turtle

The bet: classic Donchian breakout.

Turtle is the only strategy that leaned positive at all — and we are saying plainly it still did not pass. In the tuned run the deployable long side reached a pooled t of +2.09, the one positive, raw-significant signal in the whole program. But all eight coins are underpowered (8-17 trades each); the long pool’s significance is inflated by a few outsized winners (one ADA trade returned +75.8%); and once each system is pinned to a single configuration — removing the in-sample lift from picking the best of four configs per window — the long-side t drops to +1.76 / +1.24. Nothing comes within sight of the program’s 2.94 bar.

A faithful original Turtle on the daily clock: enter on a Donchian channel breakout, exit on the opposite shorter channel or a 2N (two-ATR) stop, letting winners run. It is the first momentum construction whose long side leans positive — daily and let-winners-run captures the rare large trend and the long-bias of the tape that the 1h-4h versions churned through. If a persistence premium survives cost on this universe, this is where it lives: long-only, daily or slower, more history and more coins. The daily clock — not the strategy — is the binding constraint (24 or fewer trades per coin even with ~5 years of data).

Pooled 98 out-of-sample trades: long side n=48, +10.4%/trade, t +2.09; short side n=50, -3.6%/trade, t -1.71. Long-only pinned (the deployable test): System 1 pooled t +1.57, equal-weight Sharpe +0.260; System 2 pooled t +1.27. Directionally right, and dropping the losing short side is the correct trade — but it still clears nothing.

⚠ = underpowered (fewer than 30 out-of-sample trades). Negative values in red.

This is the one result with a pulse, so it is running live on a small scale to gather the out-of-sample evidence the backtest could not. It is a forward test, not a proven strategy — you can watch it arm and wait for a breakout in real time, and if it fails live you will see that too.