Evidence-controlled model-risk research for SPX option pricing and validation.
DerivGuard-MRM develops and independently challenges a Heston option-pricing model against SVI/SSVI alternatives, with explicit controls for calibration uncertainty, numerical validation, economic materiality, stress testing, model disagreement, and distribution shift.
The accepted v1.1 scientific release passed an independent scientific/code re-review. It is research software only. It is not approved for live trading, production valuation, capital, or brokerage execution.
| Area | v1.1 evidence |
|---|---|
| Post-validation analytics | 9,720 contract-model observations across 126 market dates |
| Controlled stress repricing | 155,520 contract-scenario rows and 240 aggregate scenario rows |
| Ordinary locked DVE | FPR 0.069 vs 0.103 for FitOnly, with recall 1.000 for both |
| OOD challenge | Full DVE FPR 0.615 vs 0.154 for FitOnly |
| Winner uncertainty | 16 of 39 signed winner-vs-runner-up bootstrap intervals cross zero |
| Local qualification | 279 Pytest tests passed, including 16 numerical and 5 CUDA/GPU tests |
| Independent review | ACCEPTED |
| Published evidence | 64 machine-readable result files and 49 publication figures in the accepted internal release |
The central model-risk conclusion is intentionally qualified: multi-evidence DVE reduced false positives on the ordinary locked sample, but that improvement did not transfer robustly under OOD shift. The release therefore makes no claim of global DVE superiority or production readiness.
flowchart LR
A[SPX market data and synthetic controls] --> B[Data quality and settlement handling]
B --> C[Heston developer calibration]
B --> D[SVI / SSVI challengers]
C --> E[Independent numerical validation]
D --> F[Valuation and Greek disagreement]
E --> G[DVE evidence aggregation]
F --> G
G --> H[Materiality, slicing, monitoring, stress]
H --> I[Fresh LOCKED and OOD evaluation]
I --> J[Independent review and evidence integrity]
These figures show the project at the level where model-risk conclusions are actually made: not only aggregate fit, but where pricing error, Greek disagreement, and portfolio materiality concentrate across the option surface.
- Heston calibration diagnostics, bootstrap uncertainty, parameter correlation, profile-width analysis, and multistart stability.
- Economically defined moneyness, maturity, volatility, liquidity, settlement, and data-quality slices.
- Signed residual analysis and two-dimensional error surfaces.
- Qualified model-winner maps with date-block bootstrap uncertainty.
- Challenger attribution and explicit low-valuation-disagreement / high-Greek-disagreement cases.
- Portfolio-level valuation and Greek materiality analysis.
- Production-style rolling median/MAD, EWMA, and CUSUM monitoring diagnostics.
- Controlled spot, volatility, rate, Heston rho, vol-of-vol, and spread stress repricing.
- Fresh DVE confirmatory evaluation on ordinary locked and OOD samples.
- Independent scientific/code re-review and evidence-integrity verification.
- v1 technical report (historical baseline)
- Final v1.1 acceptance report
- Independent scientific/code re-review
- Post-validation analytics
- Fresh DVE confirmatory study
- Model winner map
- Greek-risk analysis
- Calibration diagnostics
- Publication provenance and data exclusions
- v1.1 release notes
The historical v1 technical report remains preserved for provenance. The accepted v1.1 extension is documented by the v1.1 acceptance, independent-review, confirmatory, and post-validation reports above.
Python 3.12 is the qualified hosted-CI baseline.
python -m venv .venv
python -m pip install --upgrade pip
python -m pip install "torch==2.14.0+cpu" --index-url https://download.pytorch.org/whl/cpu
python -m pip install -e ".[dev,oracle]"
python -m pip check
python -m ruff format --check src tests scripts
python -m ruff check src tests scripts
python -m mypy src
python -m pytest -q -m "not gpu and not numerical"
python -m pytest -q -m numerical
python scripts/verify_publication_integrity.pyGitHub-hosted CI is intentionally CPU-only. It runs source-quality checks, 257 CPU-safe tests, all 16 numerical tests, and read-only verification of the published DVE evidence. The five accepted CUDA/GPU tests, one CUDA-dependent confirmatory integration path, and the full calibration/stress research program remain qualified local-only workflows.
| Path | Purpose |
|---|---|
src/derivguard/ |
Pricing, calibration, challengers, validation, materiality, monitoring, and research workflows |
tests/ |
Unit, integration, numerical, regression, and GPU qualification tests |
configs/ |
Model, calibration, data-quality, and experiment configuration |
experiments/ |
Experiment registries and confirmatory specifications |
governance/ |
Hypothesis registries, controls, and frozen validation definitions |
reports/ |
v1 technical, validation, limitation, and reproducibility reports |
reports/v1_1/ |
Post-validation analytics and accepted v1.1 review package |
results/ |
Publication-safe aggregate and synthetic evidence |
artifacts/ |
Publication-safe freeze, release, and integrity manifests |
docs/ |
Numerical methods, publication provenance, and release notes |
The public Git history is a sanitized publication history assembled from the authoritative research snapshots. Vendor-derived quote data, intermediate market data, empirical checkpoints, raw/contract-level real-data tables, and contract-scenario stress rows are intentionally not distributed.
The following scientific limitations remain part of the accepted result:
- OOD robustness of Full DVE failed materially.
- Calibration bootstrap and profile-width methods are local diagnostic approximations.
- SVI/SSVI stress outputs use the disclosed sticky-strike convention.
- The Heston CPU comparison is a bounded 12-row deterministic spot-scenario check, not a comprehensive CPU comparison across every stress family.
- 0DTE pricing remains unavailable.
- Synchronized temporal OOS evaluation remains unavailable.
- Hedge/P&L conclusions remain unavailable.
See Publication Provenance for the exact public/internal boundary.
derivguard-v1.0 and derivguard-v1.1 are immutable sanitized publication snapshots. The derivguard-v1.1 tag is the accepted scientific release. The main branch may receive documentation, metadata, and repository-presentation improvements that do not modify accepted scientific evidence.
The repository is publicly viewable for research and portfolio evaluation, but it is not released under an open-source license. See LICENSE.md for the code-use terms and Publication Provenance for data-rights boundaries.
If you reference the software, use the repository's CITATION.cff. GitHub can render it through the repository's Cite this repository control.