CECL Bond Loss Model

M-040 implements the ASC 326 Current Expected Credit Loss (CECL) standard for the fixed-income (bond) portfolio. CECL replaced the prior incurred-loss model with a lifetime expected-loss measure: at every reporting date the allowance equals the credit losses expected over the entire remaining contractual life of each instrument, recognized up front rather than when a loss event becomes probable.

The engine is FinView-resident (cross-repo): the implementation is finlib.cecl.CECLCalculator (governance id finview.L3.cecl_calculator) in [redacted], consuming the credit-parameter library finlib.credit_data (governance id finview.L0.credit_data). The CECL allowance it produces is consumed by the insurance capital/reserve stack (the bond-portfolio expected-loss feeds invested-asset impairment and capital adequacy), but the source code does not live in InsModel — see Limitation 1.

The model uses a discounted-cash-flow loss method built on a PD × LGD × EAD decomposition rendered period-by-period (ASC 326-20-30-4 permits a DCF method; the period losses are PV-discounted at the instrument's effective interest rate):

1. Government/agency carve-out. Instruments rated AAA/AA+ and in sector Government/Agency are assigned zero ECL (_is_government_bond), reflecting the ASC 326 practical recognition that US sovereign/agency exposures carry no expected credit loss.

2. LGD (loss given default). LGD = 1 − recovery_rate. Recovery defaults by seniority/sector from RECOVERY_RATES: senior-secured 0.65, senior-unsecured 0.45 (the corporate-bond default), subordinated 0.25, government/agency 0.90. A generic BBB corporate is therefore LGD = 0.55.

3. PD / hazard rate. With pd_source='moody', get_counterparty_pd maps the S&P-scale rating to its Moody's equivalent (BBB → Baa) and reads the Moody's published cumulative-default table MOODY_CUMULATIVE_PD (1y/3y/5y/10y), interpolating to the bond's maturity. A continuous hazard rate is backed out as λ = −ln(1 − PD_cum) / T. With pd_source='oas' the hazard is instead derived from live FRED/OAS spreads (λ = spread / (1 − recovery)); the snapshot uses the deterministic Moody's path so no live data is touched.

4. Stress multiplier. The hazard is scaled by a scenario factor from PD_MULTIPLIERS (base 1.0, mild 1.5, stress 2.0, severe 3.0) before cash-flow generation, giving CCAR-style scenario sensitivity.

5. Cash-flow loss schedule (EAD). BondCashflows.generate_schedule(hazard_rate, recovery_rate) produces a period-by-period schedule carrying default_amount and recovery per coupon period (semiannual by default). The period loss is default_amount − recovery, and each is PV-discounted at the effective interest rate (the coupon rate as the ASC 326 EIR proxy): pv_loss = loss / (1 + r)^t.

6. Lifetime ECL & allowance. lifetime_ecl = Σ pv_loss over all periods to maturity; ecl_rate = lifetime_ecl / face_value. At portfolio level compute_portfolio_ecl sums per-bond ECL and rolls it up ecl_by_rating and ecl_by_sector. The lifetime ECL is the CECL allowance.

Prose. The two load-bearing assumptions are the PD table (K-1) and LGD (K-3). The Moody's cumulative-default table is the spine of the lifetime measure: PD at maturity is interpolated from the four published tenor points (e.g. Baa 10y = 3.27%) and converted to a continuous hazard. Because these are through-the-cycle averages, the model satisfies CECL's reversion requirement (K-6) almost by construction — there is no explicit R&S-then-revert two-segment curve; the single Moody's curve is already the long-run mean. The reasonable-and-supportable horizon (K-5) is the full contractual life rather than the more common 1–2 year R&S window, which is conservative and avoids a reversion-segment modeling choice but reduces sensitivity to near-term macro forecasts. LGD is seniority-bucketed, not name-level — a known simplification (Limitation 2).

Deterministic run of CECLCalculator (finview.L3.cecl_calculator, FinView) — reproducible, requires no live firm data (BV-032 immune); Moody's published PD table, base scenario, no FRED/OAS live spreads. Reproduce with python scripts/model_snapshots.py M-040; the underlying engine is asserted by FinView tests/test_cecl.py (6 tests, all passing — test_bbb_10y_ecl_reasonable bounds this exact BBB 10Y bond's ECL to (5k, 30k) and asserts LGD = 0.55).

Input: canonical 2-bond portfolio — UST $1M par · 4.00% cpn · 10y · AAA/Government (zero-ECL) + BBB corporate $1M par · 5.25% cpn · 10y · senior-unsecured.

The undiscounted PD×LGD×EAD product for the BBB bond is 0.0327 × 0.55 × $1M ≈ $17,985; the reported $13,387 is lower because losses are spread across 20 semiannual periods and PV-discounted at the 5.25% EIR — exactly the ASC 326 DCF mechanic. The Treasury contributes zero by the government carve-out, so the portfolio allowance equals the single corporate ECL.

Captured 2026-06-04 · deterministic, no live data.

1. Cross-repo FinView ownership. The engine is FinView-resident (finlib.cecl.CECLCalculator, finview.L3.cecl_calculator), not InsModel. The insurance capital/reserve stack consumes its CECL allowance but does not own the code, version, or test suite. Governance crosses a repo boundary; any change-control, MRM ratification, and registry binding for M-040 must reference FinView artifacts, and the model card's authority over the implementation is by reference only.

2. LGD is seniority-bucketed, not name-level. Recovery is a static lookup by seniority/sector (RECOVERY_RATES), not calibrated per issuer, vintage, or collateral. Senior-unsecured 0.45 is a long-run market average; idiosyncratic recovery dispersion is not captured.

3. 14 S&P notches collapse to 7 Moody's buckets. The _SP_TO_MOODY crosswalk maps e.g. BBB+/BBB/BBB- all to Baa, so sub-notch credit differentiation is lost in the Moody's PD path. (The OAS path preserves notch detail via interpolation, but the deterministic snapshot uses Moody's.)

4. Q-07 — Moody's PD is vendor data, not a promoted vendor model. The cumulative-default table is hard-coded vendor data in finlib.credit_data.MOODY_CUMULATIVE_PD. It has not been promoted through the vendor-model governance process (no vendor-model registry entry, no independent validation packet, no provenance/version pin to a specific Moody's Annual Default Study release). It is consumed as a constant. This is an open governance gap (Q-07).

5. EIR proxied by coupon rate. ASC 326 requires DCF discounting at the instrument's effective interest rate; the engine uses the coupon rate (fallback 5%) as the EIR proxy. For bonds bought at a material premium/discount to par the true EIR diverges from the coupon, biasing the PV of losses.

6. No explicit reasonable-and-supportable / reversion segmentation. CECL contemplates a forecast-based R&S horizon that reverts to historical experience. The model uses through-the-cycle Moody's averages over the full life — effectively "reverted everywhere." This is defensible but means near-term macro forecasts and a documented reversion technique are absent; the scenario multiplier (K-8) is the only forward-looking lever.

7. No 10-K / live-firm validation. The snapshot is a deterministic engine demonstration on a canonical synthetic portfolio. No claim is made that these figures reconcile to any issuer's 10-K disclosed CECL allowance (BV-032 firm-data divergence is avoided by not using live data).

8. Mid-period timing / amortization simplifications. BondCashflows recognizes default and recovery at period end (not mid-period) and the canonical input is a bullet (amortization='bullet'); amortizing or prepaying (CPR) structures are supported by the engine but not exercised in the canonical snapshot.

Sanity invariant: PV-discounted lifetime ECL ($13,387) < undiscounted PD×LGD×EAD ($17,985), consistent with ASC 326 DCF discounting over the bond's life.

Aggregate-risk view (per MRM §10.5 item 8): the CECL allowance is consumed downstream by the insurance capital/reserve stack (invested-asset impairment / capital adequacy). M-040 shares upstream component-model families M-130 (bond pricing) and M-133 (credit-pricing hazard-rate) with M-052 (Counterparty Concentration & Network Risk) per model_registry.yaml — so a change in those shared pricing/credit-spread components propagates to both the CECL allowance here and counterparty exposure in M-052. (The finview.L0.credit_data engine itself is bound to M-040 only.)

Open MRM items: Q-07 (Moody's PD vendor-data promotion) and Limitation 1 (cross-repo FinView ownership / change-control binding) are unresolved and should be tracked as governance issues. A full validation evidence pack with outcomes back-testing of the CECL allowance vs realized credit losses (COND-001 / RAT-040), independent effective-challenge / ratification closure (RAT-040 → approved), Q-07 vendor-model promotion, and the INV-032 packaged-interface / CI drift gate for the FinView-resident engines are gated modeling/governance items tracked for ratification — out of scope for this documentation pass.

Regulatory / accounting: - FASB ASC 326 (CECL) — Financial Instruments — Credit Losses; ASC 326-20 (measurement of expected credit losses on amortized-cost assets), 326-20-30 (DCF method, discount at effective interest rate). - ASU 2016-13 — the standard that introduced CECL.

Vendor data: - Moody's Annual Default Study (1970–2023) — cumulative default rates by rating and the 1-year transition matrix, encoded in finlib.credit_data.MOODY_CUMULATIVE_PD / MOODY_TRANSITION_MATRIX. (Vendor data, not a promoted vendor model — Q-07.)

Engine source (FinView-resident, cross-repo): - [redacted] — CECLCalculator (finview.L3.cecl_calculator): compute_bond_ecl, compute_portfolio_ecl. - [redacted] — finview.L0.credit_data: MOODY_CUMULATIVE_PD, RECOVERY_RATES, _SP_TO_MOODY, get_counterparty_pd, _pd_from_moody, _pd_from_oas. - [redacted] — BondCashflows.generate_schedule (period-by-period default/recovery cash-flow loss schedule).

Tests: - [redacted] — 6 tests (government zero-ECL, BBB band, stress monotonicity, portfolio additivity, OAS path, rating monotonicity).

Internal: - BV-032 — firm-data divergence (avoided here by deterministic, no-live-data snapshot). - Q-07 — Moody's PD as vendor data not yet promoted to a governed vendor model.

Card change history. Code-side change history lives in git log of the component files.

• 2026-05-08 — stub created from registry data per Decision 023 Phase 5 / B-07.
• 2026-05/06 — Tier-1 hand-authoring of Methodology, Key Assumptions and Their Justification, Output Snapshot, Limitations and Known Gaps, Validation Packet, and References from the FinView CECL engine code (finlib/cecl.py, finlib/credit_data.py, finlib/bond_cashflows.py) + the deterministic model_snapshots.py M-040 run. Stub marker advanced to ``.
• 2026-06-06 — documentation pass: added Standards Coverage (ASC 326-20 / 326-20-30-4, ASOP 56, PCAOB AS 2501 mapped to the engine) and Dependencies (no upstream M-NNN models; relies on finview.L3.cecl_calculator + finview.L0.credit_data; shares M-130 / M-133 component families with M-052; allowance consumed by the capital/reserve stack) sections, and an Aggregate-risk view note (§10.5 item 8). Doc-currency baseline confirmed: stamp present (model_doc_stamps.yaml), status current, fingerprint 39e093d0aca90fe6, no identity/currency drift. Registry-grounded only; no model outputs changed.