Status: Research complete / v1 research cycle concluded Frozen research endpoint: 2026-05-01
Market Strats Lab is a reproducible research system for testing whether transparent market rules and point-in-time machine-learning signals can improve long-term investing outcomes without overstating what historical evidence proves.
Can systematic ETF risk controls or a point-in-time individual-stock ranking process outperform SPY Buy & Hold after costs while remaining robust, operationally credible, and investable?
Not on all required dimensions.
- SPY Buy & Hold remained the raw-return and terminal-wealth winner.
- A transparent defensive SPY overlay retained most long-run CAGR while materially reducing historical drawdown and improving the return/drawdown trade-off.
- A controlled stock-ranking pilot produced encouraging out-of-sample rank-correlation evidence, but its hand-selected survivor universe prevents credible generalisation.
- The canonical stock experiment could not be qualified from the permitted free sources because historical identity, membership, delisting, terminal-value, and complete price evidence remained unresolved. That is a data-foundation result, not a model failure.
The project therefore ends with a useful risk-control result, a promising but noncanonical stock signal result, and a clear demonstration that point-in-time data integrity became the binding constraint on credible free-data equity research.
No tested strategy dominated SPY Buy & Hold simultaneously on raw wealth, risk, and practical liveability. The ETF figures below are frozen historical research results for 2006-04-28 through 2026-05-01, not forecasts.
| Strategy | End value | CAGR | Calmar | Maximum drawdown |
|---|---|---|---|---|
| SPY Buy & Hold | $79,306.63 | 10.90% | 0.197 | -55.19% |
| SPY 3D Overlay + deep-drawdown guard + loose relief | $71,779.16 | 10.35% | 0.429 | -24.12% |
Conclusion: SPY Buy & Hold remained the raw-return and terminal-wealth winner, while the final ETF overlay materially reduced historical drawdown and improved the return/drawdown trade-off.
The controlled 16-stock pilot was explicitly:
NONCANONICAL
SURVIVORSHIP-BIASED
RESEARCH-ONLY
NOT INVESTABLE PERFORMANCE
Its strictly out-of-sample Ridge rankings produced:
| Diagnostic | Frozen result |
|---|---|
| Mean Spearman IC | 0.1225 |
| Median Spearman IC | 0.1485 |
| Positive-IC date fraction | 62.9% |
| Top-k 20-day average excess return | 1.2166% |
| Top-minus-bottom rank spread | 2.8345% |
| Within-date permutation IC p-value | 0.0099 |
The moving-block-bootstrap IC interval was 0.0069 to 0.2420, but the rank-spread interval was
-0.0031 to 0.0576 and crossed slightly below zero. The result supports scientific interest, not
an investable-performance claim.
The final universe contract prohibited current-survivor filtering, ticker-only identity, same-close execution, missing-price imputation, and zero-valued missing delisting returns. The final bounded real-source rerun preserved those rules and produced:
blocked_identity_reconciliation_failure
SEC acquisition succeeded with HTTP 200 and returned 10,398 current mapping rows. Even so, only 42 of 1,126 historical identities resolved under the strict name/ticker evidence rule. The frozen run also retained 36 membership conflicts, 348 price-coverage failures, 648 unresolved delisting outcomes, and zero qualified monthly decisions. The final engineering rerun completed and persisted all 12 Parquet evidence tables plus the source manifest, licence audit, and JSON/Markdown summaries. The persisted summary records the verdict above and explicitly leaves canonical model training unauthorized. The bounded zero-cost research path is closed.
Increasing model complexity was not the main determinant of credible evidence. Point-in-time availability, security identity, costs, benchmark discipline, purging and embargoes, robustness tests, and honest failure states mattered more.
point-in-time sources
-> immutable local snapshots and availability evidence
-> source qualification and coverage audits
-> frozen features, targets, and model comparisons
-> research-only signal exports
-> portfolio, cost, and robustness diagnostics
-> prospective shadow records with delayed outcomes
-> explicit stop gates before broker or real-money use
market_strats.intelligence owns point-in-time source and signal evidence. The broader
market_strats package owns strategy evaluation, portfolio diagnostics, costs, robustness, and
benchmarks. Neither layer authorizes live execution.
- ETF buy-and-hold, trend, momentum, allocation, and defensive-overlay research;
- transaction-cost, spread, impact, turnover, tax, and behavioural diagnostics;
- walk-forward, rolling-window, bootstrap, permutation, and ablation controls;
- a point-in-time market, macro-vintage, FOMC, BLS, and SEC evidence layer;
- interpretable Ridge ranking and a fixed tree comparator;
- purged and embargoed out-of-sample stock evaluation;
- immutable source snapshots, hashes, availability records, and fail-closed contracts;
- manual paper and prospective shadow infrastructure with no broker authority;
- a strict point-in-time S&P 500 universe qualification system.
- SPY Buy & Hold remained the primary raw-return benchmark.
- Historical endpoint and rules were frozen before closeout.
- Transaction costs and execution timing were explicit.
- Same-close execution and lookahead data were prohibited.
- Stock labels used chronological purging and embargoes.
- Current-survivor filtering and ticker-only identity were prohibited.
- Missing prices and delisting values were never silently set to zero.
- Artifact-backed tests remained separate from portable CI.
- Raw provider data, credentials, generated reports, and restricted evidence remained local.
| Research track | Final v1 status |
|---|---|
| ETF raw-return benchmark | SPY Buy & Hold remained the winner |
| ETF risk-adjusted candidate | Useful historical return/drawdown trade-off; not universally dominant |
| Individual-equity model | Encouraging noncanonical pilot; no broad-universe claim |
| Canonical point-in-time universe | Blocked on unresolved evidence |
| Individual-equity prospective shadow | Proposal generated; first entered session remained pending |
| Broker/live/real money | Never authorized |
| Additional strategy development | Intentionally stopped for v1 |
Future work would be a new research cycle, not continuation by default. A canonical equity v2 would require a new preregistered objective and appropriately licensed historical constituent, security-master, delisting, and terminal-value evidence if the free-data boundary remains.
Python 3.11 is supported.
py -3.11 -m venv .venv
.\.venv\Scripts\python -m pip install -e ".[dev]"
.\.venv\Scripts\python -m ruff check --select E,F,I,UP `
src/market_strats/intelligence src/market_strats/universe `
tests/test_mi*.py tests/universe
.\.venv\Scripts\python -m pytest -q -m "not artifact"Rebuild the public-safe aggregate figures with:
.\.venv\Scripts\python scripts\build_public_closeout_figures.pyThe plotting script uses only frozen aggregate values already published in
docs/research_history.md; it does not read or redistribute provider time series.
src/market_strats/intelligence/ point-in-time source and signal research
src/market_strats/universe/ canonical-universe qualification
src/market_strats/analysis/ evaluation and diagnostics
src/market_strats/strategies/ transparent rules and benchmarks
src/market_strats/global_multi_asset/ multi-asset contracts and tournaments
configs/ frozen research contracts
tests/ portable and artifact-marked validation
docs/ conclusions, architecture, methods, and detailed history
- Final research conclusion
- Project closeout status
- Detailed research history
- Architecture
- Model objective
- Point-in-time universe plan
- Free-source universe result
- Free-data limit
- Reproducibility
- v1.0.0 release notes
Source code is MIT licensed. Third-party data is governed by its provider terms and is not covered by the code license. Raw SEC filings, provider responses, large generated panels, release archives, credentials, personal paths, and redistribution-restricted evidence remain excluded from Git.
Research and education only. Nothing in this repository is financial advice, an investment recommendation, a live signal, or authorization to place an order.