Market Wizards, wave 2: four classic trading rules traded mechanically
Do the Turtle, Darvas box, Minervini VCP and Weinstein stage-2 rules make money after costs on US daily bars from January 2016 to May 2022?
No. None of the four rules passed. Turtle came closest: it made +1,003.5% in total (45.4% a year), beat SPY by +3.12% a month on average and had a Sharpe of 0.96, but its deflated Sharpe was 0.08 against the 0.80 bar (0.66 even at the pre-registered family of four), so the edge does not survive the number of trials it was judged against. It also fell 51.1% from its peak on daily marks (48.6% on month ends). The Darvas box lost 46.0%, Minervini's VCP 44.4% and Weinstein stage 2 33.7%; each lost money after costs and failed every return check and the deflated Sharpe. SPY made +105.4% over the same years. All four trials are counted; none passed, so none went on to validation or the holdout. The binding run is a re-run on pinned data that reproduced every return and check of the earlier run exactly.
Backtested by QS Research
Pre-registration
Four rules were fixed and hashed before any run: Turtle System 1 (a 20-session breakout, a 10-session exit, a 2N stop and up to four units); a Darvas box (a 55-session breakout, a 20-session exit, a 2.5N stop and one unit); Minervini's volatility-contraction entry (a 20-session breakout above a rising 50-session average, a 2N stop and two units); and Weinstein stage 2 (a rising 150-session average with a 40-session breakout, a 2N stop and two units). Each is long-only, risks 1% of equity per N, holds at most 20 names and runs on US daily bars from 4 January 2016 to 31 May 2022. Each trial counts whatever its result. The bar was the standard promotion gate fixed at the time: costs by liquidity tier, a deflated Sharpe at the pre-registered family of four trials, the first and second half of the months, the best month removed and stressed costs. Only a trial that passed discovery could go on to validation or the one holdout.
Registered 2026-11-02Amendments
- 2026-11-02 After the first run, the gate was found to average only months holding three or more names, to re-simulate the stressed-cost check on different trades, and to deflate the Sharpe only against the trials run so far. All three were fixed without changing any rule, spec, window, data or pass bar, and all four trials were rerun on the same frozen specs: every month is now judged, the stressed cost is charged on the same trades, and the deflated Sharpe uses the pre-registered family of four. That rerun was the binding result until the pinned re-run of Amendment 2; run 1 stays on record. approved: Recorded with the binding rerun after run 1; accepted in independent review, 2026-11-21, as procedural only.
- 2026-11-21 Before the run: the four frozen rules are re-run, on the same window and gate, on pinned data snapshots of the daily bars and the monthly panel, from a signed release, and that run becomes the binding result whatever it shows. The earlier panel file had been rebuilt in place, so this is a new run on fixed inputs, not a byte replay. It reproduced every return, check and backtest id; only the deflated Sharpe's trial count changed. approved: Committed before the run on the owner's instruction to close the unpinned-data gap; accepted in independent review, 2026-11-22.
A holdout sample is sealed: its results stay unread until the study is final.
Trials · 4 trials · 0 passed · judged at N = 4 trials; 2972 trials run in total
| Trial | Verdict | |
|---|---|---|
| Turtle 20/10 | FAIL | counted |
| Darvas box | FAIL | counted |
| Minervini VCP | FAIL | counted |
| Weinstein stage 2 | FAIL | counted |
| Turtle 20/10 | FAIL | not counted |
| Darvas box | FAIL | not counted |
| Minervini VCP | FAIL | not counted |
| Weinstein stage 2 | FAIL | not counted |
Limitations
- Prices are split-adjusted, not total-return: dividends are excluded, which understates a long book by roughly the dividend yield of the names held.
- Survivorship: the store holds the names it has bars for, so a breakout system on a survivor universe is flattered. A pass would have been re-checked on the survivorship-free universe before promotion.
- The binding run is the 22 Nov re-run on pinned data snapshots (daily bars and the monthly panel, each by content hash). It reproduced the earlier unpinned run exactly: every return, every check and every backtest id. Its deflated Sharpe was computed by the engine against the shared daily record of 1,741 effective trials, a stricter bar than the pre-registered family of four; Turtle's is 0.08 there and 0.66 at the family of four, below 0.80 either way.
- Stop fills pay half the liquidity tier's per-side cost as slippage; a gap through a stop fills at the open.
- The earlier 2 Nov run did not record whether its build was clean; the binding re-run records a clean, signed build.
- The rules approximate the books' methods: Darvas's weekly box becomes a 55-session high, and Minervini's entry omits the trend template's relative-strength and 52-week-high filters. The authors' discretion, full method and their own position sizing are not tested.
- One window and one market regime: January 2016 to May 2022, mostly rising. The rules were not tested after May 2022, and nothing passed, so no validation or holdout was run.
- Drawdowns are the daily peak-to-trough: 51.1% for Turtle, 71.2% for Darvas, 67.8% for Minervini and 61.6% for Weinstein. The month-end figures are shallower (48.6%, 63.8%, 65.2% and 54.8%).
- The promotion gate's positive control (a planted edge it should pass) had not been run when these trials were judged; the pre-registration says every run before it carries that caveat. A fail is evidence about these rules, not a validation of the gate.
- The engine recorded the four trials under its shared daily family label, not the pre-registered family name. The counted set is exactly the four trials described here.
- The strategies are long-only, capped at 20 names and sized at 1% of equity per N; these exact parameter choices are what was tested, not a range of them.
The numbers: every trial's checks, metrics, curve and yearly results, and the provenance
3 published versions · current d6c04deea09d0e4e, published 2026-11-22