Cross-pattern survey: classical chart patterns

Thesis (mostly rejected): pennants, flags, head-and-shoulders, cup-and-handle, double bottoms/tops — the patterns every TA book opens with — produce a meaningful directional edge on a clean event-study run with same-name matched controls.

Code: research/flag_pennant.py, research/head_shoulders.py, research/head_shoulders_refine.py, research/cup_handle.py, research/double_bottom_top.py Reports: corresponding *_report.json files

Two patterns show real, if small, edge: bull_flag weakly positive, bear_flag positively signed (contrarian — a textbook bearish pattern is actually a long signal). The bear-flag result lines up with the much stronger contrarian finding in MA-cross / drawdown-bounce (see below): a stock that has been selling enough to form a bearish chart pattern is oversold, not on its way down further. Bull pennant and bear pennant samples are too small to read.

H&S top — sub-cohort with real edge: when filtered to OOS-2020+ with high head-prominence and tight shoulder symmetry, edge widens to roughly -2 % to -2.5 % over 10-20 days (n ≈ 362). The directionally correct signal is real but the magnitude is below what we'd productize alone.

Verdict on the family: Nearly all classical chart patterns are noise after frictions. The only useful artefacts: - Bear flag's contrarian edge → confirms the drawdown-bounce thesis (see Working signal section). - Bull flag's weak positive edge → live with the existing coil-breakout cohort which has the same shape and bigger n.