Per-market Turtle lengths, walked forward against classic 20/10/20
On US stocks and ETFs, does fitting the three Turtle lengths (entry channel, exit channel and the average-true-range length, N) separately for each market on the five years before each year trade that year better than the classic 20/10/20, and better than one set shared by every market?
Not a counted test. The study was not pre-registered: its record was written on 2026-11-21, after every result, and nothing was fixed in advance. It fits the three Turtle lengths (entry channel, exit channel and N) separately for each US stock or ETF on the five years before each year, then trades that year with the fitted set, and compares it with classic 20/10/20 and with one set shared by every market. On the final run (574 names asked, 531 fitted, 529 with a walk-forward), the Sharpe-objective fit made +0.396 R a trade over 24,062 out-of-sample trades, against -0.201 R over 42,905 trades for classic; one R is twice the average true range at the unit's fill. It did better than classic on 73% of the 529 names (mean +34.3 R a name, t 11.4) and better than one shared set on 57% (mean +6.8 R a name, t 2.5). Its pooled daily Sharpe was -0.004 against -0.310 for classic, so the edge is per trade, not a profitable portfolio. The win/loss score did better than classic (mean +16.4 R a name, t 7.6) and was no clear difference from one shared set (mean -5.8 R a name, t -1.9). The first 382-name run pointed the same way (+0.372 R against -0.139; pooled Sharpe 0.30 against -0.02), and extending the grid on the same names lifted the Sharpe fit's net R from +0.37 to +0.50. Those early pooled Sharpes were most likely lifted by 13 near-cash funds that the later quiet-market filter removed. Sixteen comparisons were read without adjustment, and the t-statistics treat names as independent, so they overstate the evidence.
The report's own label is exploratory, so this page states a descriptive verdict: a parameter fit and a walk-forward description, not a counted test, with no pre-registered bar. Sixteen pooled comparisons were read without adjustment, and the t-statistics treat names as independent, so they overstate the evidence; the fitted rule's pooled daily Sharpe in the final run is about zero (-0.004). The family of 16 is the number of comparisons the study read, a floor on the search, not a pre-registered family; no holdout was set aside.
Backtested by QS Research
Pre-registration
No pre-registration. The date 2026-11-21 is when this retrospective record was written, after every result; nothing was fixed in advance. It records the study as run: split- and dividend-adjusted daily US stock and ETF bars from January 2016; a per-market sweep of the entry channel (10 to 200; 10 to 100 in the first runs), the exit channel (5 to 80; 5 to 50 in the first runs) and N (10, 14, 20, 30, 40), with the exit shorter than the entry; the classic 20/10/20, a 2N stop, one unit added every 0.5N up to four, and gaps filled at the open; costs of 2.5 basis points of the fill plus 0.05N a unit on both sides, with no borrow charge. A cell is scored only with at least 15 closed trades in the window; its plateau is the mean of the cell and its grid neighbours, and needs at least half of those neighbours scored; the classic set is kept unless another plateau is higher. For each test year the pick is made on the prior 1,260 sessions and applied to that year, against classic and against one set shared by every market. Each pooled comparison was read alone at t of 2 or more, a screen the method calls not a test; there was no cost stress, multiple-testing adjustment, holdout or recorded trial. Sixteen pooled comparisons were looked at and 11 were written up before the record; the grid was extended and a quiet-market filter added after out-of-sample results were seen. There was no pass bar, and none is invented here.
Registered 2026-11-21Trials · 0 trials · 0 passed · judged at N = 16 trials; 2965 trials run in total
| Trial | Verdict | |
|---|---|---|
| Trades | RESULT | not counted |
| Net R per trade | RESULT | not counted |
| Win rate | RESULT | not counted |
| Win/loss | RESULT | not counted |
| Sharpe of the pooled daily curve | RESULT | not counted |
| sharpe | RESULT | not counted |
| win/loss score | RESULT | not counted |
| 2026-11-21 01:23 and 01:51, sharpe | RESULT | not counted |
| 2026-11-21 01:23 and 01:51, win/loss score | RESULT | not counted |
| 2026-11-21 08:11 and 14:10, sharpe | RESULT | not counted |
| 2026-11-21 08:11 and 14:10, win/loss score | RESULT | not counted |
| sharpe · per-market fit | RESULT | not counted |
| sharpe · classic 20/10/20 | RESULT | not counted |
| win/loss score · per-market fit | RESULT | not counted |
| win/loss score · classic 20/10/20 | RESULT | not counted |
| sharpe | RESULT | not counted |
| win/loss score | RESULT | not counted |
| Net R per trade (classic: -0.14) | RESULT | not counted |
| Sharpe of the pooled daily curve | RESULT | not counted |
| Per-name difference vs classic, t | RESULT | not counted |
| Per-name difference vs one shared set, t | RESULT | not counted |
| Names at the longest entry tried | RESULT | not counted |
Limitations
- Not pre-registered. The record was written after every result; nothing was fixed before the data, so no result here is a pre-registered test.
- The grid was extended to entry 200 and exit 80, and the quiet-market filter was added, after out-of-sample results were seen; the later runs report on rules that the earlier results helped choose.
- Survivorship: every universe is a November 2026 liquid list, so names that fell out of it since 2016 are missing.
- The t-statistics treat names as independent; names move together, so they overstate the evidence. Sixteen pooled comparisons were read without adjustment, and the family of 16 is a floor on the search, not a pre-registered family.
- No data snapshot and no binary build are recorded, so the runs cannot be reproduced exactly; the first two 382-name runs kept no output file at all, and only the 2026-11-21 logs and walk-forward extracts can be checked.
- The quiet-market filter selects names on their last 1,260 sessions, which include years the walk-forward tests, so early test years use later information; it also removed 13 near-cash funds, and the pooled Sharpe figures before it were most likely lifted by them (an inference from the logs, not a separate run).
- The edge is per trade: in the final run the pooled daily Sharpe of the fitted sets is about zero (-0.004 for the Sharpe fit, -0.006 for the win/loss fit), so this is not a profitable portfolio as tested.
- No holdout was set aside; the walk-forward is the only out-of-sample read, and the rule choices used it.
- The one shared set is a single length set chosen per year from the fit windows, not a portfolio; the comparison is with that baseline, not with a live strategy.
The numbers: every trial's checks, metrics, curve and yearly results, and the provenance
1 published version · current 6bf59c52df08afb7, published 2026-11-21