The Edge That Wasn't
All 81 experiments › #26

#26 — H4 Donchian breakout (backtest)

2026-05-13 backtest + deploy (Sessions 053–054); live until stopped 2026-05-26
Verdict: positive · ✅ backtest — refuted live
H4 Donchian(10) breakouts with a 1×ATR trail backtested at 16–25 p/d/pair with 84% WR, then went 0-for-3 live and were stopped.

Hypothesis — what it tests

Do H4 Donchian-channel breakouts with an ATR trailing stop capture multi-day trends profitably after spread?

Description

A classic Donchian-channel breakout on 4-hour bars: buy when the close breaks the 10-bar highest high, sell on the 10-bar lowest low, initial stop 2×ATR, trail 1×ATR(14) from the peak. The Session 053 backtest (part of a two-strategy H4 trend study alongside SMA crosses) passed bootstrap gates on all four JPY-heavy pairs at 16.4–24.8 p/d OOS with Calmar 70–145, and a follow-up duration analysis of 3,434 backtest trades showed the seductive asymmetry: 84.2% win rate, +41.2p average winner vs −12.1p average loser, losses dying within 24h while winners ran (6+ bars = 99% WR). It was deployed on account 010 at 10 units/pair. Live it was refuted outright — 3 trades, −61.6 pips, 0% win rate — and stopped 2026-05-26 (audit Entry 53). The book's own editorial audit flags the '84% WR' as backtest-only, mechanically inflated by conditioning on trade duration; a later H1 sweep (Entry 29, 0/2400 configs positive) confirmed no Donchian variant generalizes.

Key result

Backtest OOS: GBP_JPY 24.8, USD_JPY 24.0, EUR_JPY 20.2, GBP_USD 16.4 p/d (85.4 total), Calmar 70–145; duration study: 3,434 trades, 84.2% WR, +41.2p avg win / −12.1p avg loss, 99.8% of losses exit ≤24h; live: 3 trades, −61.6p, 0% WR, stopped 2026-05-26

Indicators

Donchian channelWilder ATR(14)ATR-ratio (atrr_20)

Algorithms

grid searchbootstrap validation (P5>0, P(+)>0.95, 2000 samples)ATR trailing stoptrade-duration analysisLightGBM feature screening (H4 8A-D pipeline)

Code

research/experiments/technical/backtest_h4_trend.pyresearch/experiments/technical/backtest_h4_atrr_filter.pyresearch/experiments/technical/h4_feature_screening.pyresearch/experiments/h4_donchian_duration.pyservices/strategy_h4_donchian/main.py

Interactive version (search, filters, figures): the experiment explorer.

← #25 EUR/USD microstructure / tick-pace#27 Zone Recovery trail-lock redesign →
This page is one experiment from the audited record behind the book The Edge That Wasn't — get the book · about the project.