GOES Economic Scenario Generator

M-120 is the GOES Economic Scenario Generator (Generator Of Economic Scenarios) — the prescribed multi-factor real-world ESG that the NAIC adopts in the 2026 Valuation Manual to replace the Academy Interest Rate Generator (AIRG) for principles-based reserve and capital work. GOES is the NAIC's custom calibration of Conning's GEMS engine; M-120 is the platform's independent, open re-implementation of that calibration (it does not embed Conning's proprietary GEMS code). It produces coherent stochastic paths for Treasury yields, equity, corporate credit, and inflation that feed nearly every Tier-1 reserve/capital model: VM-20 stochastic reserves (M-001), VM-21 CTE (M-002), and C-3 Phase I/II capital (M-050).

GOES lives in its own repository (ecosystem/goes, package goes) and is wired to InsModel through the goes_output_v1 contract: GOES writes six CSV files (treasury_yields, equity_gwf, bond_gwf, equity_returns, bond_returns, inflation), and InsModel's firmmodel/data_providers/goes_data_provider.py (GOESDataProvider) reads them and serves them to the calculation engines through the standard EngineDataProvider interface. The same contract is also consumed by Gold Copy (gold_copy/scenarios/goes_reader.py).

The generator (goes.L5.orchestrator, GOESGenerator in src/goes/generator.py) composes five sub-models on a single monthly clock (dt = 1/12), default horizon 30 years, default 10,000 scenarios, seeded for reproducibility:

1. Treasury — 3-factor CIR with deterministic shift (goes.L5.cir_treasury, CIRThreeFactorModel). Three independent Cox-Ingersoll-Ross factors dX_i = (κ_i − λ1_i)((θ_i + λ0_i) − X_i) dt + σ_i √X_i dW_i combine additively into the overnight rate, plus a deterministic shift function l(s) calibrated so the month-0 curve reproduces the initial market curve exactly (verified by test_initial_curve_match). Yields come from the closed-form CIR zero-coupon bond coefficients P_i(τ) = A_i(τ)·exp(−B_i(τ)·X_i). The factors carry NAIC-Q&A risk premia (λ0, λ1) splitting risk-neutral from real-world dynamics, an exponential shift decay (Decay = 3 per NAIC Q&A Q8), a permanent maturity-graded term premium, a fractional rate floor, weekly sub-stepping (4 sub-steps/month) for integration accuracy, and optional 2-regime volatility switching plus a B-coefficient yield_amplifier for tail widening. This is the sub-generator snapshotted below.
2. Equity — Stochastic-Volatility-with-Jumps (SVJ) (goes.L5.svj_equity, SVJEquityModel). A Bates-1996 model (Heston stochastic variance + Merton jumps): dS/S = (r(t) + ERP − λμ_J − ½V) dt + √V dW_S + J dN, dV = κ_v(θ_v − V) dt + σ_v √V dW_V, with leverage correlation ρ = −0.70. Drift is linked to the CIR overnight rate; dividends are linked to the CIR 10-year yield. Five secondary equity indices (Mid/Small Cap, NASDAQ, EAFE, EM) are derived from S&P 500 by a Sharpe-ratio / vol-ratio scaling.
3. Corporate credit (CreditSpreadModel). Mean-reverting (CIR-like, non-negative) credit spreads with an 8×8 Markov rating-transition matrix (AAA…CCC + Default), producing total/income/price returns for 8 bond funds layered on the Treasury curve.
4. Inflation / CPI (goes.L5.cpi_inflation, InflationModel). Mean-reverting inflation (θ ≈ 2.5%) linked to the short rate, producing a CPI index path and a real-short-rate path.
5. Global jump process (GlobalJumpProcess). A cross-asset Poisson tail event that simultaneously shocks equity down, widens spreads, and spikes volatility — reproducing the empirical rise in cross-asset correlation during crises.

The orchestrator runs them in dependency order on one RNG stream (global jump → Treasury → equity → credit → secondary equity → wealth factors → inflation), so the resulting scenario set is internally coherent across asset classes. Output is emitted in NAIC GOES CSV format and validated against the downloaded official NAIC/Conning scenario files via src/goes/naic_comparison.py (the --compare CLI flag).

GOES is parameter-driven via GOESConfig. The load-bearing assumptions, with the values shipped in the default production config:

Operational notes: The basic CIR model produces narrower yield distributions than full GEMS, so the shipped default applies a calibrated yield_amplifier = 4.3 (centered B-coefficient tail widening) together with a public 20% rate ceiling (rate_ceiling = 0.20, per GOES Acceptance Criteria 1A/1I; confirmed by the genuine NAIC 2024-12-31 file where the 3M P99 maxes ~15.4%). This lands the 10Y P5–P95 at year 10 at ~875bps vs the official NAIC target of ~891bps (within reconciliation tolerance), keeps all 26 primary acceptance criteria met (the rate_ceiling is a safety guardrail whose threshold traces to the public 20% criterion, not load-bearing — with the cap off, amp=4.3's 3M tail tops out ~14%, so pct_3m_above_20pct passes either way; 2L NB-120-01), and preserves the median (the amplifier is centered, so it does not shift P50). Optional 2-regime volatility switching (regime_params) remains off by default. The 10,000-scenario / 30-year defaults are the production run size; the snapshot below uses a reduced set for fast, deterministic documentation.

Deterministic run of the GOES CIR 3-factor Treasury sub-generator (goes.L5.cir_treasury, CIRThreeFactorModel) using the production default CIR factors and initial curve — reproducible, requires no live firm data (python scripts/model_snapshots.py M-120 in InsModel; the underlying engine is asserted by goes/tests/test_treasury.py, 32 passing tests). Because goes is a separate package, the snapshot inserts the GOES source tree on sys.path before importing.

Input: seed 42 · 200 scenarios · 120 monthly steps (10-year horizon) · default GOES initial curve (10Y = 4.400%)

The terminal 10Y distribution mean-reverts toward the CIR long-run level and stays positive (Feller + fractional floor). The P5–P95 band is now ~887bps at the 10-year horizon — landing on the official NAIC GOES 2024-12-31 target of 891bps — because the shipped default now applies the calibrated yield_amplifier = 4.3 plus the public 20% rate ceiling (COND-002 / INV-009 resolved; see Limitations item 3). The median (~4.6%) is preserved — the amplifier is centered on E[X_i(t)], so it widens the tail without shifting the median. These are scenario inputs only — they become reserves/capital only when consumed downstream by M-001/M-002/M-050.

Diff vs prior snapshot (2026-06-04, pre-calibration default yield_amplifier=1.0, no ceiling): 10Y P5 3.866% → 0.850%, P50 4.728% → 4.556% (preserved), P95 5.928% → 9.717%, P5–P95 ~206bps → ~887bps, min/max 3.584%/6.724% → 0.010%/13.139%. The widening to the NAIC band is the intended, output-changing calibration. Re-froze via python3 scripts/model_snapshots.py M-120 (InsModel) after setting the calibrated default.

Captured 2026-06-07 · deterministic, no live data · calibrated default (COND-002).

1. Status: under_development — not production-ratified. M-120 is an active re-implementation of the NAIC GOES calibration, not a frozen, 2L-ratified model. Its parameters, default toggles, and tail calibration are still moving. No production reserve or capital number should be certified off M-120 output until it is promoted to active and ratified.
2. Upstream-dependency cascade. M-120 is the prescribed scenario source for the full stochastic legs of M-001 (VM-20 CTE70), M-002 (VM-21 CTE), and M-050 (C-3 Phase I/II). Because M-120 is under_development, those consumers currently run their deterministic / 16-scenario-exclusion paths only and cannot certify a held full stochastic CTE reserve. M-120's status therefore cascades: until it is ratified, the downstream stochastic reserves of M-001/M-002/M-050 are blocked at the deterministic floor (see M-001 Limitation 6 / Output Snapshot). This is the single highest-leverage gap in the Tier-1 chain.
3. Tail width calibrated to the NAIC band (COND-002 resolved); residual extreme-tail shape gaps remain. The basic 3-factor CIR produces narrower yield distributions than full GEMS, so the shipped default now applies a calibrated yield_amplifier = 4.3 plus a public 20% rate ceiling (rate_ceiling = 0.20). At this default the 10Y P5–P95 is ~875bps vs the official NAIC ~891bps target (was ~206bps at the prior yield_amplifier=1.0 default), the official-file reconciliation is 444/481 (Treasury 60/60), and the acceptance scorecard is 35/40, 26/26 primary (the rate_ceiling is a public-criterion safety guardrail, not load-bearing for this calibration — the centered amp=4.3 3M tail tops out ~14%; 2L NB-120-01). The 5 remaining acceptance misses are extreme (1st/99th-percentile 20Y) tail-shape features that need time-varying κ/θ or regime switching, which the centered amplifier does not fully reproduce. A documented residual: bond-fund excess returns now reconcile 0/4 (the widened rate fan inflates the rate-driven return component, ~38–43bps over NAIC for Short/Long IG) — a known CIR-vs-GEMS limitation (see item 5 and docs/naic_validation_report.md §5), bounded overshoot, not a defect.
4. Risk premia λ0/λ1 are approximate, not Conning's. The exact Conning Lambda0/Lambda1 calibration is proprietary and unavailable; M-120 uses fitted approximations. This is a structural source of divergence from the official NAIC files at the distribution edges.
5. Bond-fund GWF residuals at long horizons. Short-duration bond funds (Money Market, Short Govt, Short IG Corp) overshoot NAIC GWF at the 20–30-year horizon because the CIR yield median runs ~10–14bps above NAIC and compounds; this is a calibration residual, not fixable by duration tuning alone.
6. No live-data calibration; immune to but separate from BV-032. The snapshot is purely synthetic (seeded RNG, no firm filings), so it is not exposed to the BV-032 firm-data divergence — but it also makes no 10-K / firm-specific claim. M-120 generates market scenarios only; it provides no mortality, lapse, expense, or firm-financial assumptions (GOESDataProvider.get_assumptions returns empty by design).
7. Cross-venv coupling. GOES is a separate repo/venv and is not installed in the InsModel venv; the InsModel-side snapshot only runs by path-inserting the GOES source tree. Production consumption is decoupled correctly via the goes_output_v1 CSV contract, but any in-process snapshot has this packaging caveat.

Ratification status. Resolved in the Decision 053 re-review loop: INV-009 / COND-002 (default-config calibration to the NAIC band) is applied — the shipped default is now yield_amplifier=4.3 + rate_ceiling=0.20, giving 10Y P5–P95 ~875bps vs the ~891bps target, with the gold snapshot re-frozen and the diff recorded above. COND-001 (official-file backtest) is cleared — the public NAIC/Conning 2024-12-31 statistics XLSX was acquired into ecosystem/goes/data/naic_official/20241231/, the 11 previously-skipped test_naic_comparison.py tests are live and passing, and --compare reconciles 444/481 (Treasury 60/60). COND-003 (doc-currency surface mapping) was cleared in the prior tranche (GOES engine files now resolve to the real ecosystem/goes repo shas). Still pending 2L: unlocking the held full stochastic reserves of M-001/M-002/M-050 (INV-010 cascade — gated on M-120 promotion + re-validation of the held reserves) and the lifecycle under_development → active promotion itself (2L re-review of this remediation; last_validated_on to be set on promotion).

Governance binding: engine IDs are stamped via the Decision 023 @engine("goes.L5.*") decorator (src/goes/governance.py); SR 11-7 narrative lives in src/goes/legacy_metadata/<ClassName>.yaml. No 2L production ratification exists yet (consistent with under_development).

Regulatory / actuarial: - NAIC GOES (Generator of Economic Scenarios) — prescribed ESG replacing the Academy Interest Rate Generator (AIRG) in the 2026 Valuation Manual for PBR reserve and capital work. - NAIC ESG Q&A (April 2021) — CIR factor dynamics, real-world constraints (Q11), shift decay (Q8, Decay=3). - American Academy of Actuaries ESG / AIRG — predecessor scenario generator being superseded by GOES. - Conning GEMS — the commercial ESG that the official NAIC GOES calibration is built on; M-120 is an independent re-implementation (GEMS code not embedded). - C-3 Phase I/II capital framework — downstream consumer of GOES scenarios (M-050).

Academic: - Cox, Ingersoll & Ross (1985) — the square-root short-rate model underlying the 3-factor Treasury generator. - Bates (1996) / Heston (1993) + Merton (1976) — stochastic-volatility-with-jumps equity dynamics.

Engine source (ecosystem/goes/src/goes/): - generator.py — GOESGenerator / GOESConfig orchestrator (goes.L5.orchestrator). - treasury/cir_three_factor.py — CIRThreeFactorModel (goes.L5.cir_treasury). - equity/svj_model.py — SVJEquityModel (goes.L5.svj_equity). - corporate/credit_model.py — CreditSpreadModel (8-fund / rating-transition credit). - inflation/cpi_model.py — InflationModel (goes.L5.cpi_inflation). - correlation/jump_process.py — GlobalJumpProcess (cross-asset tail jumps). - naic_comparison.py — official-file comparison harness.

Consumer / contract: - ecosystem/contracts/goes_output_v1.json (v1.0.0) — output CSV contract (producer goes; consumers insmodel, gold_copy). - InsModel/Models/firmmodel/data_providers/goes_data_provider.py — GOESDataProvider (InsModel ingestion). - gold_copy/scenarios/goes_reader.py — Gold Copy ingestion.

Tests: - goes/tests/test_treasury.py, test_equity.py, test_inflation.py, test_integration.py, test_acceptance.py, test_naic_comparison.py. - InsModel/Models/tests/test_goes_data_provider.py, test_projection_goes_integration.py, test_vm20_stochastic_goes.py.

Internal: - Decision 023 (Policy-Based Binding Architecture) — @engine decorators / goes.L5.* engine IDs. - BV-032 — firm-data divergence note; M-120 is synthetic-only and makes no firm-specific claim.

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.
• 2026-06-04 — Tier-1 hand-authoring of Methodology, Key Assumptions and Their Justification, Limitations and Known Gaps, Output Snapshot, Validation Packet, and References from GOES engine source (cir_three_factor.py, svj_model.py, credit_model.py, cpi_model.py, generator.py) + legacy_metadata + the model/contract registries. Stub marker advanced from to.
• 2026-06-06 — code-grounded documentation pass. Corrected the Components output-contract id from the non-existent goes_scenarios_v1 to the canonical goes_output_v1 (v1.0.0), reconciling Components with the card's existing Methodology / Validation Packet / References and with contract_registry.yaml (the registry model_registry.yaml line 1557 still lists goes_scenarios_v1 — registry fix left as governance work). Added a dedicated Standards Coverage section (asop_56 / naic_ag53 / sr_26_2 / internal, each grounded in the implementing source). Added a dedicated Dependencies section enumerating the downstream consumers from model_registry upstream_models edges (M-001/M-002/M-004/M-005/M-020/M-021/M-050/M-051/M-070/M-110). Expanded the stub Description to a self-contained paragraph. Noted gated/ratification items (INV-009, INV-010 cascade, validation-evidence pack + lifecycle promotion, doc-currency surface mapping) under Limitations. No model outputs, back-test numbers, or validation results were fabricated or changed.