Release date: 2026-07-16
This release closes the current experimental cycle and reframes the repository as an evidence-controlled portfolio research platform. It preserves negative and inconclusive results instead of promoting development point estimates as confirmed investment performance.
- Corrected return accounting, fill timing, and spread/slippage double counting.
- Added incremental market-data caching and optional Reddit/X connectors.
- Replaced the mismatched one-day target with a five-session sector-neutral residual-return rank.
- Expanded from a hindsight-selected ten-stock universe to effective-dated S&P 500 membership.
- Added nested purged model selection, SHAP feature selection, persistent OOS forecasts, and permanent shuffled controls.
- Reduced development turnover using a frozen 40-name hold zone and 1% trade threshold.
- Ran the shortened public confirmation, strict weekly NLP, exact-time FinBERT event, point-in-time NMF topic, and pair-trading experiments.
- Added resumable pair formation and placebo checkpoints.
- Revised the architecture so every module's portfolio authority follows its evidence gate.
| Module | Release disposition |
|---|---|
| Data, feature store, execution, experiment governance | Retained |
| Frozen numeric 40-name portfolio | Reproducible research baseline; not confirmed alpha |
| Slow sentiment and topics | Disabled by default after shuffle robustness failure |
| Tactical FinBERT events | Disabled after timing/sign and return gates failed |
| Pair trading | Disabled after confidence-bound and random-pair gates failed |
| Structured LLM events | Proposal only; no allocation authority |
archive-experimentspreserves the previously publishedmaintipe15e51dbefore this consolidation.maincontains the final proposal, pair experiment, revised project status, and release notes.
Both branches should be pushed so the historical project remains directly accessible while the repository default branch presents the consolidated research platform.
- Complete test suite: 128 passed.
- New pair module and tests: Ruff clean.
- Python source compilation: passed.
- Git whitespace validation: passed.
- Pair result reproduced from cached formations and placebo checkpoints.
Generated datasets, model artifacts, permutation paths, and reports remain ignored. Their frozen contracts and conclusions are represented by tracked manifests, implementation code, tests, and dated Markdown logs.