Evidence correction, July 30, 2026: performance estimates built from the pre-correction repaired price panel are invalid, and the intermediate corrected-price replay was later superseded by the common-fold, framework-clean rerun. The design record remains useful, but superseded quantitative claims must not be cited as current evidence. See the framework-clean comparison.
Status: built and verified; not validated live alpha
This final-reference package remains the conservative fail-closed build from the earlier frozen v4 contract; its build metrics below are retained as historical evidence for that frozen design. The later framework-clean comparison is the current model evidence. It identifies Ridge as the strongest model, with 8.22% net CAGR, 0.443 net Sharpe, and -34.29% maximum drawdown from 2020 through 2023 H1, versus 10.73%, 0.583, and -31.62% for full-investment SPY. Ridge's zero-cost sensitivity is 14.99% CAGR and 0.693 Sharpe. It is a candidate to test prospectively, not a passed alpha claim.
The final version turns the project's separate experiments into one auditable reference build. It does not retune a model, activate a failed module, or open the 2026 lockbox. Instead, it checks that persisted evidence still matches the frozen v4 contract, calculates Fama-French attribution, hashes every required input, and writes one coherent result package.
| Component | Final setting |
|---|---|
| Universe | Point-in-time liquid S&P 500 candidates |
| Target | Five-session sector-neutral residual-return rank |
| Model | Ridge, alpha 10, 40 frozen features |
| Portfolio | 40 equal-weight names, 70% invested |
| Turnover control | Rank-200 exit, 26-week minimum hold, 1% trade threshold |
| Sector cap | 25% |
| Execution | Next open, 8 bps one way plus $10 per order |
| FF3 | Attribution enabled; alternative target authority zero |
| Slow sentiment and topics | Authority zero |
| Tactical news | Authority zero |
| Pair trading | Authority zero |
| LSTM and Transformer | Excluded |
| Transparent smart-universe v7 | Inactive after failing 4 of 5 development gates |
| Lockbox | Closed |
| Period and portfolio | CAGR | Sharpe | Max drawdown |
|---|---|---|---|
| Development reference, 182 weeks | 12.19% | 0.71 | -27.40% |
| Shortened observed confirmation, 64 weeks | 11.34% | 1.11 | -5.73% |
| Confirmation 70% SPY / 30% cash | 11.05% | 1.14 | -6.26% |
The portfolio's confirmation CAGR was 0.29 percentage points above the exposure-matched benchmark, but its Sharpe was slightly lower. Earlier confirmation controls also failed the shuffle and confidence-bound gates, so this is an observed portfolio result rather than a market-beating claim.
Fama-French attribution reinforces that conclusion:
- development annualized alpha: +4.96%, p=0.320;
- shortened-confirmation annualized alpha: -1.93%, p=0.672;
- shortened-confirmation benchmark alpha: -1.67%, p=0.584.
None is statistically distinguishable from zero.
build-final-version refuses to run when:
- any failed NLP, pair, or FF3-target module has nonzero authority;
- LSTM or Transformer is enabled;
- the model or portfolio manifest differs from the frozen artifacts;
- a source artifact says the lockbox was opened;
- weekly returns reach the sealed lockbox period;
- required inputs or the failed FF3-target decision are missing.
python -m src.cli download-fama-french
python -m src.cli build-final-versionThe default output is reports/final_reference_v1/:
resolved_manifest.json: complete executed contract;metrics.json: performance, factor attribution, and module authority;weekly_returns.csv: aligned development and observed-confirmation returns;ARTIFACTS.sha256: hashes for the manifest and every evidence input;FINAL_BUILD.md: concise human-readable result.
Generated reports remain ignored by Git. The manifest, implementation, tests, and this evidence summary are tracked.
This command packages and validates already frozen artifacts. It is not a raw-data-to-live-order pipeline and does not claim deployable alpha. A future production version would still require official point-in-time sectors, stronger delisted-security coverage, participation-aware costs, an untouched confirmation period, monitoring, and an order-approval layer.
University WRDS/CRSP access is not available for this project phase. The public-data pipeline may continue to support development and sensitivity analysis, but its missing-security effect is only partially identified. A future CRSP or equivalent licensed-data integration is therefore a data quality upgrade and a prerequisite for any fully survivorship-corrected performance claim, not a current runtime dependency.
The later transparent-universe v7 experiment does not alter this reference. Its weekly top-120 filter reduced CAGR and increased turnover; any successor must use a slower universe target and monthly or quarterly membership bands on a new development boundary.