BTC Donchian Channel Breakout — Long-Short, Classic 2-Parameter Turtle-Style Trend Capture, Wide Slow Channel, ATR Stop, No Regime Gate (BTCUSDT.BINANCE USD-M, 1D)

A LONG-SHORT, single-instrument, single-leg directional strategy on BTCUSDT.BINANCE USD-M using the classic Donchian channel breakout (the Turtle mechanism) — deliberately the SIMPLEST, most parameter-sparse trend design possible, in direct response to the dominant failure mode in the log: OVERFIT. My gated/complex momentum variants (alt-breadth, vol-regime-band) died with IS Sharpe 7.6 collapsing to OOS -2.1 and deflated_sharpe 0, precisely because extra parameters and reduced trade counts created a large best-of-N noise surface. The lone survivor is the simple ungated BTC momentum book. This strategy pushes simplicity to the limit: only TWO structural parameters (entry channel length, exit channel length), NO regime gate, NO vol filter, NO breadth confirm — nothing for an optimizer to curve-fit. It also uses a DIFFERENT entry trigger than everything in the pipeline (which is all EMA-momentum-confluence or mean-reversion): a breakout enters only on a genuinely NEW N-day extreme, which by construction fires rarely and only when a real directional move is underway, giving large per-trade moves that make fees negligible and reducing the whipsaw that a fast EMA cross suffers in ranges. It trades BTC because BTC is the one asset where directional trading has survived (alts die as beta/outlier artifacts). Single perp leg with a hard ATR stop — NO hedge leg, NO leverage stacking — so the cross-account liquidation flaw that killed the carry/pair books is structurally impossible. Long-short (fills the under-weight direction bucket) and daily/multi-week horizon (fills the sparse long/macro horizon). This is NOT the dead pure-TSMOM (sign-of-trailing-return) book — it is a channel-breakout with an explicit channel-based exit and ATR stop, a distinct, empirically robust, low-parameter classic.

Implements the hypothesis literally and minimally. (1) Two structural parameters only — entry_len and exit_len — with the entry trigger being a genuinely NEW N-day extreme: the channel is computed over the N bars BEFORE the current bar, so a breakout is not tautological. (2) No regime/vol/breadth gate, addressing the overfit failure mode: the optimizer's search surface is deliberately tiny, and entry_thresh is clamped to [0.95, 1.15] (declared in _param_bounds) so it cannot be curve-fit into a mid-channel always-in rule; entry_len/exit_len/atr_len/atr_stop_mult clamps are likewise declared. (3) The signal is CONTINUOUS and varies on every bar (normalised channel position), not gated behind the entry condition, so it is inspectable and never frozen. (4) Single perp leg, no hedge, no leverage stacking (leverage=1.0, and sizing references config.leverage only in the notional ceiling), so the cross-account liquidation failure mode is structurally impossible. (5) On daily BTC a 30/12 channel is wide/slow enough that entries fire only on real directional moves (per-trade moves of many percent, dwarfing the ~0.10% round-trip taker cost) yet still fires roughly 8-15 times a year, well above the zero-trade sandbox floor — I chose 30 rather than the classic 55 precisely so the Layer-3 window and the 15-day holdout remain measurable, with 55-80 still inside the declared optimizer bounds. (6) Risk-per-trade is capped at 2.5% of equity over a 3-ATR stop rather than sizing off equity*leverage every bar, which is the compounding-sizing pattern that produced past 80-100% drawdowns. One caveat, stated as required: a channel breakout on BTC daily is a well-known, widely-traded mechanism, so any surviving edge is likely modest and regime-dependent (it will give back open profit in every range) — but the low parameter count is exactly what should make its walk-forward OOS hold up rather than collapse.

Outlier-mirage OHLCV breakout, the L9/L41 class (0/213 survival). The +119.6% total return is carried almost entirely by ONE day — 2021-03-24 +82.1% (2021 annual +75.8%) — with return_kurtosis 287.5 and skew 10.3 confirming a single observation dominates the whole distribution; remove it and there is no edge. Every year outside that 2021 spike is negative or flat (2022 -4.6%, 2023 -5.4%, 2024 -6.4%, then modest 2025 +9.0%/2026 +3.6%), so this is captured luck on one breakout, not a distributed edge. It fires only 49 times in 6.6 years, so the Sharpe (0.57, CI-low -0.04 straddling zero) is statistically unmeasured (L10) and the 7.78% avg trade is itself the one outlier trade. benchmark_meaningful is true and alpha is only 0.058 with IR -0.57 — it underperforms simply holding BTC. The simplicity/low-parameter framing does not cure this: a 49-trade, kurtosis-287 series sent into the 3-phase run would best-of-N overfit and collapse in walk-forward OOS/holdout exactly as the vol-regime and alt-breadth siblings just did (DSR ~0). Additionally, a +82% single-day return at ~29% avg position is implausible from BTC price action, indicating a compounding/attribution concentration that inflates the metrics. Abandon rather than spend the optimization budget.