Does a stock's own Turtle record predict its next Turtle trade?
If a stock's Turtle long trades worked over the last year, does its next Turtle long trade make money after costs?
Both trials fail. The top third by each name's own past-year record lost 0.174 R a trade after costs (R is the trade's risk unit) on 4,765 trades, with a gross edge of only 0.083 R; its month-clustered lower bound was -0.451 R and only 1 of 6 years was positive, so it failed all six checks. The top-minus-bottom spread made 0.052 R a month over 66 months, with 5 of 6 years positive, but its lower bound was -0.147 R and its deflated Sharpe was 0.015 against the 0.80 bar, so the score does not reliably sort trades. The middle third lost less than the top third in discovery. The top third's mean is flattered by a few winners: the median top trade lost 1.296 R and only 23.4% of top trades made money. Three in four discovery entries (41,896 of 56,073) could not be scored, and the median trailing score across the scored entries was -0.731 R: the top third was mostly names that lost less, not names that won. The holdout stays sealed. Judged at the pre-registered family of 5, not against every trial QS Research has run. The live past-record sort uses a noisier version of the idea and gets no support from this result.
Backtested by QS Research
Pre-registration
Approved on 2026-11-14, before the counted run. The rule is the earlier Turtle long rule: N is the 20-day average true range and R is twice N, the risk unit; buy a 20-day high, exit on a 10-day low or a 2N stop, add up to four units at half-N, on a 10,695-name survivorship-free US stock universe from 2016 to October 2026, with the earlier study's $20M daily dollar-volume and $1B market-cap entry filters. Each entry is scored by the mean net R of the same name's long Turtle trades that closed in the 252 sessions before it, needing at least 8; entries with less history or fewer trades are unscored. On each entry date the scored entries are split into thirds by the scores of the previous 21 sessions: top above the 2/3 cut, bottom below the 1/3 cut, middle otherwise; a session with fewer than 3 scores is unranked. The two trials are the top third and the monthly top-minus-bottom difference. The split is 60/20/20 by exit date; the holdout from 2024-08-29 was never read because no trial passed discovery. The family is 5 trials: the two new ones plus the earlier Turtle long, Turtle short and Wave 2 Turtle. The bar is the approved policy: at least 30 observations (trades, or months for the spread), then six checks — mean net R above zero, a one-sided 0.025 month-clustered lower bound above zero, at least half the years positive with at most two losing, double-cost stress above zero, the top 1% of observations removed above zero, and a deflated Sharpe of at least 0.80. The pre-registration flagged that the round-trip cost was about 0.26 R against a +0.044 R unconditional gross edge; the study was run anyway to decide the live sort's label. The implementation choices were fixed in code before the run; there were no amendments after approval.
Registered 2026-11-14A holdout sample is sealed: its results stay unread until the study is final.
Trials · 2 trials · 0 passed · judged at N = 5 trials; 2965 trials run in total
| Trial | Verdict | |
|---|---|---|
| Top third by the name's own past-year Turtle record | FAIL | counted |
| Top third minus bottom third, by month | FAIL | counted |
Limitations
- The live past-record sort is not the tested rule: it uses a 6-month state window and a 1-year backtest that mixes long and short trades. This study tested the cleaner long-only version; the result gives the noisier live version no support.
- Long side only. The persistence of short-side records was not tested.
- The score needs 8 closed long trades in the prior 252 sessions. 41,896 of 56,073 discovery entries (three in four) were unscored, including every 2016 entry because the window had not yet built 252 sessions; the score covers names that break out often, a minority.
- The top third is mostly names that lost less: the median trailing score across the scored entries was -0.731 R and the 2/3 quantile about -0.205 R.
- The top third's mean is flattered by a few winners: the median top trade lost 1.296 R and only 23.4% of top trades made money.
- The detailed table's per-trade mean, lower bound, stress mean and median are in R, the trade's risk unit (twice the 20-day average true range), not fractions of capital. The shared report reader labels the unit fraction and the study page shows those values as percentages: read -17.4% as -0.174 R, the median -129.6% as -1.296 R, and the spread's +5.2% as +0.052 R. Use the R figures in the summary.
- Costs are a model: liquidity-tier fees by dollar volume plus 0.05N slippage per fill, with no market impact. The ranking is computed on net R under this model; it was not re-run under lighter app costs, so the ordering could change. The top third's gross edge was 0.083 R against an average round-trip cost of about 0.257 R, so the fail is not close.
- The universe is the earlier Turtle study's: names need $20M daily dollar volume and $1B market cap on the prior day; smaller names are not tested. 184 names with bars do not match the market-cap panel and are never entered.
- One regime: discovery is 2016 to mid-2022, mostly rising. The descriptive validation span (2022-2024) is shown but not judged, and the holdout from 2024-08-29 was never read. Showing the validation span after no trial passed discovery was not pre-registered; it uses up that split for any future version of this rule.
- The deflated Sharpe is judged at the pre-registered family of 5, which includes the earlier Turtle long and short and the earlier Wave 2 Turtle; the effective family count from the trial record was 3. It is not deflated against every trial QS Research has run.
- Delisted names exit at their last close; prices are split- and dividend-adjusted, so channel levels differ slightly from a raw chart.
- No per-trade list is published; the result is aggregate.
- Six discovery entries had too few scores in their 21-session windows to rank; they are in the result but not in the write-up's table.
The numbers: every trial's checks, metrics, curve and yearly results, and the provenance
1 published version · current a9ddd5525cb444dd, published 2026-11-21