Repository navigation
fix(notebook,#18124): research_long_short_harvest -- real VIX quintiles + measured Hurst + trained ML core replace placeholder - #18164
Conversation
…es + measured Hurst + trained ML core replace placeholder SOTA repair (last of the placeholder set in the #18124 audit). The notebook was QuantBook-only with un-executed analysis cells and a fake backtest placeholder at the end. - Local research path: SPY + GLD + the real ^VIX index via yfinance (tz-normalized), so the whole notebook executes locally. - VIX quintile -> next-21d SPY return table measured on the real index: Q1 +0.82% ... Q5 +2.87% (n~600 per bucket) -- the volatility-risk-premium structure the strategy builds on, measured instead of asserted. - Hurst exponent measured (aggregated variance) on SPY 0.009 / GLD 0.020: daily-scale series sit far below the H > 0.85 short-screen threshold -- informative about how selective that screen is. - The promised RandomForestClassifier (11 VIX/SPY features) trained for real (temporal 70/30 split): AUC 0.386, accuracy 61.2% < 69.8% majority baseline -- honest NO EDGE verdict written in. - The full long-short engine (top-4 market-cap longs, weekly Hurst shorts, 3-stage stops, margin) explicitly routed to QC Cloud, not faked. Executed 9/9 cells, 0 errors, all execution_count set (C.1/C.2). Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
|
Scope = notebooks CHANGED in this PR, not the whole corpus. Explicit |
github-actions
Bot
added
the
variation-tag-prev-absent
Notebook outputs-required (H.4 schema): PASS (every code cell carries an
|
Golden-Set Execution (H.7 P3)✅ 8/8 notebooks passed (certified reproducible)
Pinned lockfile: |
Notebook PR Validation: PASS
Checks: H.1 (no errors), H.3 (execution_count), C.1 (no banned patterns) |
…, 3.11 -> 3.13 to match series env)
…thon3 (3.11) kernel to match base stamp
github-actions
Bot
added
the
large-pr-no-review
|
Cette PR depasse le seuil de couverture review (par defaut 300 additions) et n'a recu aucune review -- ni bot, ni humaine. Le label Le label est retire au balayage suivant (quotidien) des qu'une review arrive -- dans Seuil, historique et exceptions : cf. |
|
[ADJOINT PREFLIGHT] Note de lecture (informatif) : advisory sticky |
2 participants
Grain: MED/notebook-python -- lane myia-po-2026:CoursIA -- prev: #18163
Summary
SOTA repair of
Research-Executor/research_long_short_harvest.ipynb(last ofthe placeholder set in the #18124 audit). Was: QuantBook-only (unrunnable
locally), analysis cells un-executed, fake "BACKTEST RESULTS" placeholder.
tz-normalized, notebook executes end-to-end locally.
Q1 +0.82% / Q2 +0.40% / Q3 +0.90% / Q4 +1.04% / Q5 +2.87% (n~600 per
bucket) -- the volatility-risk-premium structure the strategy builds on,
measured instead of asserted.
daily-scale series sit far below the H > 0.85 short-screen threshold;
the notebook now says how selective that screen really is.
features, temporal 70/30 split): AUC 0.386, accuracy 61.2% < 69.8%
majority baseline -- honest NO EDGE verdict written in, not hidden.
3-stage stops, margin) explicitly routed to QC Cloud.
Validation
execution_countset (C.2)raise NotImplementedError/assert False/1/0(verified)model, honest out-of-sample verdict
See #18124. With this PR the tranche claimed by this lane is complete:
6 of the 10 audited notebooks repaired (fallback-banner trio + placeholder
trio); remaining 4 (non-executed-cells family: defensive 4 done in #18163,
macro_factor_rotation + any stragglers) to re-audit against the current state.
🤖 Generated with Claude Code