BrainBytesandClaude Opus 4.6
9a8fcfcaac
The complete trading and quantitative finance system for Claude Code. - 18 specialized agents (quant-researcher, algo-strategist, risk-manager, derivatives-analyst, etc.) - 82 domain skills across 10 categories (quant methods, strategies, risk, derivatives, portfolio, execution, data, crypto, macro, backtesting) - 20 slash commands for trading workflows - 5 always-active rules (risk limits, data integrity, model governance, regulatory compliance, ethical trading) - 4 hook scripts (risk check, backtest validation, data quality, session start) - 3 context modes (research, live-trading, backtesting) - 3 example configs (quant fund, crypto trader, retail algo) - 8 MCP server configurations - The Complete Guide (setup + advanced patterns) - Viral-ready README, CONTRIBUTING, CI, PR template Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
3.0 KiB
3.0 KiB
description
| description |
|---|
| Model governance rules for validation, documentation, and review of quantitative models |
Model Governance
All quantitative models used for trading, risk management, or portfolio construction must follow these governance standards.
Model Validation Requirements
- Every model must have a documented validation report before deployment.
- Validation must be performed by an agent or reviewer independent of the model developer.
- Validation must include: theoretical soundness review, implementation verification, and performance benchmarking.
- Models must be re-validated annually or when market regime changes are detected.
Backtesting Standards
- Walk-forward analysis with at minimum 3 non-overlapping out-of-sample periods is required.
- In-sample and out-of-sample performance must be reported separately. Never report blended results.
- Transaction costs, slippage, and market impact must be included in all backtest results.
- Backtest results must include confidence intervals, not just point estimates.
- The deflated Sharpe ratio (accounting for number of strategies tested) should be reported.
Overfitting Checks
- Monitor the ratio of in-sample to out-of-sample Sharpe. Decay greater than 40% suggests overfitting.
- Limit the number of free parameters relative to the number of independent observations.
- Apply regularization (L1, L2) to any model with more than 5 parameters.
- Use combinatorial purged cross-validation for strategies with overlapping holding periods.
- Document every strategy variant tested. The more variants tested, the higher the bar for significance.
Parameter Stability
- Key model parameters must be stable across rolling estimation windows.
- A parameter that changes sign or magnitude by more than 50% across windows is unreliable.
- Perform sensitivity analysis: model performance should degrade gracefully with parameter perturbation.
- Avoid cliff effects where small parameter changes cause large performance shifts.
Documentation Requirements
Every model must have documentation covering:
- Purpose — What decision does this model support?
- Methodology — Mathematical specification, assumptions, and limitations.
- Data — Input data sources, cleaning procedures, and quality checks.
- Validation — Out-of-sample results, robustness checks, and known weaknesses.
- Parameters — All parameters, how they were estimated, and their stability.
- Monitoring — What metrics are tracked to detect model degradation.
- Contingency — What happens if the model fails or produces unreasonable outputs.
Model Review Process
- New models require sign-off from the risk-manager agent before deployment.
- Material model changes (methodology, parameters, data sources) require re-validation.
- Monthly model performance monitoring reports must be generated.
- Models that underperform their validation benchmarks for 3 consecutive months must be reviewed.
- A model inventory must be maintained listing all active models, their owners, and last validation date.