research: expand executable evaluation contracts across panels

[magnaquant]

Summary

• freezes a prospective multi-panel evaluation protocol before provider retrieval
• evaluates four fixed strategy families across U.S. sector and country-equity ETF panels
• adds a repeated-seed walk-forward GBRT family, causal event-driven accounting, and vectorized sensitivity
• estimates one-switch effects for execution timing, residual cash, transaction costs, matched comparison, and engine semantics
• publishes aggregate results, five figure pairs, exact input/protocol/source hashes, and public and anonymous paper builds
• updates dependency locks for jupyter-server 2.20.0 to address CVE-2026-44727

Research design

The protocol was frozen in commit 4018f4063f46889f41d6981db5a71079e1dbd713 before the two provider requests. It covers:

• 2 complete-row real-data panels
• 4 fixed strategy families
• 8 family-panel cells
• 5 one-switch diagnostics
• 5,000 circular-block bootstrap draws at 5, 21, and 63 sessions
• 5 frozen learned-model seeds
• event-driven pseudo-share accounting as primary and daily re-pegging as sensitivity

This is a repository-frozen historical evaluation, not an externally registered or temporally out-of-sample study.

Main empirical findings

• removing 13 bps modeled cost raises annualized arithmetic return in all 8 cells by 0.57 to 1.81 percentage points
• same-close assignment is not uniformly optimistic: its 21-session pointwise interval is below zero in 3 cells and overlaps zero in 5
• zero residual-cash return is below zero in 5 cells and overlaps zero in 3
• only 1 of 8 strategy-minus-costed-comparator intervals is below zero; the other 7 overlap zero
• same-close and zero-cost diagnostics each change 4 within-panel family ranks; comparator substitution changes 5
• event-driven versus vectorized conventions change no family ranks but reach 93.21 basis points in maximum one-day return difference and 3.89% in maximum absolute terminal-wealth gap
• the original negative worked case remains unchanged at 1.40% net CAGR and -0.15 conventional cash-excess Sharpe

Intervals are pointwise, conditional on the selected panels and frozen models, and are not multiplicity-adjusted. These results do not establish profitable alpha or future performance.

Reproducibility and data policy

Raw provider matrices remain untracked and are identified only by SHA-256. Provider publication permission has not been independently verified. Open schemas, canonical tapes, conformance fixtures, aggregate tables, figures, manifests, and software reproduce the evaluation semantics; exact empirical reproduction requires owner-supplied matrices matching the recorded hashes.

Clean source commit: e0443b8f77cd23aee8f1fa64a2bc237e47626c47

Public PDF SHA-256: 9bf806b296b6bade747de580142545a9825e2d3107c6a55531ea1456932ad2a5

Anonymous PDF SHA-256: cbcee2583ea2da700b259c108d9991061033a17dd97c11d0aab676396eabef7b

Both release wrappers were rerun from the committed artifact release and left the working tree byte-clean.

Verification

• ruff check .
• pytest tests/ -q --tb=short --cov=quantcortex --cov-report=term --cov-fail-under=60: 455 passed, 68.44%
• poetry check --lock and full dependency export drift check
• sdist and wheel build
• 17/17 broker adapter behavioral checks
• all 5 research notebooks execute offline; 4/4 expected figures render
• all 23 public PDF pages inspected; anonymous build inspected and identifier-scanned
• public and anonymous Tectonic builds reproduce byte for byte

Known boundaries

• no live broker authentication, account-permission, funded-order, or venue test
• no claim that provider terms authorize redistribution; raw matrices are not committed
• panels are related U.S.-listed ETFs and the interval is historical, not a temporal holdout
• no third-party engine comparison or independent external preregistration
• no alpha, investment-advice, or deployment-readiness claim