Turtle Trading System (Donchian breakout)

The claim

Richard Dennis bet William Eckhardt that trading could be taught to anyone. His 'Turtles' — recruited by newspaper ad, trained for two weeks — traded a fully mechanical Donchian breakout: buy a 20-day high, sell a 20-day low, risk a fixed fraction per trade, cut at 2N. Several went on to run funds. It is the origin story of systematic trend following.

Where it comes from

Richard Dennis & William Eckhardt, 1983–84. The exact rules were published by former Turtles (Curtis Faith's 'Way of the Turtle' and the free 'Original Turtle Trading Rules' PDF), making this one of the few legendary systems with a canonical, checkable spec.

The exact rule we graded

Long when price breaks above the 20-bar high, short when it breaks below the 20-bar low (System 1), on 1-hour candles. Exits: 2-ATR catastrophic stop (the Turtles' 2N rule), 2-ATR trailing stop.

• ✕OOS net expectancy > 0
• ✕Clears the cost hurdle
• ✕Robust, not lucky
• ✕Survivable drawdown
• ✓Not overfit

The backtest above is history. Since publication, this exact spec also runs in our nightly forward-test harness (the same one behind /forward-tests): only trades entered after publication count, open positions are never force-closed, and the record publishes either way.

No closed forward trades recorded yet — the record accrues nightly as trades complete their full exit windows.

Honest port notes

• The full system has parts our exit engine can't express, all disclosed: the 10-bar-opposite-extreme exit is approximated with a 2-ATR trail, pyramiding (up to 4 units added at ½N spacing), N-based unit sizing, and System 1's skip-after-a-winning-breakout filter are dropped. Our stop ATR is 14-period; the Turtles' N was 20-day.
• System 2 (55-bar breakout, the Turtles' slower variant) graded F at −3.3bp net across 4,009 out-of-sample trades — less bad than System 1's −8.0bp, consistent with the library-wide pattern that slower trend signals decay less at the 1-hour clock. Public verify id: 858b5117-da9e-4450-a0c0-e15696663437.
• The context grade is the story. The literal daily-bar System 1 (BTC/ETH/SOL) passes FOUR of five gates — +52.0bp net out-of-sample across 111 trades, clears the fee hurdle, positive on 2 of 3 assets, 7.8% max drawdown at 1% sizing. It fails only the overfit gate: +259bp per trade in the training window decayed to +56bp gross out-of-sample. No classic in this library has come closer to passing. Forty-year-old rules, still net-positive out-of-sample on an asset class that didn't exist when they were written — and still not deployable by our gates, because a 4.6× train→test decay is exactly what the overfit gate exists to catch. We even ran our Strategy Doctor on it: zero repair candidates — a two-condition system has nothing left to simplify, and overfit is the one verdict honest tuning can't fix. Public verify id: c83c794b-ab45-4fc4-91ea-fec37b225886.

{
  "name": "Turtle Trading System 1 (20-bar breakout)",
  "universe": [
    "BTC/USDT:USDT",
    "ETH/USDT:USDT",
    "SOL/USDT:USDT",
    "BNB/USDT:USDT",
    "XRP/USDT:USDT",
    "DOGE/USDT:USDT",
    "LINK/USDT:USDT",
    "AVAX/USDT:USDT"
  ],
  "risk": {
    "risk_per_trade_pct": 1,
    "leverage_cap": 3,
    "max_daily_loss_pct": 5,
    "max_dd_halt_pct": 25
  },
  "sleeves": [
    {
      "name": "main",
      "timeframe": "1h",
      "entry": {
        "type": "rule",
        "logic": "all",
        "long": [
          {
            "indicator": "price",
            "period": 1,
            "op": ">",
            "vs_indicator": "donchian_hi",
            "vs_period": 20
          }
        ],
        "short": [
          {
            "indicator": "price",
            "period": 1,
            "op": "<",
            "vs_indicator": "donchian_lo",
            "vs_period": 20
          }
        ]
      },
      "exit": {
        "cat_stop_atr": 2,
        "trail_atr": 2
      }
    }
  ]
}