
**Last Updated:** 2026-04-12  
**Status:** Phase 3 Remediation Complete  
**Prepared By:** Sentinel Credit Analytics Team  
**Approved By:** Chief Risk Officer (pending final sign-off)

---

## EXECUTIVE SUMMARY

The Sentinel Predictive Intelligence Engine (Probability of Insolvency) is a quantitative credit rating model designed to estimate default probability and assign credit ratings to large-cap, publicly-traded corporations. The model replicates the published methodologies of S&P Global Ratings, Moody's Investors Service, and Fitch Ratings, synthesizing inputs from financial statement analysis, market data, and sector-specific adjustments to produce a final rating on a 1-21 notch scale (1=AAA, 21=D).

This document describes the model's architecture, validation approach, known limitations, and governance framework in compliance with the Federal Reserve's Supervision & Regulation (SR) Letter 11-7, which establishes expectations for credit risk models used in supervisory capital calculations and risk management decisions.

**Key Model Characteristics:**
- **Scope:** Large-cap equities (>$1B market cap), investment-grade & high-yield corporates
- **Inputs:** Audited GAAP financial statements, market data (equity prices, CDS spreads), sector classification
- **Output:** Credit rating (1-21 notch scale) + confidence interval + sensitivity analysis
- **Methodology:** Multi-dimensional (business risk, financial risk, liquidity, leverage) with sector-specific adjustments
- **Validation:** 32-company backtesting (2010-2025 vintages), out-of-sample holdout (15 large-cap names), sector-stratified testing
- **Limitations:** Relies on historical data patterns; cannot model unprecedented shocks (pandemics, geopolitical disruptions); infrastructure/solar FFO treatment requires judgment

---

## SECTION 1: MODEL DESCRIPTION & PURPOSE

### 1.1 Model Overview

The Sentinel Predictive Intelligence Engine synthesizes three complementary rating methodologies:

1. **S&P Global Ratings Anchor Matrix Methodology**
   - Business Risk Profile (BRP): 1-6 scale based on industry risk, country risk, competitive position
   - Financial Risk Profile (FRP): 1-6 scale based on leverage, coverage, cash flow generation
   - Anchor Rating: BRP × FRP matrix produces 6×6 grid of baseline ratings (1-19 notches)
   - Adjustments: liquidity, financial flexibility, leverage trends

2. **Moody's Analytical Approach**
   - Industry risk assessment (cyclicality, leverage norms, default history)
   - Company-specific metrics (scale, geographic diversity, brand value, R&D intensity)
   - Financial metrics: Debt/EBITDA, FCF/Debt, Interest Coverage, ROIC
   - Comparable company positioning within sector peer sets

3. **Fitch Ratings Quantitative Framework**
   - Leverage ratio assessment (Debt/EBITDA, Debt/Assets)
   - Profitability & cash flow (EBITDA margin, FCF conversion, FCF/Debt)
   - Liquidity position (cash on hand, undrawn facilities, maturity ladder)
   - Capital structure (equity quality, subordination, covenant protections)

The engine **does not replace** published agency ratings; rather, it provides an independent benchmark for internal credit assessment, portfolio risk monitoring, and peer comparison. All ratings generated by the model are labeled "HELIOS ENGINE ESTIMATES" to distinguish them from official agency ratings.

### 1.2 Intended Use Cases

1. **Portfolio Risk Monitoring:** Track rating drift for holdings; identify credit quality deterioration before agency downgrades
2. **Relative Value Analysis:** Compare Sentinel-estimated ratings to published agency ratings to identify arbitrage opportunities
3. **Deal Assessment:** Rate new issuers or private credit opportunities using published financial statements
4. **Scenario Analysis:** Model impact of leveraged transactions (M&A, LBOs, debt issuance) on credit ratings
5. **Covenant Monitoring:** Predict rating/covenant covenant breach probability using financial projections
6. **Regulatory Compliance:** Meet supervisory expectations for quantitative credit models (SR 11-7)

**Non-Intended Uses:**
- Official credit underwriting (must incorporate qualitative judgment, management quality, industry dynamics)
- Sole basis for loan approval (must supplement with analyst judgment, bank policy, relationship context)
- Real-time trading signals (model updates quarterly/annually, not designed for intra-quarter signals)

### 1.3 Model Scope & Constraints

**In Scope:**
- Public companies with audited GAAP/IFRS financials
- Debt securities rated by major agencies (S&P, Moody's, Fitch)
- Investment-grade and high-yield corporates
- North American & Western European issuers (97% of calibration sample)
- Fiscal years 2010-2025 (with pre-2010 data for defaults calibration)
- Large-cap threshold: >$1B market cap

**Out of Scope:**
- Private companies (market cap data unavailable)
- Sovereign debt (separate rating model)
- Sub-investment-grade below CCC (model less reliable; default imminent for most CCC- and D)
- Distressed/Chapter 11 companies (model assumes going-concern valuation)
- Emerging markets <$100M market cap (market data unreliable)
- Insurance companies, banks (specialized financial risk models required)

### 1.4 Rating Scale

The Sentinel engine produces ratings on a 1-21 notch scale, aligned with S&P Global Ratings:

| Notch | Label | Category | Interpretation |
|-------|-------|----------|-----------------|
| 1 | AAA | Investment Grade | Highest quality, minimal default risk |
| 2 | AAA- | Investment Grade | Extremely strong financial position |
| 3 | AA+ | Investment Grade | Very strong; very low default risk |
| 4 | AA | Investment Grade | Stable, strong fundamentals |
| 5 | AA- | Investment Grade | Strong with some vulnerability |
| 6 | A+ | Investment Grade | Upper-medium grade |
| 7 | A | Investment Grade | Solid fundamentals |
| 8 | A- | Investment Grade | Lower-medium grade |
| 9 | BBB+ | Investment Grade | Upper speculative |
| 10 | BBB | Investment Grade | Adequate financial position |
| 11 | BBB- | Investment Grade | Lower investment grade |
| 12 | BB+ | High-Yield | Non-investment grade; significant risk |
| 13 | BB | High-Yield | Material default risk |
| 14 | BB- | High-Yield | Substantial default risk |
| 15 | B+ | High-Yield | Speculative; highly leveraged |
| 16 | B | High-Yield | Very speculative |
| 17 | B- | High-Yield | Distressed/vulnerable |
| 18 | CCC+ | High-Yield | Near-term default risk (distressed) |
| 19 | CCC | High-Yield | Imminent default risk |
| 20 | CCC- | High-Yield | Default expected within months |
| 21 | D | Default | In default |

**Investment-Grade Threshold:** Ratings 1-11 (AAA through BBB-)  
**High-Yield Threshold:** Ratings 12-21 (BB+ through D)

---

## SECTION 2: MODEL DEVELOPMENT

### 2.1 Methodology Overview

The Sentinel Predictive Intelligence Engine synthesizes three methodological pillars:

#### **Pillar 1: Business Risk Assessment (20% of final rating)**

Business risk captures the durability of competitive position, industry dynamics, and management quality using a 1-6 scale:

**Score 1 (Excellent):** Dominant market position, high barriers to entry, diversified customer base, counter-cyclical demand  
*Examples:* Microsoft, Amazon, Coca-Cola

**Score 2 (Strong):** Strong market position, meaningful differentiation, stable industry structure  
*Examples:* Apple, Johnson & Johnson, Procter & Gamble

**Score 3 (Satisfactory):** Adequate market position, normal cyclicality, moderate competitive pressure  
*Examples:* Ford, General Motors, Southwest Airlines

**Score 4 (Fair):** Weak market position, high cyclicality, commodity-like competition  
*Examples:* Regional banks, discount retailers, commodity producers

**Score 5 (Weak):** Distressed competitive position, declining industry, single-product dependency  
*Examples:* Struggling mall retailers, legacy telecom players with churn

**Score 6 (Vulnerable):** Untenable competitive position, likely to exit market or consolidate  
*Examples:* Companies in terminal decline (Blockbuster era)

**Calculation:** Business Risk = weighted average of:
- Industry Risk (40%): derived from Moody's sector-level default rates
- Country Risk (30%): for non-US exposure (assumes 0% for pure domestic US issuers)
- Competitive Position (30%): based on market share, brand strength, innovation capacity

```
Industry Risk: 1 = <1% annual default rate; 6 = >10% annual default rate (per Moody's)
Country Risk: 1 = AAA sovereign; 6 = CCC sovereign (proxies to corporate risk)
Competitive Position: 1 = >#50% mkt share or monopoly; 6 = <5% market share
```

#### **Pillar 2: Financial Risk Assessment (60% of final rating)**

Financial risk is the dominant factor, reflecting leverage, profitability, cash flow generation, and coverage ratios. Decomposed into a 1-6 Financial Risk Profile (FRP):

**Score 1 (Minimal):** Net leverage <1.5x EBITDA, EBITDA interest coverage >8x, FCF positive & growing  
**Score 2 (Modest):** Net leverage 1.5-2.5x, coverage 5-8x, stable FCF  
**Score 3 (Intermediate):** Net leverage 2.5-4.0x, coverage 3-5x, adequate FCF generation  
**Score 4 (Significant):** Net leverage 4.0-5.5x, coverage 2-3x, FCF pressure emerging  
**Score 5 (Aggressive):** Net leverage >5.5x, coverage <2x, negative FCF trending  
**Score 6 (Highly Leveraged):** Net leverage >8x, coverage <1x, cash burn imminent  

**Key Financial Ratios (Composite 35/35/30 weighting):**

1. **Leverage Metrics (35% weight):**
   - Debt/EBITDA (industry-adjusted thresholds; infrastructure/utilities 0.5x higher)
   - Net Debt/EBITDA (excludes cash; preferred for financial analysis)
   - Debt/Assets (reflects solvency; less sensitive to cyclical EBITDA)
   - Adjusted Debt (adds back operating leases per ASC 842, adjusts for non-recourse)

2. **Coverage Metrics (35% weight):**
   - EBITDA Interest Coverage (EBITDA / Total Interest Expense)
   - EBIT Interest Coverage (EBIT / Total Interest Expense)
   - FFO Interest Coverage (FFO / Total Interest Expense)
   - FCF Coverage (FCF / Interest Expense)

3. **Profitability & Cash Generation (30% weight):**
   - EBITDA Margin (EBITDA / Revenue; sector-adjusted norms)
   - EBIT Margin (Operating profit margin)
   - FCF/Debt (Free Cash Flow as % of total debt; measures deleveraging capacity)
   - Return on Invested Capital (ROIC; EBIT × (1 - Tax Rate) / IC)
   - FFO/Debt (Funds From Operations as % of debt; alternative cash measure)

**Calculation:** Final Financial Risk Score = weighted composite of above ratios, with score 1-6 assigned based on percentile ranking within peer cohort.

#### **Pillar 3: Liquidity Assessment (15% weight, minimum score cap)**

Liquidity assesses ability to service debt over a 12-month horizon and refund maturing debt:

**Score 1 (Exceptional):** >$2B available liquidity, 5+ year average maturity, investment-grade bond access  
**Score 2 (Strong):** $500M-$2B liquidity, 4+ year maturity, regular market access  
**Score 3 (Adequate):** $200M-$500M liquidity, 3-4 year maturity, access conditional  
**Score 4 (Less than Adequate):** $50M-$200M liquidity, <3 year maturity, limited refinancing  
**Score 5 (Weak):** <$50M liquidity, covenant pressure, refinancing risk  

**Components:**
- Cash on hand (unrestricted)
- Undrawn revolver commitments (discounted for covenant capacity)
- Expected operating cash flow (next 12 months)
- Maturing debt (next 12 months)
- Refinancing risk adjustment

**Calculation:** Liquidity Score = function of (Available Liquidity / (Debt Due in 12M + Interest Accrual))

**Application:** Liquidity score is a "floor" adjustment; may cap financial risk score from going higher (better) if liquidity deteriorates below threshold.

### 2.2 Anchor Rating Matrix

The S&P Anchor Matrix combines Business Risk Profile (rows 1-6) and Financial Risk Profile (columns 1-6) to produce a baseline rating (1-19 notches):

```
         FRP1   FRP2   FRP3   FRP4   FRP5   FRP6
BRP1      1      2      4      7     10     13
BRP2      2      3      5      8     11     14
BRP3      4      5      7      9     12     15
BRP4      6      7      9     11     13     16
BRP5      8      9     11     13     15     17
BRP6     10     11     13     15     17     19
```

**Example:** Company with BRP=3 (Satisfactory) and FRP=3 (Intermediate) → Anchor Rating = 7 (A)

### 2.3 Adjustments to Anchor Rating

After deriving the anchor rating, apply the following adjustments (±0 to ±3 notches):

#### **A. Leverage Trend (+/- 1-2 notches)**
- **Rising leverage** (2nd derivative >0): Downgrade 1-2 notches (signal of distress acceleration)
- **Declining leverage** (2nd derivative <0): Upgrade 1 notch (deleveraging commitment credible)
- **Stable leverage** (±0.1x year-over-year): No adjustment

#### **B. Profitability Trend (+/- 1 notch)**
- **Margin expansion** (>200 bps YoY): Upgrade 1 notch
- **Margin contraction** (>200 bps YoY): Downgrade 1 notch
- **Stable margins**: No adjustment

#### **C. Financial Flexibility (+/- 0-1 notches)**
- **High flexibility** (large revolvers, investment-grade status): Upgrade 0-1 notch
- **Low flexibility** (covenant-constrained, sub-IG): Downgrade 0-1 notch

#### **D. Acquisition/Leverage Integration (-1-3 notches, temporary)**
- **Recent LBO**: Downgrade 2-3 notches temporarily (assess post-integration)
- **Large add-on acquisition**: Downgrade 1-2 notches until integration risk clears

#### **E. Covenant Quality & Structure (+/- 0-1 notches)**
- **Strong covenants** (low thresholds, springing features): Upgrade 0-1 notch
- **Weak covenants** (loose maintenance, incurrence-only): Downgrade 0-1 notch

#### **F. Sector-Specific Overlays**
- See Section 2.4 for infrastructure/REIT/utility/solar specific adjustments

### 2.4 Sector-Specific Methodology Adjustments

#### **Infrastructure & Toll Roads**
- **FFO Treatment:** Cap FFO at max(EBITDA, 0) to exclude financing gains
- **Leverage Threshold:** Debt/EBITDA up to 5-6x acceptable (vs. 3.5x general industrial)
- **Coverage:** EBITDA/Interest as low as 1.5x acceptable (vs. 2.5x minimum general)
- **Rationale:** Long-term, inflation-linked contracts provide stable cash; high leverage supported by non-recourse debt structure

#### **Solar/Renewable Energy**
- **FFO Treatment:** Conservative; use min(NI+D&A, EBITDA)
- **Leverage Threshold:** Debt/EBITDA 4-8x acceptable; contract-dependent
- **Coverage:** Fixed PPA rates provide stable interest coverage even if equity-thin
- **Key Metrics:** PPA contract tenor (15-20 year standard); counterparty risk of off-taker
- **Rationale:** Project finance structures; non-recourse debt; PPAs provide cash flow certainty

#### **Utilities (Electric, Water, Gas)**
- **FFO Treatment:** Standard (NI + D&A); regulatory environment suppresses non-operating items
- **Leverage Threshold:** Debt/EBITDA 3-4x acceptable (regulatory capital requirements, stable returns)
- **Coverage:** EBITDA/Interest >2.5x expected (regulatory guarantee)
- **Key Metrics:** Regulatory ROE expectations (8-10%), rate base growth, service territory demographics
- **Adjustment:** Slight upgrade (+0-1 notch) vs. industrial peer due to regulatory stability

#### **REITs (Real Estate)**
- **FFO Treatment:** Use FFO (funds from operations) directly; FFO/Debt ratio instead of traditional leverage ratios
- **FFO Definition:** GAAP earnings + D&A - gains/losses on asset sales (REIT industry standard)
- **Threshold:** FFO/Debt 8-15% range (equivalent to 6-12x debt/FFO) for investment-grade
- **Key Metrics:** Occupancy rate, lease spreads, same-store NOI growth, debt maturity ladder
- **Dividend Constraint:** Assess ability to maintain >90% payout (REIT requirement) while deleveraging
- **Rationale:** REITs are required to distribute 90% of taxable income; different capital structure dynamics

#### **Financial Institutions (Banks, Non-Bank Financials)**
- **Model Limitation:** Standard model NOT SUITABLE for banks; use separate financial institution model
- **Reason:** Balance sheet structure (loans vs. corporate debt), capital ratios (Tier 1, CET1), deposit stability
- **Excluded from Sentinel:** Insurance, investment banks (partially—equity-financed businesses)
- **Alternative:** Use regulatory stress tests (CCAR, DFAST for US banks); ECB SSM framework (EU)

#### **Technology & Software**
- **FFO Treatment:** Standard; check for non-operating income (stock-based comp, IP licensing)
- **Leverage Threshold:** Debt/EBITDA 2-3x (lower than industrials due to volatility)
- **Key Metrics:** Customer concentration, contract duration (SaaS: multi-year favorable; software licensing: annual)
- **R&D Sustainability:** Ensure R&D spending sustainable; R&D cuts signal distress
- **Market Cap Dependency:** High equity volatility; stress-test under 30-40% equity price decline

#### **Energy (Oil & Gas)**
- **Commodity Exposure:** Extreme cyclicality; require "trough case" scenario (oil $35-50/bbl)
- **FFO Treatment:** Conservative; use min(NI+D&A, CFO); exclude hedging gains
- **Leverage Threshold:** Highly variable by reserve life and hedging; range 2-5x Debt/EBITDA
- **Key Metrics:** Proved reserves, reserve replacement ratio, hedging ratio (% of next 2 years production)
- **Sensitivity:** Mandatory stress test under -$20/bbl shock to commodity prices
- **Covenant Triggers:** Many energy credits have commodity-linked covenants; monitor closely

#### **Consumer Cyclical (Retail, Automotive, Hospitality)**
- **Cyclicality Adjustment:** Use "trough" EBITDA estimate (cycle low) for leverage ratios
- **Coverage:** Require 2.5-3x minimum coverage even in peak year
- **FFO Treatment:** Exclude one-time store closures, restructuring, inventory write-downs
- **Trend Analysis:** Rising capex as % of sales (healthy reinvestment) vs. cutting capex (distress signal)
- **Key Metrics:** Comparable store sales, inventory days, payables period

### 2.5 Financial Data & Normalization

#### **Data Sources**
- **Primary:** SEC EDGAR filings (10-K annual reports, 10-Q quarterly)
- **Secondary:** Bloomberg Terminal, FactSet, Thomson Reuters Eikon
- **Market Data:** S&P Capital IQ (equity prices), Markit (CDS spreads), Fed funds rates
- **Validation:** Cross-check 3 sources for major balance sheet items; report discrepancies

#### **Accounting Adjustments**
The model converts GAAP data to "Rating Agency Adjusted" financials:

1. **Add-backs:**
   - Operating lease obligations (ASC 842 capitalization): ROU asset + obligation
   - Non-recourse debt interest (for infrastructure): Add back to interest expense
   - Stock-based comp (in operating margin calculation): Add back to operating expenses
   - Contingent consideration (M&A earnout): Treat as debt obligation

2. **Reclassifications:**
   - Preferred stock: Treat as hybrid debt (50% equity, 50% debt for leverage ratios)
   - Sale-leaseback proceeds: Adjust for off-balance-sheet financing impact
   - Operating JVs: Consolidate pro-rata for debt and EBITDA (not just equity carry)

3. **Normalizations:**
   - One-time items: Exclude from EBITDA if >5% of operating profit
   - Restructuring charges: Add back as "recurring" operational cost (amortize over 3 years)
   - Pension liabilities: Use funded status net of asset smoothing (if underfunded >10%)

#### **FFO Calculation**

Standard Definition:
```
FFO = Net Income + Depreciation & Amortization + Non-recourse Interest Add-back
    = NI + D&A + (Non-recourse Debt / Total Debt) × Interest Expense
```

Rationale for Non-Recourse Adjustment:
- Non-recourse debt is serviced from project cash flows, not consolidated cash
- Interest expense includes non-recourse interest (deducted from project NI)
- For rating consolidated credit, add back to reflect true recourse cash generation

**Example:**
- Net Income: $100M
- D&A: $50M
- Interest Expense: $30M (includes $10M non-recourse)
- Non-recourse Debt: $200M / Total Debt $400M = 50%
- FFO = $100 + $50 + ($10M) = **$160M** (not $150M)

### 2.6 Model Calibration & Testing

#### **Backtesting Approach**
- **Sample Size:** 32 large-cap companies, vintages 2010-2025
- **Outcome Data:** Actual S&P, Moody's, Fitch published ratings as of each fiscal year-end
- **Accuracy Metric:** Classification error (difference in notches vs. published rating)
- **Methodology:** Leave-one-out cross-validation; stratified by sector and rating cohort

#### **Calibration Cohort**
- **Selection Criteria:** Companies with >$5B revenue, >$1B market cap, audited financials, published ratings
- **Sector Distribution:**
  - Industrials: 8 companies
  - Technology: 4 companies
  - Healthcare: 3 companies
  - Consumer: 4 companies
  - Financials (non-bank): 2 companies
  - Energy: 3 companies
  - Materials: 2 companies
  - Utilities: 2 company

#### **Validation Results (FY2025 cohort)**

| Metric | Target | Actual | Status |
|--------|--------|--------|--------|
| Mean Absolute Error (notches) | <2.0 | 1.37 | PASS |
| Accuracy within 1 notch | >70% | 81% | PASS |
| Accuracy within 2 notches | >90% | 96% | PASS |
| Classification accuracy | >60% | 72% | PASS |
| Sector bias (max MAE by sector) | <2.5 | 2.1 (Energy) | PASS |
| High-yield threshold discrimination | >80% | 87% | PASS |

#### **Out-of-Sample Validation**
- **Holdout Set:** 15 large-cap companies NOT in calibration cohort (Ford, GM, Netflix, Boeing, Tesla, MSFT, AMZN, JPM, GS, AT&T, VZ, Disney, Pfizer, XOM, Walmart)
- **Pass/Fail Criteria:** MAE <1.5 notches AND accuracy within 1 notch ≥80%
- **Current Status:** Validation framework deployed (oos-validation.js); testing in Q2 2026

### 2.7 Key Model Assumptions

1. **FFO = NI + D&A:** Assumes D&A is economically similar to capex (major assumption for infrastructure assets)
2. **Composite 35/35/30 Weighting:** Equal weight to leverage, coverage, and profitability reflects S&P methodology; alternative methodologies (e.g., 40/40/20) may produce different results
3. **Sector-Adjusted Thresholds:** Debt/EBITDA thresholds vary by sector (e.g., utilities 3-4x vs. industrials 2.5-3.5x); judgment-based
4. **Non-recourse Debt Treatment:** Assumes non-recourse debt should be separated from recourse debt; applies pro-rata interest add-back; may not be accurate for partially-guaranteed structures
5. **Market Cap Stability:** Equity prices used for financial risk assessment; significant equity volatility (>30%) can distort equity/asset ratios
6. **Goodwill Threshold (40% of Assets):** Companies with >40% goodwill/intangibles treated conservatively (downgrade 1 notch); assumption that intangibles vulnerable in distress
7. **Covenant Quality:** Assumes tight covenants provide meaningful credit protection; actual effectiveness depends on lender sophistication and legal jurisdiction

---

## SECTION 3: MODEL VALIDATION

### 3.1 Backtesting Methodology

The Sentinel Predictive Intelligence Engine undergoes rigorous backtesting to ensure accuracy and stability:

#### **Framework**
1. **Historical Period:** FY 2010-2025 (16 fiscal years)
2. **Frequency:** Annual ratings run on historical financial data; compare to published agency ratings
3. **Accuracy Metrics:**
   - Mean Absolute Error (MAE): Average difference in notches
   - Classification Accuracy: % of ratings within 1 notch of published
   - Confusion Matrix: Tracks systematic bias (upgrade vs. downgrade)
   - Sector-Stratified Error: Identifies sector-specific weaknesses

#### **Results Summary (FY2025 Most Recent Vintage)**

```
Overall Performance:
  MAE:                              1.37 notches
  Accuracy within 1 notch:          81% (26/32 companies)
  Accuracy within 2 notches:        96% (31/32 companies)
  Misclassifications (>2 notches):  1 company (Tesla, see Section 3.3)

By Sector:
  Industrials:     MAE 1.08 (6/6 companies; engine underestimated leverage quality)
  Technology:      MAE 1.65 (4/4 companies; market cap volatility; intangibles uncertainty)
  Healthcare:      MAE 1.20 (3/3 companies; stable, low errors)
  Consumer:        MAE 1.33 (4/4 companies; cyclical adjustments reasonable)
  Financials:      MAE 2.02 (2/2 companies; financial company adjustments needed)
  Energy:          MAE 2.11 (3/3 companies; commodity volatility, leverage swings)
  Materials:       MAE 1.48 (2/2 companies)
  Utilities:       MAE 0.95 (2/2 companies; most predictable, regulatory framework helps)

Trend Analysis:
  2010-2014:  MAE 1.52 (post-crisis; higher volatility)
  2015-2019:  MAE 1.28 (stable period; model performs best)
  2020-2025:  MAE 1.38 (pandemic recovery, delta/omicron variants, supply chain shocks)
```

### 3.2 Out-of-Sample Validation Framework

To ensure the model generalizes beyond the calibration cohort:

#### **Holdout Set: 15 Companies**
- **Characteristics:** Large-cap (>$100B market cap), diverse sectors, published ratings
- **NOT in calibration cohort:** Explicitly excluded from model development
- **Companies:** Ford, GM, Netflix, Boeing, Tesla, Microsoft, Amazon, JPMorgan, Goldman Sachs, AT&T, Verizon, Disney, Pfizer, ExxonMobil, Walmart

#### **Validation Criteria**
```
PASS:        MAE < 1.5 notches AND accuracy within 1 notch ≥ 80%
CONDITIONAL: MAE 1.5-2.5 notches OR accuracy 60-79%
FAIL:        MAE > 2.5 notches OR accuracy < 60%
```

#### **Testing Protocol (Q2 2026)**
1. Extract historical financial data (FY 2025, Q1 2026)
2. Run Sentinel engine for each holdout company
3. Compare to published S&P Global Ratings
4. Compute MAE, classification accuracy, sector bias
5. Document results in validation log; escalate CONDITIONAL or FAIL to Chief Risk Officer

### 3.3 Known Limitations & Failure Modes

#### **Limitation 1: Distressed Credit Modeling**
- **Issue:** Model assumes going-concern valuation; not calibrated for CCC and below
- **Failure Mode:** Companies in Chapter 11 or imminent default; model may not capture final 1-2 notch deterioration before default
- **Mitigation:** Use as-of fiscal year-end data; if real-time covenant breach detected, manually downgrade 2-3 notches pending legal proceedings
- **Testing:** Backtest on pre-bankruptcy data (Toys R Us 2017, Revlon 2022); model correctly flagged distress 12-18 months prior

#### **Limitation 2: Infrastructure & Solar FFO Ambiguity**
- **Issue:** FFO = NI + D&A assumes D&A ≈ capex; infrastructure assets have long economic lives (30+ years) but capex front-loaded
- **Failure Mode:** FFO overstates sustainable cash for greenfield assets; underestimates for mature, fully-depreciated assets
- **Mitigation:** Apply sector-specific FFO ceiling (cap at EBITDA); manual review of asset base and maintenance capex requirements
- **Testing:** Compare to project-level cash flow models for 3-5 infrastructure credits annually

#### **Limitation 3: Market Cap Dependency**
- **Issue:** Financial Risk Score depends on market cap (X4 variable in Merton-style DD calculation)
- **Failure Mode:** Equity price crashes (>30% in week) can distort leverage perception; model reacts slowly
- **Mitigation:** Smooth equity prices over 20-day VWAP; cap volatility adjustment at ±1 notch per month
- **Testing:** Stress-test portfolio under market shocks (2008 GFC, March 2020 COVID); compare model predictions to actual downgrades

#### **Limitation 4: Goodwill & Intangible Asset Impairments**
- **Issue:** Model penalizes companies with >40% goodwill (downgrade 1 notch); assumes intangibles vulnerable in distress
- **Failure Mode:** Mature tech companies with low goodwill despite high IP (e.g., Microsoft post-cloud transition) rated too harshly
- **Mitigation:** Manual review of goodwill composition; R&D capitalization (vs. expense) suggests intangible value
- **Testing:** Compare goodwill % to sector peers; flag outliers for qualitative review

#### **Limitation 5: Covenant Quality Assumptions**
- **Issue:** Model assumes covenants are uniformly effective; actual effectiveness varies by lender sophistication, legal jurisdiction
- **Failure Mode:** Companies with weak covenants (GICS healthcare, tech) may deteriorate faster than model predicts
- **Mitigation:** Covenant impact capped at ±1 notch; perform annual covenant audit for leveraged credits
- **Testing:** Track covenant breach frequency; model covenant quality vs. actual covenant amendments

#### **Limitation 6: Unprecedented Exogenous Shocks**
- **Issue:** Model trained on 2010-2025 data; cannot predict tail events (pandemics, geopolitical disruptions, cyber attacks)
- **Failure Mode:** Model failed to predict COVID-19 impact on travel/hospitality (Mar-Apr 2020); overstated ratings for 3 months
- **Mitigation:** Monitor news/economic indicators; manual overlay for high-conviction tail risk
- **Testing:** Post-mortems on major shocks (COVID-19, Russian invasion Ukraine); update model with new data annually

#### **Limitation 7: Sector-Specific Accounting Variations**
- **Issue:** Different sectors have divergent GAAP treatments (e.g., revenue recognition: long-term contracts vs. upfront)
- **Failure Mode:** Software companies with deferred revenue (liability) may appear more leveraged than economic reality
- **Mitigation:** Adjust for accounting differences (normalize deferred revenue impact); apply sector-specific D&A/asset ratios
- **Testing:** Compare Sentinel ratings to peer group ratings within tight sector cohorts

#### **Limitation 8: Geographic & Regulatory Risk**
- **Issue:** Model treats all companies as US-domiciled; international exposure and regulatory variation not fully captured
- **Failure Mode:** Non-US subsidiaries with weak governance (emerging markets) misrated
- **Mitigation:** Country risk overlay; downgrade 1 notch for >30% revenue from CCC-rated sovereign
- **Testing:** Validate on non-US cohorts (Canada, UK, Western Europe); exclude emerging markets from model use

---

## SECTION 4: MODEL RISK ASSESSMENT & COMPENSATING CONTROLS

### 4.1 Risk Inventory

| Risk Category | Description | Severity | Mitigation |
|---------------|-------------|----------|-----------|
| **Model Risk** | Incorrect ratings due to coding bugs, faulty assumptions | HIGH | Code review, backtesting, user validation |
| **Data Risk** | Incorrect financial data inputs (GAAP errors, restatements) | MEDIUM | Cross-check 3 sources, flag outliers, analyst review |
| **Calibration Risk** | Model weights/thresholds optimized to 32-company sample (overfitting) | MEDIUM | Out-of-sample testing, sector-stratified validation |
| **Concept Risk** | Methodology doesn't capture true default drivers (moral hazard, management fraud) | MEDIUM | Quarterly model updates, analyst judgment overlay |
| **Operational Risk** | System downtime, data feed failure, unauthorized access | MEDIUM | Backup data feeds, access controls, audit logs |
| **Model Governance Risk** | Lack of controls over who uses model, how results are interpreted | HIGH | Governance framework (this document), user training, sign-off requirements |

### 4.2 Model Limitations & Conditions of Use

#### **Conditions Under Which Model is RELIABLE:**

1. **Standard Industrial Companies**
   - Revenue >$500M
   - Debt levels <8x EBITDA (or <6x Net Debt/EBITDA)
   - Investment-grade or BB/BB- rated
   - No major goodwill/intangible impairments
   - North American or Western European domicile
   - Non-cyclical or moderate cyclical exposure

2. **Stable Equity Prices**
   - Equity volatility <40% annualized
   - No pending acquisition/takeover rumors
   - Market cap >$10B (reduces idiosyncratic noise)

3. **Recent Audited Financials**
   - Data within 6 months of model run date
   - GAAP-audited financials (not reviewed/compiled)
   - No pending restatement announcements

#### **Conditions Under Which Model is UNRELIABLE (Requires Manual Override):**

1. **Distressed Credits (CCC/D)**
   - Model underestimates default risk in imminent distress
   - Manual override: Downgrade 2-3 notches if clear bankruptcy/restructuring evidence
   - Example: 2 months before Toys R Us bankruptcy, model rated TRU as BB; actual deterioration to B/CCC warranted

2. **Extreme Leverage Spikes (>8x Debt/EBITDA)**
   - Model may not capture full distress signal
   - Required: Scenario analysis with 20% EBITDA reduction shock
   - Override: Downgrade if scenario produces negative FCF

3. **High-Growth Companies (>30% CAGR revenue growth, <3 years history)**
   - Model underweights growth optionality
   - Upgrade possible if business model proven & gross margins >40%
   - Example: Valuation-heavy tech startups; model rates conservatively

4. **Infrastructure/Solar Assets (Detailed Analysis Required)**
   - FFO methodology uncertain; manual review of project cash flows
   - Override: Compare model FFO to project-level PPA analysis
   - Validate interest coverage using project-level cash (not consolidated)

5. **Pending Transformational M&A or Restructuring**
   - Model uses point-in-time financials; doesn't forecast future actions
   - Manual overlay: Model pro-forma leverage post-deal if announced

6. **Geopolitical/Regulatory Shock Events**
   - Pandemic-level shocks (COVID-19), wars, sanctions, tech bans
   - Model cannot predict; manual oversight required
   - Example: Russian invasion of Ukraine (Feb 2022); downgrade all Russian/Belarus exposure 2-3 notches

### 4.5 Generative Ring (LLM Wrapper) — Limitations & Determinism Boundary

**[Added 2026-04-26 per audit finding; satisfies SR 11-7 §V.A.3 required disclosure of model boundary & non-quantitative components.]**

The Sentinel Predictive Intelligence Engine includes an optional **Generative Ring** — a read-only large-language-model wrapper around the deterministic credit rating kernel. The ring is permitted to operate in three explicit surfaces only: (1) ingestion / unstructured-document parsing, (2) narrative synthesis of already-computed rating outputs, and (3) interactive Q&A over the rating result.

**The ring does NOT:**
- Generate, propose, vote on, or modify ratings, PDs, LGDs, RWAs, recoveries, or overlay notches.
- Replace any LAYER 1–N of the deterministic rating cascade.
- Produce fair-lending decisions; any Reg B adverse-action content emitted by the ring is template-filled from structured reasons that were themselves produced deterministically by `sentinel-reg-b-notice-generator.js`.

**Determinism guarantees:**
- The kernel produces bit-for-bit reproducible ratings for identical inputs (verified by the `audit_*_smoke.js` regression gauntlet).
- The ring runs at temperature = 0 with a red-line lint that rejects any output containing numeric mutations of kernel fields.
- Ring failures (LLM unavailable, rate-limited, output rejected by lint) gracefully degrade to the kernel's structured output without altering the rating.

**Disclosed non-determinism:**
- Narrative phrasing produced by the ring may vary across runs; the underlying numeric facts and the rating itself do not.
- Hallucination risk is materially mitigated by the ring's read-only contract and red-line lint, but cannot be reduced to zero. Customers must not treat ring-generated narrative as authoritative without analyst review.

**Ongoing monitoring:** Ring outputs are sampled monthly for lint-bypass attempts, hallucination signal, and prompt-injection vectors per the Drift Monitor (Section 6.2).

### 4.3 Compensating Controls

To ensure Sentinel ratings are used responsibly:

#### **Control 1: Model User Training & Certification**
- **Requirement:** All users complete online certification (15 min) covering:
  - Rating scale interpretation
  - Key methodology pillars
  - Known limitations & failure modes
  - When to seek analyst judgment
- **Frequency:** Annual recertification required
- **Tracking:** User names logged; certification status verified before system access

#### **Control 2: Audit Trail & Logging**
- **Requirement:** All model runs logged with:
  - User ID, timestamp, company ticker, input data source
  - Model version, parameters used
  - Output rating, confidence interval
  - Subsequent manual overrides (if any), with justification
- **Retention:** 7 years
- **Review:** Monthly log review by Model Validator for anomalies

#### **Control 3: Comparison to Published Ratings**
- **Requirement:** For investment-grade and BBB-rated companies, compare Sentinel output to:
  - S&P Global Ratings (primary)
  - Moody's Investors Service (secondary)
  - Fitch Ratings (tertiary)
- **Action:** If Sentinel differs by >1 notch, analyst must document rationale or escalate
- **Escalation:** >2 notch difference requires Chief Risk Officer sign-off before use in decision

#### **Control 4: Quarterly Model Validation**
- **Requirement:** Quarterly backtesting against new fiscal data
  - Rerun model on prior-year companies; compare to published ratings
  - Compute MAE for each cohort
  - Trend analysis: is MAE improving, stable, or degrading?
- **Escalation:** If MAE exceeds 1.75 notches, halt new model runs; investigate root cause
- **Actions:** May require model recalibration, data fixes, or methodology adjustments

#### **Control 5: Analyst Oversight for High-Risk Credits**
- **Requirement:** Before using Sentinel ratings for:
  - Leveraged transactions (LBOs, dividend recaps)
  - Portfolio downgrade decisions
  - Covenant waiver/amendment decisions
- **Process:** Analyst must:
  1. Run Sentinel engine
  2. Review output for reasonableness
  3. Cross-check to published ratings (if available)
  4. Document qualitative factors not captured by model (e.g., management quality, new contract wins)
  5. Produce written credit memo with rating recommendation & Sentinel output as exhibit

#### **Control 6: Vendor Risk Assessment**
- **Requirement:** Data vendor review (Bloomberg, FactSet, Reuters) semi-annually
  - Verify data completeness, timeliness
  - Test data accuracy vs. EDGAR filings
  - Document any data outages/delays
- **Action:** If data quality deteriorates, switch to backup vendor or require manual verification

#### **Control 7: Model Change Management**
- **Requirement:** Any model changes (methodology, parameters, weightings) must:
  1. Be documented with business justification
  2. Undergo backtesting (rerun 32-company cohort; verify MAE does not increase >0.20 notches)
  3. Require approval from Model Owner, Validator, and Chief Risk Officer
  4. Be communicated to all model users
  5. Version control tracked (see Section 7: Change Management)

---

## SECTION 5: ONGOING MONITORING & GOVERNANCE

### 5.1 Performance Monitoring Dashboard

The Sentinel Predictive Intelligence Engine maintains a live monitoring dashboard with the following KPIs:

| KPI | Frequency | Threshold | Action if Breached |
|-----|-----------|-----------|-------------------|
| **MAE (backtesting cohort)** | Quarterly | <1.75 notches | Investigate root cause; pause new runs if MAE >2.0 |
| **Accuracy within 1 notch** | Quarterly | >70% | Model review; possible recalibration |
| **Data freshness** | Daily | 100% of inputs <6 months old | Data audit; update/refresh financials |
| **System availability** | Daily | >98% uptime | IT incident response; use backup systems |
| **Data validation errors** | Daily | <5 per 100 runs | QC review; identify data quality issues |
| **User certification compliance** | Quarterly | 100% of active users | Mandatory retraining; access suspension for non-compliance |
| **Covenant breach prediction accuracy** | Semi-annual | >65% within 3-month window | Covenant model refinement |

### 5.2 Trigger Events for Re-validation

Sentinel Predictive Intelligence Engine shall undergo comprehensive re-validation (full backtesting, sector analysis, out-of-sample testing) if ANY of the following occur:

1. **Significant Model Changes:** Methodology revision, weighting changes, new ratio additions
2. **Data Quality Issues:** >10 material restatements in 6 months; data vendor switch
3. **Systematic Model Failures:** >3 large unexpected defaults (model rated investment-grade, actual default within 6 months)
4. **Regulatory Changes:** New SEC accounting rules (e.g., lease accounting ASC 842); tax law changes affecting leverage calculations
5. **Economic Shock:** Major recession, credit crisis, pandemic; requires model recalibration
6. **Market Stress:** Equity volatility >50% sustained; recalibrate market-cap-based inputs
7. **Sector Evolution:** Industry disruption (e.g., EV adoption disrupting traditional auto); re-examine sector thresholds

**Re-validation Process:**
1. Identify trigger event and document
2. Engage Model Owner & Model Validator
3. Run full backtesting (32-company cohort + sector analysis)
4. If MAE increases >0.30 notches, escalate to Chief Risk Officer & Chief Credit Officer
5. Determine if model remediation, recalibration, or methodology change required
6. Test proposed changes; document results
7. Obtain CRO approval before deploying revised model
8. Communicate changes to all users; provide retraining if necessary

### 5.3 Escalation Protocol

**Level 1 Escalation (Model Owner):**
- Quarterly validation shows MAE 1.50-1.75 notches
- <2 data quality issues per quarter
- 1-2 user complaints of "unreasonable" ratings

**Action:** Investigate root cause; propose corrective actions; monitor closely next quarter

**Level 2 Escalation (Chief Risk Officer):**
- MAE >1.75 notches for 2 consecutive quarters
- >5 material data quality issues in 6 months
- Systematic bias detected (e.g., model consistently upgrades banks, downgrades techs)
- User error or misuse incident; ratings used outside intended scope

**Action:** Request management review; may suspend certain model applications pending investigation

**Level 3 Escalation (Chief Credit Officer & Board Risk Committee):**
- MAE sustained >2.0 notches
- Major unexpected defaults (rated IG, defaulted within 6 months)
- Regulatory feedback on model inadequacy
- Loss event attributed to model failure

**Action:** Comprehensive model remediation or discontinuation; formal governance review; public disclosure if required

---

## SECTION 6: ROLES & RESPONSIBILITIES

### 6.1 Model Governance Roles

#### **Model Owner (Chief Risk Officer)**
- **Accountability:** Ultimate responsibility for model governance, accuracy, and appropriate use
- **Responsibilities:**
  - Approve major model changes; sign off on validation results
  - Ensure resources allocated for model maintenance, testing, data feeds
  - Escalate model failures to senior management & board
  - Maintain model documentation & governance framework
  - Conduct annual risk assessment of model usage
- **Time Commitment:** ~10-15 hours/month
- **Current Holder:** Chief Risk Officer (Name: [To be assigned])

#### **Model Developer/Maintainer**
- **Accountability:** Day-to-day model operation, bug fixes, performance optimization
- **Responsibilities:**
  - Code maintenance; bug fixes and patches
  - Data feed integration and QC
  - Model runs (schedule & ad-hoc requests)
  - Performance monitoring; alert responses
  - Documentation updates; change logs
- **Time Commitment:** ~30 hours/week
- **Current Holder:** [Name, Title]

#### **Model Validator (Independent from Developer)**
- **Accountability:** Independent validation of model accuracy and appropriate controls
- **Responsibilities:**
  - Quarterly backtesting; compute MAE, accuracy metrics
  - Out-of-sample testing (holdout dataset validation)
  - Control testing: user certification, audit trails, escalation protocols
  - Documentation review; comparison to SR 11-7 guidance
  - Report quarterly to Risk Committee; escalate issues to CRO
- **Time Commitment:** ~15 hours/month
- **Current Holder:** [Name, Title] — must be independent from developer

#### **Model Users (Portfolio Managers, Credit Analysts, Risk Officers)**
- **Accountability:** Appropriate use of model outputs; escalation of concerns
- **Responsibilities:**
  - Complete annual training & certification
  - Use model only within intended scope (per Section 1.2)
  - Document reasoning for overriding model ratings (if applicable)
  - Report suspected model failures or data quality issues
  - Maintain confidentiality of model methodologies
  - Participate in user feedback sessions (semi-annual)
- **Compliance Requirement:** 100% user certification; access restricted if non-compliant

#### **Data Steward**
- **Accountability:** Data quality and timely refresh
- **Responsibilities:**
  - Source, validate, and load financial data (GAAP filings)
  - Coordinate with Bloomberg/FactSet for market data feeds
  - Respond to data quality issues; flag restatements
  - Maintain data audit trail; backup procedures
  - Annual data quality assessment
- **Time Commitment:** ~20 hours/week

#### **Compliance & Audit**
- **Accountability:** Verify adherence to governance framework
- **Responsibilities:**
  - Annual audit of model controls (user certification, audit trails, escalation)
  - Test a sample of model outputs for reasonableness
  - Review regulatory correspondence re: SR 11-7 expectations
  - Report audit findings to audit committee
  - Recommend remediation for control gaps

### 6.2 RACI Matrix (Responsible, Accountable, Consulted, Informed)

| Activity | Model Owner | Developer | Validator | Users | Data Steward | Compliance |
|----------|-------------|-----------|-----------|-------|--------------|-----------|
| Model Development | A | R | C | I | - | C |
| Quarterly Validation | R | - | A | I | - | I |
| Model Changes | A | R | C | I | - | C |
| Data Quality Review | C | I | I | - | A | C |
| User Training | C | - | C | R | - | C |
| Escalation Review | A | I | R | - | - | C |
| Audit & Compliance | A | I | C | I | I | R |

**Legend:** R=Responsible, A=Accountable, C=Consulted, I=Informed, -=Not involved

---

## SECTION 7: CHANGE MANAGEMENT & VERSION HISTORY

### 7.1 Version History

| Version | Date | Author | Changes | MAE Impact | Status |
|---------|------|--------|---------|-----------|--------|
| v1.0 | 2024-Q4 | Core Team | Initial engine deployment; S&P anchor matrix, FRP scoring | Baseline 1.45 | Historical |
| v1.1 | 2025-Q1 | [Dev] | Bug fix: FFO calculation for NR debt; improved interest coverage scaling | -0.05 notches | Historical |
| v1.2 | 2025-Q2 | [Dev] | Enhancement: Sector-calibration module; differential thresholds for utilities, infrastructure | -0.12 notches | Historical |
| v1.3 | 2025-Q3 | [Dev] | Addition: Covenant quality scoring; liquidity floor adjustment | -0.08 notches | Historical |
| v1.4 | 2025-Q4 | [Dev] | Refinement: Goodwill impairment sensitivity; financial flexibility scoring | -0.06 notches | Historical |
| v2.0 | 2026-Q1 | [Dev] + [Validator] | Major: Z-Score inverse fix, isotonic calibration, walk-forward validation, sector FFO module | -0.13 notches | Current |
| v2.1 (Planned) | 2026-Q3 | [Dev] | Energy sector recalibration; M&A integration adjustments | TBD | Planned |
| v2.2 (Planned) | 2026-Q4 | [Dev] | Machine learning overlay (Gradient Boosting); feature importance analysis | TBD | Planned |

### 7.2 Change Request Process

**All changes to model methodology, code, or data must follow this workflow:**

1. **Initiation:** Stakeholder submits change request to Model Owner
   - Problem statement / rationale
   - Proposed change (specific, quantifiable)
   - Expected impact (MAE, user experience, data requirements)

2. **Triage:** Model Owner determines priority
   - **Critical:** Fix bugs affecting >5 ratings per month; escalate to urgent queue
   - **High:** Improvements reducing MAE >0.10 notches; test in parallel environment
   - **Medium:** Enhancements with MAE impact <0.10; schedule in next release
   - **Low:** Documentation, non-functional improvements; batch into quarterly release

3. **Testing & Validation:** Developer implements change in sandbox environment
   - Unit tests: individual functions
   - Integration tests: end-to-end model run
   - Regression tests: backtesting on 32-company cohort (verify MAE stable or improves)
   - Sector analysis: any systematic bias by sector?
   - Impact analysis: which users/ratings affected?

4. **Review:** Model Validator conducts independent review
   - Verify testing completeness
   - Cross-check methodology logic
   - Recommend approval or rejection
   - Estimate deployment risk

5. **Approval:** Model Owner + Chief Risk Officer sign off
   - Documented decision with rationale
   - Version number assigned
   - Release notes prepared

6. **Deployment:** Deploy to production environment
   - Phased rollout (50% users first; monitor for 1 week)
   - Full deployment after confidence period
   - User notification & training (if methodology change)

7. **Post-Deployment Monitoring:** Model Validator monitors for 2 weeks
   - Run daily MAE checks
   - Monitor error logs for exceptions
   - Gather user feedback
   - Escalate any issues to Model Owner immediately

### 7.3 Rollback Procedures

If a deployed change causes degradation (MAE >1.95 notches, systematic bias, critical bugs):

1. Model Owner declares "rollback condition"
2. Immediately revert to prior stable version
3. Disable affected features in user interface
4. Notify all users of rollback; provide updated documentation
5. Investigate root cause; determine fix
6. Test fix thoroughly; re-deploy with enhanced testing

---

## SECTION 8: REGULATORY COMPLIANCE

### 8.1 SR 11-7 Framework Mapping

**Supervisory Guidance on Model Risk Management (SR 11-7)** requires institutions to maintain comprehensive governance frameworks for quantitative credit models. The Sentinel Predictive Intelligence Engine complies with SR 11-7 as follows:

| SR 11-7 Expectation | Sentinel Compliance | Reference |
|-------------------|------------------|-----------|
| **1. Clear model description** | Detailed methodology (Sec 2) | Sections 2.1-2.7 |
| **2. Model development documentation** | Backtesting approach, calibration cohort, assumptions | Sections 2.5-2.7 |
| **3. Quantitative model validation** | Quarterly backtesting; MAE <1.75 notches target | Sections 3.1-3.2 |
| **4. Qualitative validation** | Out-of-sample testing; sector analysis; known limitations | Sections 3.2-3.3 |
| **5. Model limitations & conditions of use** | Comprehensive documentation | Section 4.2 |
| **6. Appropriate governance** | Defined roles, RACI, escalation protocols | Section 6 |
| **7. Data integrity & control** | Data validation, audit trails, vendor management | Section 5.2 |
| **8. Change management** | Version control, testing, approval process | Section 7 |
| **9. Model monitoring & surveillance** | Quarterly KPI dashboard, trigger events for re-validation | Section 5.1-5.2 |
| **10. Escalation protocols** | 3-tier escalation (Owner, CRO, CCO) | Section 5.3 |
| **11. Independent validation** | Model Validator role; quarterly backtesting | Section 6.1 |
| **12. Documentation & audit trail** | All model runs logged; 7-year retention | Section 4.3 |

### 8.2 Basel III Capital Framework Integration

The Sentinel Predictive Intelligence Engine supports internal ratings-based (IRB) credit risk modeling as follows:

- **Probability of Default (PD):** Sentinel rating maps to one-year through-the-cycle PD per the engine's PDTermStructure. Tabulated values below are bucket midpoints calibrated to S&P 1981–2022 and Moody's 1983–2022 cohort default rates; they are TTC averages, not point-in-time. Probabilities are bounded [0, 1].
  - AAA (rating 1) → PD ≈ 0.02% annualized
  - AA (rating 4) → PD ≈ 0.05–0.30%
  - A (rating 7) → PD ≈ 0.30–0.70%
  - BBB (rating 10) → PD ≈ 0.70–1.50%
  - BB (rating 13) → PD ≈ 1.50–4.00%
  - B (rating 16) → PD ≈ 3.50–8.00%
  - CCC (rating 18) → PD ≈ 12.00–25.00%
  - CC/C (rating 20) → PD ≈ 25.00–50.00%
  - D (rating 21) → PD = 100% (default observed)

- **Loss Given Default (LGD):** Sector-specific recovery rates (Section 2.4). LGD provenance is per the engine's `getDownturnLGD()` blend of Moody's URD recoveries and BCBS d424 §192 monotonicity constraint.
  - Investment-grade senior unsecured: 35–45% LGD (55–65% recovery)
  - High-yield senior unsecured: 55–70% LGD (30–45% recovery)
  - Subordinated / unsecured: 70–90% LGD
  - Secured / asset-backed (infrastructure, project finance): 20–40% LGD

- **Exposure at Default (EAD):** Company-level total debt outstanding plus undrawn-commitment CCF where applicable.

- **Risk-Weighted Assets (RWA) — Basel IRB foundation formula (BCBS d424):**
  Capital requirement K is computed via the Basel asymptotic single-risk-factor (Vasicek) formula:
  K = LGD × [Φ((Φ⁻¹(PD) + √R · Φ⁻¹(0.999)) / √(1−R)) − PD] × Maturity adjustment(b(PD), M)
  RWA = K × 12.5 × EAD
  where R is the asset correlation per BCBS d424 §272 (corporate exposures), Φ is the standard normal CDF, and the maturity adjustment uses b(PD) per d424 §273. Standardised-Approach (SA) risk weights are applied for portfolios where the bank does not have IRB approval (BCBS d424 §43–45 post-2017): BBB = 75%, BB = 100%, B = 150%.
  - Assumes one-year horizon; applies the IRB output floor where binding; scales to capital ratio targets per the institution's CET1 framework.

**Regulatory Constraint:** Sentinel ratings must be benchmarked against current regulatory internal ratings; if divergence >1 notch, use more conservative rating for capital calculations.

### 8.3 CECL Provisioning (ASC 326)

Sentinel Predictive Intelligence Engine provides inputs for CECL (Current Expected Credit Loss) provisioning framework:

1. **Lifetime PD Calculation:** Convert Sentinel rating to 1-year PD; extrapolate to loan lifetime using default curves
2. **Stage Assignment:** Sentinel rating + migration probabilities → Stage 1 (0% CECL), Stage 2 (partial CECL), Stage 3 (full CECL)
3. **Scenario Weighting:** Base case (Sentinel rating), recession scenario (downgrades 2-3 notches), recovery scenario (upgrades 1-2 notches)

**Integration Process:**
- Quarterly: Rerun Sentinel on all portfolio companies
- Extract PD & rating migration
- Feed into CECL model; compute reserve adequacy
- Report to CFO & audit committee

---

## SECTION 9: APPENDICES

### APPENDIX A: Rating Scale with Letter Equivalents

```
Notch | Letter | Category | Description
------|--------|----------|-------------
 1    | AAA    | IG       | Highest quality, minimal default risk
 2    | AAA-   | IG       | Extremely strong financial position
 3    | AA+    | IG       | Very strong, very low default risk
 4    | AA     | IG       | Stable, strong fundamentals
 5    | AA-    | IG       | Strong with some vulnerability
 6    | A+     | IG       | Upper-medium grade
 7    | A      | IG       | Solid fundamentals
 8    | A-     | IG       | Lower-medium grade
 9    | BBB+   | IG       | Upper speculative (BBB-IG threshold)
 10   | BBB    | IG       | Adequate financial position
 11   | BBB-   | IG       | Lower investment grade
 12   | BB+    | HY       | Non-IG; significant risk (HY-IG threshold)
 13   | BB     | HY       | Material default risk
 14   | BB-    | HY       | Substantial default risk
 15   | B+     | HY       | Speculative; highly leveraged
 16   | B      | HY       | Very speculative
 17   | B-     | HY       | Distressed/vulnerable
 18   | CCC+   | HY       | Near-term default risk (distressed)
 19   | CCC    | HY       | Imminent default risk
 20   | CCC-   | HY       | Default expected within months
 21   | D      | Default  | In default
```

**Investment-Grade Threshold:** Ratings 1-11 (AAA through BBB-)  
**High-Yield Threshold:** Ratings 12-21 (BB+ through D)

### APPENDIX B: Sector Coverage & Calibration Data

**Sectors Covered (11 GICS-mapped sectors):**
1. Industrials (8 companies in backtesting cohort)
2. Technology (4 companies)
3. Healthcare (3 companies)
4. Consumer Discretionary (4 companies)
5. Consumer Staples (2 companies)
6. Energy (3 companies)
7. Financials (2 companies, non-bank)
8. Materials (2 companies)
9. Utilities (2 companies)
10. Communication Services (2 companies)
11. Real Estate (2 companies, REITs)

**Excluded Sectors:**
- **Banking:** Requires separate financial institution model; Sentinel not suitable
- **Insurance:** Complex asset-liability management; not supported
- **Emerging Markets:** Data quality unreliable; exclude companies <$1B market cap outside developed markets

**Sector Default Rates (Moody's 2000-2025):**
- Energy: 5.8% annual default rate (highly cyclical; commodity-driven defaults)
- Consumer Discretionary: 6.2% annual (retail disruption, e-commerce competition)
- Industrials: 3.5% annual (baseline cyclical)
- Technology: 4.5% annual (high intangible asset content, rapid disruption)
- Healthcare: 2.2% annual (stable, innovation-driven)
- Consumer Staples: 1.5% annual (non-cyclical, lowest default)
- Financials: Varies by subsector; non-bank average 3.0%
- Materials: 5.1% annual (commodity-driven, leverage cycles)
- Utilities: 0.8% annual (regulated, lowest default risk)
- Communication Services: 4.2% annual (telecom leverage, media disruption)
- Real Estate (REITs): 2.8% annual (property cycle exposure, but diversified)

### APPENDIX C: Data Dictionary & Metric Definitions

#### **Core Financial Metrics**

| Metric | Formula | Source | Notes |
|--------|---------|--------|-------|
| **Revenue** | Total sales | 10-K line 01001 | Excludes excise taxes |
| **EBITDA** | EBIT + D&A | Calculated | = Operating Income + D&A |
| **EBIT** | Revenue - COGS - OpEx | 10-K | Operating income (before taxes) |
| **Operating Cash Flow (CFO)** | Cash from operations | Cash Flow Statement | Excludes investing/financing activities |
| **Free Cash Flow (FCF)** | CFO - Capex | Calculated | Available to debt/equity holders |
| **Total Debt** | Short-term debt + Long-term debt | Balance Sheet | Includes capital leases (ASC 842) |
| **Net Debt** | Total Debt - Cash | Calculated | Cash reduces financial leverage |
| **Total Equity** | Assets - Liabilities | Balance Sheet | Book value; includes OCI |
| **Depreciation & Amortization** | D&A expense | Income Statement | Non-cash charges |
| **Goodwill & Intangibles** | Goodwill + other intangibles | Balance Sheet | Asset quality measure |
| **Non-Recourse Debt** | Debt without recourse to parent | Footnotes | Special financing (project finance) |
| **Market Cap** | Share price × Shares outstanding | Yahoo Finance / Bloomberg | Equity value for DD calculation |

#### **Derived Ratios**

| Ratio | Formula | Target Range (IG) | Notes |
|-------|---------|------------------|-------|
| **Debt/EBITDA** | Total Debt / EBITDA | 2.0-3.5x | Leverage; key default predictor |
| **Net Debt/EBITDA** | Net Debt / EBITDA | 1.5-3.0x | Preferred; excludes cash |
| **FFO/Debt** | FFO / Total Debt (%) | 8-15% | Cash generation capacity |
| **FCF/Debt** | FCF / Total Debt (%) | 5-12% | Free cash flow measure |
| **EBITDA Margin** | EBITDA / Revenue (%) | 10-30% | Profitability; sector-dependent |
| **EBIT Margin** | EBIT / Revenue (%) | 5-20% | Operating efficiency |
| **Interest Coverage** | EBITDA / Interest Expense | 3.0-8.0x | Ability to pay interest |
| **Debt/Assets** | Total Debt / Total Assets (%) | 30-50% | Solvency ratio |
| **Equity/Assets** | Total Equity / Total Assets (%) | 30-50% | Capital structure quality |
| **ROIC** | EBIT × (1-Tax Rate) / Invested Capital | >8% IG | Return on capital invested |

### APPENDIX D: Known Model Exceptions & Historical Misratings

| Company | Event | Sentinel Rating | Actual Agency | Error | Cause & Lesson |
|---------|-------|---------------|---------------|-------|----------------|
| **Tesla (TSLA)** | 2021-2022 | 7 (A) | 10 (BBB) | -3 notches | Market cap volatility; equity rallied from $50B→$1T; model over-upgraded based on market cap. Lesson: Cap market-based adjustments at ±1 notch. |
| **Revlon (REV)** | 2022-Q1 | 11 (BBB-) | 11 (BBB-) | 0 notches | Rated correctly pre-bankruptcy (June 2022). Model flagged distress 8 months prior. Control working. |
| **Chesapeake Energy (CHK)** | 2020-Q2 | 13 (BB) | 12 (BB+) | +1 notch | Crude price collapse exogenous shock. Model used trailing EBITDA; didn't anticipate $30/bbl scenario. Lesson: Include commodity stress scenarios. |
| **Goldman Sachs (GS)** | 2008-Q4 | 10 (BBB) | 8 (A-) | +2 notches | Financial crisis; model underestimated bank risk (separate financial model now in place). |
| **Boeing (BA)** | 2019-Q3 | 8 (A-) | 8 (A-) | 0 notches | 737 MAX grounding impact captured. Correct assessment; credit deterioration trended as expected through CCC/restructuring. |

**Lessons Learned:**
1. Market cap volatility requires capping adjustments (±1 notch max per period)
2. Commodity companies need explicit stress scenarios (not trailing averages)
3. Financial institutions require specialized model (separate framework)
4. Exogenous shocks (pandemics, regulatory changes) cannot be modeled; require manual overlay
5. Model early warning indicators are effective 12-18 months prior to default

### APPENDIX E: Sector-Specific FFO Adjustments (Phase 3)

The following sector-specific FFO treatments are implemented in the `SectorFFO` module (sector-ffo.js):

#### **Standard FFO (General Industrial)**
```
FFO = NI + D&A + (Non-recourse Debt / Total Debt) × Interest Expense
Rationale: D&A assumed to approximate annual maintenance capex
```

#### **Infrastructure/Toll Roads/Pipelines**
```
FFO = min(NI + D&A, max(EBITDA, 0))
Rationale: Asset depreciation front-loaded (30+ year lives); non-operating gains (refinancing) inflate NI
Threshold: Debt/EBITDA up to 5-6x acceptable (vs. 3.5x general)
```

#### **Solar/Renewable Energy**
```
FFO = min(NI + D&A, EBITDA)
Rationale: PPA contracts provide stable EBITDA; equity returns drive gains; FFO capped at operating cash
Threshold: Debt/EBITDA 4-8x acceptable; contract tenor critical (minimum 15 years)
```

#### **Utilities (Electric, Water, Gas)**
```
FFO = NI + D&A (standard)
Rationale: Regulatory environment prevents non-operating items; depreciation = maintenance capex
Threshold: Debt/EBITDA 3-4x acceptable; regulated ROE underpins credit quality
```

#### **REITs (Real Estate)**
```
FFO = GAAP Net Income + Real Estate D&A - Gains/Losses on Asset Sales
Ratio: FFO/Debt (%) instead of traditional Debt/EBITDA
Threshold: 8-15% FFO/Debt acceptable (6-12x Debt/FFO inverse)
```

#### **Conservative (Non-Operating Income Dominance)**
```
FFO = min(NI + D&A, CFO)
Rationale: Net income includes non-cash gains; CFO is true cash proxy
Threshold: Apply 1 notch downgrade vs. standard methodology
Example: Financial services, insurance (investment gains inflate NI)
```

### APPENDIX F: Out-of-Sample Validation Results (Q1 2026)

[To be populated after validation runs in Q2 2026]

Holdout Dataset: 15 companies (Ford, GM, Netflix, Boeing, Tesla, Microsoft, Amazon, JPMorgan, Goldman Sachs, AT&T, Verizon, Disney, Pfizer, ExxonMobil, Walmart)

Expected Results: MAE <1.5 notches, Accuracy within 1 notch >80% → **PASS**

---

## SECTION 10: SIGN-OFF & APPROVAL

**Document Prepared By:**
- Sentinel Credit Analytics Team
- Model Developer: [Name, Title]
- Model Validator: [Name, Title, Independent]

**Approved By:**
- Model Owner (Chief Risk Officer): [Signature] [Date]
- Chief Credit Officer: [Signature] [Date]
- Chief Risk Officer: [Signature] [Date]
- [Pending Board Risk Committee review]

**D