[laurentvv]

Résumé

Phase 3 du plan, empilée sur #111 (elle-même sur #110). Le banc est construit et son verdict, calculé par le code, est FAIL : la porte du plan (§2.1) n'est pas franchie, donc rien n'est construit côté exécution T212.

Résultat (test hors échantillon 2021-01-04 → 2026-10-12, net de frais 25 pb/sens et d'impôt 30 %)

	CAGR	Sharpe	Drawdown max	Espérance
Pullback (variante retenue au calibrage 2015-2020)	−22,6 %	−1,93	−78,9 %	−0,55 R (680 trades)
Entrées au hasard (médiane de 40)	−15,0 %	−1,25	−64,6 %	−0,45 R
B&H Nasdaq-100 EUR	+14,6 %	0,74	−31,3 %
MA200 Nasdaq-100 EUR	+9,7 %	0,65	−31,1 %

• Les 6 variantes pré-enregistrées perdent toutes, au calibrage comme au test.
• La stratégie est battue par 95-98 % des tirages au hasard : consolidation, support et AVWAP n'ajoutent rien dans cette traduction journalière.
• Pas un artefact des hypothèses : avec l'ordre intrajournalier le plus favorable, elle reste négative (−7,8 %/an, −0,16 R). Le biais du survivant joue en sa faveur.
• Limite : l'intraday de la source (bougies 1 à 15 min) n'est pas testé. La méthode d'origine est donc non testée, pas réfutée.

Contenu

• src/pullback/market.py : pré-filtre vectorisé, identique à rules.detect_consolidation (0 écart sur 12 089 contrôles réels) et au screener.
• src/pullback/portfolio.py : moteur événementiel OHLC (gaps payés, pire cas intrajournalier, change, frais, impôt).
• src/pullback/bench.py : variantes, walk-forward, contrôles, bootstrap, Sharpe déflaté, verdict, rapport.
• docs/BACKTEST_PULLBACK_2026-10-12.md (généré) + §9 du plan.
• 15 tests moteur ; suite complète 466 passed, 3 skipped.

Décision à prendre

Phases 4 à 7 suspendues. Options : tester l'intraday (données payantes), arrêter la méthode, ou revenir à une base passive.

🤖 Generated with Claude Code

[kilo-code-bot]

Phase 3 pullback gate review. Inline findings below; no logic/numerical bugs found in the engine, prefilter, bench, or gate.

[kilo-code-bot]

WARNING: Unused import replace from dataclasses

replace is imported at src/pullback/bench.py:15 but is never referenced in the new file (the tests build frozen Variant instances via the keyword constructor, e.g. bench.Variant("b", max_dist=0.04), rather than dataclasses.replace). Under the repo's Ruff configuration in pyproject.toml, whose default rule set includes F401, ruff check flags this — remove it before merge.

Suggested change

	from dataclasses import dataclass, replace
	from dataclasses import dataclass

---

Reply with @kilocode-bot fix it to have Kilo Code address this issue.

[kilo-code-bot]

SUGGESTION: Dead assignment to Bl.liquid

Bl = Bench(mkt, refs, sig_large) is only driven through Bl.evaluate(...), which calls build_orders/run_portfolio — neither reads Market.liquid. Market.liquid is consumed only by Bench.ranks() (via momentum_masks); Bl.ranks() is never called in run_bench, so this assignment has no effect on the verdict. The large-cap filter is already applied at signal generation (market.candidate_setups(mkt, mkt.liquid(min_dollar_volume=LARGE_CAP_DOLLAR_VOLUME), ...)), so the line is misleading. Remove it, or keep it with a note if Bl.ranks() is intended later (note: removing is safe only while ranks() stays uncalled on Bl).

---

Reply with @kilocode-bot fix it to have Kilo Code address this issue.

[kilo-code-bot]

Code Review Summary

Status: 2 Issues Found | Recommendation: Address minor non-behavioral issues before merge

Overview

Severity	Count
CRITICAL	0
WARNING	1
SUGGESTION	1

The pullback bench is a careful, well-tested (466 passed, 3 skipped) implementation of the daily pullback method and correctly computes the FAIL verdict from the plan gate (§2.1). I read the three new modules in full plus their dependencies (src/pullback/rules.py, store.py, universe.py; src/backtest/engine.py, metrics.py, strategies.py) and traced the timing, position-sizing units, tax/fee accounting, R-denominator consistency, gate/p-value logic, and reference-metrics keys. No logic or numerical bugs were found in the execution engine, the vectorized consolidation prefilter, the bench/walk-forward logic, or the gate. Two non-impacting cleanliness items are flagged inline.

Issue Details (click to expand)

WARNING

File	Line	Issue
src/pullback/bench.py	15	Unused import replace (from dataclasses) — flagged by Ruff F401 under the project's configured Ruff rules.

SUGGESTION

File	Line	Issue
src/pullback/bench.py	312	Dead assignment Bl.liquid = mkt.liquid(...): Bl.ranks() is never called, so .liquid is never read for Bl; the large-cap universe is already applied via sig_large at signal generation.

Files Reviewed (7 files)

• src/pullback/bench.py — 2 issues (new)
• src/pullback/market.py — 0 issues (new)
• src/pullback/portfolio.py — 0 issues (new)
• tests/test_pullback_portfolio.py — 0 issues (new)
• src/pullback/rules.py — read for context (unchanged)
• src/backtest/engine.py — read for context (unchanged)
• src/backtest/metrics.py — read for context (unchanged)

---

_{Reviewed by free · Input: 312.6K · Output: 61.9K · Cached: 1.2M}