The study store did not answer; this is the study as last read.
FAILThe Turtle 20-day breakout with a 2N stop, net of costs, on listed and delisted US stocks
Does the classic Turtle rule (buy a 20-day high or short a 20-day low, exit on a 10-day break or a 2N stop, add up to four units) make money after costs on US stocks from 2016, delisted names included?
No. Neither side passed. Long trades lost 0.22 R each after costs (56,073 trades, 2016 to mid-2022) and short trades lost 0.56 R (48,762 trades). Every pre-registered check failed: the mean, its month-clustered lower bound, the year-by-year check, double costs, dropping the top 1% of trades, and the deflated Sharpe. Before costs the long side made about +0.04 R a trade and the short side lost 0.32 R. Costs of about a quarter of an R a trade (about 2.6 units, each paying slippage and fees both ways) turn the long side negative. The rule won in a few trending years (longs in 2017 and 2021, shorts in 2020) and lost more in the rest. Because nothing passed discovery, the validation and holdout periods were never judged.
Backtested by QS Research
Pre-registration
Written and approved before any data for the test was read. Turtle System 1 as the trend screen runs it: enter on a 20-day high (long) or low (short), exit on a 10-day low or high or a 2N stop, add a unit every 0.5N up to four, with N the 20-day Wilder ATR and gaps filled at the open. It runs on a survivorship-free daily universe of listed and delisted US common stocks with adjusted prices. A name may enter only with $20M 50-day dollar volume and a $1B market cap on the prior day. Costs are per-name liquidity-tier fees both ways, 0.05N slippage per fill and borrow on shorts. Long and short are two trials, judged at a family of 3 (the two trials plus an earlier, different Turtle test) on six checks, with a 60/20/20 split by exit date. The holdout opens only for a trial that passes validation.
Registered 2026-11-05A holdout sample is sealed: its results stay unread until the study is final.
Trials · 2 trials · 0 passed · judged at N = 3 trials; 1868 trials run in total
| Trial | Verdict | |
|---|---|---|
| Long: buy the 20-day high | RESULT | counted |
| Short: sell the 20-day low | RESULT | counted |
| Trades | RESULT | not counted |
| Mean net R | RESULT | not counted |
| Month-clustered lower bound (one-sided α = 0.025) | RESULT | not counted |
| Per-year: positive years / losing years (of 7) | RESULT | not counted |
| Stress mean (costs × 2) | RESULT | not counted |
| Mean without the top 1% of trades | RESULT | not counted |
| Deflated Sharpe at N = 3 | RESULT | not counted |
| Verdict | RESULT | not counted |
Limitations
- An earlier, different Turtle test (long only, entries at the close, survivor names) read 2016 to May 2022; this test's discovery span overlaps it, and that test is counted in the family.
- Prices are split- and dividend-adjusted, so channel levels differ slightly from a raw chart, and shorts pay dividends through the prices.
- One rule: no 'skip after a winner' filter and no portfolio limits; every eligible name trades on its own.
- Costs are a model (tier fees, slippage, borrow), with no market impact beyond the tier. With the screen's lighter costs, both sides still lose.
- Names under $20M daily dollar volume or a $1B market cap are not tested.
- Not a tested result; a candidate hypothesis: a positive figure for the most liquid names (long, +0.10 R after costs) is a slice found after the test failed. It goes to a separate pre-registered test, not to this verdict.
The numbers: every trial's checks, metrics, curve and yearly results, and the provenance
3 published versions · current b8b0d5e2c0c181a8, published 2026-11-11