In 1983, Chicago trader Richard Dennis bet his partner William Eckhardt that trading could be taught to anyone. He recruited a group of novices, nicknamed them the turtles, handed them a fully mechanical rulebook, and staked them with his own money. Several of them went on to run hundreds of millions. The rulebook leaked decades ago, so the turtle system is one of the few famous strategies whose rules are actually public.
That makes part of it testable, and the honest place to start is saying which part.
The rules, in full
The system the turtles were taught has six parts:
- Entry, System 1: buy when price ticks above the highest high of the previous 20 days. Skip the signal if the previous System 1 trade would have been a winner; a 55-day breakout acts as a failsafe entry so a big trend is never missed entirely.
- Entry, System 2: buy when price ticks above the highest high of the previous 55 days. Never skipped.
- Position size: N is the 20-day average true range, a volatility unit. A position ("unit") is sized so a 1N move equals about 1% of the account, so a volatile market gets a smaller position.
- Stop: exit if price moves 2N against the entry, which risks roughly 2% of the account per unit.
- Pyramiding: add another unit each time the trade moves half an N in your favor, up to four units in one market.
- Exit: System 1 exits at a 10-day low, System 2 at a 20-day low. No profit targets, ever; the whole design is to hold winners until the trend visibly ends.
The entries are Donchian channel breakouts, named for Richard Donchian, who traded them decades before Dennis. Nothing in the list is secret or clever. Dennis's argument, which the bet was designed to settle, was that the edge is in following the rules, not in knowing them.
What a backtest can and cannot isolate
Our free backtester expresses entries as rules and exits as a fixed hold, so it can test a turtle-style entry: does breaking out of a 20- or 55-day channel tell you anything about the next couple of weeks? Three approximations, named up front:
- The tool's Donchian rule breaks the highest close of the prior window (today excluded), not the highest intraday high the turtles used. The two disagree often enough to matter.
- The test takes every breakout. System 1's skip rule (stand aside after a winner) is an entry filter this test does not apply, so the tested arm is a plain Donchian breakout, not System 1.
- There is no 2N stop, no pyramiding, and no channel exit, because a fixed hold removes them by construction. Those parts are where the turtle returns actually came from: the system expects to lose on most trades and make it back on a few huge ones, so a fixed-hold win rate is precisely the statistic the turtles did not optimize.
With that boundary drawn, the entry itself is still worth measuring.
What the two entries showed
We ran both breakout lengths over 22 US-listed large-cap stocks and sector ETFs, over the window the backtester covers, buying each breakout close and holding ten sessions, one position per name, with commission and slippage. The number to beat is not 50%: over this mostly rising window, buying on any day at all and holding ten sessions finished green 55.4% of the time (24,948 windows). Against that:
| Entry | Trades | Finished green | vs base rate | p-value |
|---|---|---|---|---|
| 20-day breakout | 980 | 58.7% | +3.3 points | 0.043 |
| 55-day breakout | 654 | 56.9% | +1.5 points | 0.46 |
Read the 0.043 with both hands on the caveats. This test series has now published around nineteen signal arms (candles, oscillators, divergences, regime splits, these two), and across nineteen looks, chance alone would be expected to put about one of them under 0.05. This is that one. It is also the only arm in the series to land under the line in the beating direction, the trades are moderately spread (980 trades on 521 distinct entry days, largest single-day cluster 12 of 22 names, on 2025-05-02), and the longer hold points the same way: 671 twenty-session holds won 59.6% against a 55.9% base rate (+3.7 points, p=0.057), with a 25.3% median drawdown against the benchmark's 33.8%. So the honest summary is not "probably real". It is: the most promising number this series has produced, and exactly what one lucky arm out of nineteen would look like. A signal on correlated names also never has fully independent trades, which widens every interval further. (For calibration: the regime study's best arm printed p=0.05, which rounds from 0.053, just above the line.)
The 55-day entry is the surprise in the other direction. The turtles trusted System 2 as the reliable one, the entry you never skip. On these names and this window, the slower breakout's win rate is statistically nothing. Its returns were not: +4.8% median per name at 26.1% time in market, against the 20-day entry's +4.1% at 38.4%, with the same 20.5% median drawdown. Fewer, later entries won less often and still made slightly more per unit of exposure, which is the turtle thesis in miniature: win rate and profit are different questions.
Win rate is not profit, and the turtles knew it
Median outcome per name of trading the 20-day entry with ten-day holds: +4.1% total over four and a half years, in the market 38% of the time, with a 20.4% median drawdown. Buy-and-hold made +57.3% with a 33.8% drawdown over the same span. At twenty-day holds the entry made +15.8% at 54% exposure with a 25.3% drawdown.
So the entry wins slightly more often than chance, and converting that into buy-and-hold-beating profit is the part the fixed hold cannot do, which is exactly why the real system bolts on the 2N stop, the pyramiding, and the trailing channel exit. The turtles' returns came from occasionally riding a position for months at up to four units. A ten-day hold caps every winner at ten days; it measures the entry's information, not the system's engine.
The same entry has a regime dimension, measured in our regime study: 59.6% of ten-day holds won while the S&P 500 was above its 200-day average, against that regime's own 56.6% base rate (+3.0 points, p=0.16), and 71.3% in the rare windows below it, against a 64.2% base rate there (+7.1 points, p=0.23, on 80 trades of which 12 fired on one day). That page's conclusion applies here too: no version of this entry demonstrably beat just being long in its regime.
Test it yourself
The entry is one rule in the free backtester: open Chart setup, add Donchian channel break, set it to close above the prior 20-day extreme, and pick a hold. Switch 20 to 55 to run the slower entry. TC2000 users can approximate the same scan with a PCF like C > MAXC20.1 (the highest close of the prior 20 bars, today excluded; the turtles' own version would use MAXH20.1 and intraday highs); our EasyScan setups guide covers the scan mechanics on that platform.
The experiment worth running is the one this page ran: compare the breakout's win rate to buying on any day at all over the same window, not to 50%. And if you want to see the trend-following half that a fixed hold cannot express, run the same entry with a longer hold and watch the win rate stay flat while the return distribution stretches.
Frequently asked questions
Do the turtle trading rules still work? The entry half showed the most promising number in our test series: the 20-day breakout won 58.7% of ten-day holds against a 55.4% base rate (980 trades, p=0.043), which is suggestive but also exactly what one arm in nineteen produces by chance. The system half, volatility sizing, the 2N stop, pyramiding, and exits that ride trends for months, is not testable with fixed holds, and it is where the famous returns actually came from. "The rules are public" and "the returns are reproducible" are different claims.
What are the turtle entry rules? Buy when price ticks above the highest high of the previous 20 days (System 1, skipped if the previous System 1 signal would have won, with a 55-day failsafe) or of the previous 55 days (System 2, never skipped). Both are Donchian channel breakouts. Our test used the backtester's close-based version of the channel and took every signal, without the skip rule.
What is the 2N stop? N is the 20-day average true range, a volatility unit. The turtles sized a position so a 1N move equaled about 1% of the account, exited any trade that moved 2N against the entry (roughly a 2% loss per unit), and added units every half-N in their favor up to four. It is the risk-management half of the system, and nothing on this page tests it.
Is the 20-day or the 55-day breakout better? In our window, on large-cap US names, the 20-day entry won more often (58.7% vs 56.9% of ten-day holds) but the 55-day entry returned slightly more at two-thirds the market exposure. The turtles preferred System 2; on win rate our data leans the other way, on return per unit of exposure it leans theirs. Their markets were 1980s commodities, not 2020s mega-cap equities, so treat neither number as universal.
Before you trade this
These figures come from one run over 22 US-listed large-cap stocks and sector ETFs, December 2021 to mid-2026, a mostly rising window; the entry is the backtester's close-based Donchian channel (not the turtles' intraday-high channel), bought on the breakout close and held a fixed 10 or 20 sessions, without System 1's skip rule, which tests the signal's information and deliberately not the turtle system's stops, sizing, pyramiding, or trend-riding exits. Win rates are before costs and compared against a like-for-like base rate; the return and drawdown medians include commission and slippage. Around nineteen signal arms have been published in this series, so one p-value under 0.05 is what chance predicts, and signals on correlated names are not fully independent, which widens every interval further. The backtester's window rolls forward daily, so re-running these rules will not reproduce these exact figures. The run is committed in the repository (backend/scripts/research/candle_study.py, the --rule 'donchian:...' path with --baseline) so it can be repeated rather than trusted. Past prices are not a forecast, and nothing here is investment advice.