BtcEfficiencyGatedChannelBreakoutLS

BTC Efficiency-Gated Volatility-Channel Breakout — Trade Trends ONLY in Clean (High Kaufman-Efficiency) Regimes, Stay Flat in Chop, Ride With a Chandelier Trailing Stop (BINANCE USD-M, 4H Bars, Long-Short, Asymmetric Let-Winners-Run, 3-Parameter)

A LONG-SHORT, single-instrument, ASYMMETRIC-PAYOFF trend strategy on BTCUSDT.BINANCE (USD-M perp), deliberately built around the ONE fact this session has established beyond doubt: every SYMMETRIC mean-reversion / contrarian-fade / market-neutral sleeve dies fee_edge or no_edge (ETH funding-contrarian breakeven, SOL OI-washout negative-edge, LTC-BCH and SOL-ETH pairs below the fee floor), while the SINGLE surviving strategy in 1163 experiments is a directional BTC trend construction that lets winners run with a trailing stop. So this stays firmly in the only living family — but with a mechanism DISTINCT from both the pending dual-TF momentum-confluence survivor (which gates on 4H vs 1D momentum-sign agreement) and the pending ETH OI-confirmed breakout (which gates on open-interest expansion). Here the gate is a REGIME-QUALITY filter: the Kaufman Efficiency Ratio (ER = net move / sum of absolute moves) measures how CLEAN the recent path is. Breakouts fail overwhelmingly in CHOP; by only taking a volatility-channel breakout when ER shows a genuinely trending regime and staying FLAT otherwise, the strategy sidesteps the whipsaw that is the primary trend-following killer — and the chandelier trailing stop delivers the fat right tail that lets a modest win rate clear fees (the structural property the dead symmetric fades lacked). It avoids every graveyard: NOT symmetric reversion (dead), NOT the sprayed dual-TF momentum template on a new ticker (L56 — different mechanism, and on BTC, the one asset where trend is PROVEN to survive the analyst's gates rather than an alt where it is 0-for), NOT funding carry (L57), NOT pairs/cross-sectional (dead), NOT options (infra-broken), NOT a liquidation feed (L58). Pure OHLCV, computationally TRIVIAL: a rolling ER, a rolling channel, and an ATR — a few scalars per bar, no full-history rescans, no sorted()/min() in the hot path. Fills the under-represented long-short bucket. Exactly 3 tunable parameters: the ER trend-quality threshold, the channel lookback, and the chandelier ATR multiple.

Clean restart: the previous abandonment was pipeline-caused (Layer-2 staticmethod mis-binding, QA performance verdicts on an unoptimized smoke test, backtest-slot starvation), so the mechanism is kept exactly as verified — instance-method helpers only, fully incremental state (Wilder ATR, running abs-move sum, bounded deques, no full-history rescan), continuous every-bar signal in natural range-position units, and a directionless ER regime gate. The one change is a genuine off-by-one fix in the ER estimator: _er_closes was sized er_period + 1, so (read before the current close is appended) the numerator |close_t - close_{t-er_period-1}| spanned one more move than the er_period absolute moves in the denominator's path sum. That biased ER upward — it could exceed 1.0 and was silently clamped — which quietly loosened the chop filter that is the entire point of the hypothesis. With maxlen = er_period the numerator and denominator cover the same window and ER is a true ratio in [0, 1]. Nothing else was touched: entry/exit/sizing logic, imports, parameter set and clamps (mirrored in _param_bounds) are unchanged.

Overfit / no-generalization, at the iteration limit (attempt 2 of 2). Promote is impossible: two HARD gates fail unwaivably — PBO 0.5106 > 0.5 (selection more likely than not overfitting) and the holdout has only 3 trades (< 10 min), so its 1.67 Sharpe is uninferable. Optimization actively degraded the strategy: it drove er_period to 55 / break_thresh to 1.80, collapsing 235 trades into 14 over 6 years — PF 15.5 / win_rate 0.79 on 14 trades is a curve-fit to a handful of clean windows, not an edge. Robustness confirms noise: deflated_sharpe 0.783 << 0.95, is_significant=false, Sharpe CI [0.448, 1.582] barely off zero, fails programme FDR (keeps 3 of 231, candidate_p 0.217), and WF-OOS windows are -1.95 / 0.99 / 3.07 (avg 0.70, first window deeply negative). This is the overfit-not-dead-premise case per the decision rule (high IS collapsing to near-zero significance, thin negative-window OOS) → abandon, not revise: the identical Kaufman-ER-gated breakout/momentum mechanism is already 0-for-5 across BTC/ETH/SOL and a majors basket (all abandoned overfit or fee_edge), so re-pointing the same mechanism at another target would simply overfit again. There is no untried, robust parameter region — the honest 235-trade default config also sits below every promotion floor (Sharpe 0.98). Failure pattern: overfit.

Long-short BTCUSDT (Binance USD-M, 4H) volatility-channel breakout that only fires when Kaufman's Efficiency Ratio shows a clean, genuinely trending path, and stays flat in chop. The signal is the continuous range position (close - channel mid) / (half channel width) computed every bar from the PRIOR channel_lookback bars, so +-1 is exactly the channel edge and a breakout reads beyond it. Exits are a chandelier trailing stop only (highest high since entry minus chandelier_atr x ATR, mirrored for shorts) — no fixed take-profit — giving the asymmetric let-winners-run payoff. Sizing risks risk_frac of equity per ATR-based stop distance, capped at 1x equity notional.

CLEAN RESTART 2026-09-04 — this run's verdict history and learning records were removed and it was restarted from verification. Its previous abandonment came from the pipeline, not from the market: the Layer-2 harness mis-bound @staticmethod helpers (fixed), QA issued terminal performance verdicts on an unoptimized smoke test (removed — QA now judges correctness only), and sandbox timeouts came from backtest-slot starvation (fixed). The hypothesis and the strategy code are unchanged. Verify the code on its merits; performance is decided later by the full backtest and the optimizer.

No action required; confirm the base class honors the override (sandbox produced 231 trades, so warmup gating works).

min_bars_required overrides a framework method (Layer-1 base_shadow warning). Behavior appears intentional and correct (returns a sensible warmup covering channel/ER/ATR windows), synthetic tests and sandbox both warm up and trade normally.

Information ratio vs buy-hold is -0.6447 in the smoke test despite positive alpha (0.094) and low beta (0.037) — the strategy adds absolute return but underperforms simply holding BTC on a benchmark-relative basis over this truncated window. Worth the analyst re-checking IR / alpha significance on full history at optimized params before promotion; not a correctness issue and not blocking.

Ample, measurable sample: 235 trades over 6.6 years (128 long / 107 short) — supports a meaningful 3-parameter walk-forward without curve-fitting to noise.

Edge clears costs with margin: avg_trade_return_pct 0.878% vs the 0.15% Binance USD-M floor; commissions only 3.55% of gross, so not fee-fragile.

Solid, plausible risk-adjusted profile: Sharpe 0.98 (PSR 0.977), Sortino 1.42, Calmar 1.32, max DD 8.1%, recovery factor 11.7 — no artifact signature (no implausible Sharpe/win-rate/monotonic equity).

Trades match the hypothesized mechanism: long-short breakout with ~4.5d avg hold and asymmetric let-winners-run (avg_win 3115 vs avg_loss 1032, tail_ratio 1.23, return_skew 1.70) and a 34% win rate — exactly the fat-right-tail trend profile described.

Sits in the only surviving family (BTC directional trend with trailing stop), via a distinct regime-quality (Kaufman ER) gate; positive in 6 of 7 years, calm/normal/stressed regimes all positive.

Information ratio -0.727 vs buy-hold BTC (QA edge concern): strategy underperforms simply holding BTC on a benchmark-relative basis. Mitigated by low beta (0.046), positive alpha (0.088) and 8% DD — it deliberately avoids full-cycle drawdown — so judge on absolute risk-adjusted metrics, but analyst should re-check alpha/IR significance at optimized params before any promotion.

Modeled impact_cost_pct 2.49% and one large outlier day (2026-08-28 +12.6%, 2021-01-11 +16.1%) contribute meaningfully to the headline; robustness/deflated-Sharpe gating in the 3-phase run is the real test.

2025 was a losing year (-3.45%); a couple of parameters (er_threshold, chandelier_atr) will need sensitivity checking for cliffs.

Per-trade edge clears fees comfortably in both runs (initial avg_trade_return_pct 0.878%, optimized 6.15% vs the 0.15% Binance USD-M floor) and commissions are only 3.55%/1.55% of gross — this is NOT a fee_edge death.

The default-parameter backtest was economically honest: 235 trades (128 long / 107 short) over 6.6 years, PF 1.56, DD 8.1%, positive alpha (0.088), low beta (0.046) — a genuine directional trend construction, not market beta.

Sensitivity grid is mostly flat (only er_period flagged as a cliff) and walk-forward is_overfitted=false by the ratio test.

Optimization made the strategy WORSE, not better: the optimizer chose er_period=55 and break_thresh=1.80, which collapsed the trade count from 235 to 14 over 6 years (10 long / 4 short). PF 15.5 and win_rate 0.79 on 14 trades is the textbook curve-fit signature — the tuned config fits a handful of clean windows and is statistically empty.

Two HARD gates fail and cannot be waived: PBO 0.5106 > 0.5 (the parameter selection is more likely than not overfitting) and holdout is only 3 trades (< 10 min), so the 'passing' holdout Sharpe 1.67 is uninferable noise.

Robustness fails across the board: deflated_sharpe 0.783 < 0.95, is_significant=false, Sharpe CI [0.448, 1.582] is barely off zero, and the candidate does NOT survive programme-level FDR (Benjamini-Hochberg keeps 3 of 231; candidate_p 0.217).

Walk-forward OOS is fragile: per-window OOS Sharpes -1.95 / 0.99 / 3.07 (avg 0.70) — the first window is deeply negative and the 3.07 is a single-window outlier carrying the average; OOS 0.70 < the 1.5 in-sample floor.

The ER-gated breakout/momentum family is 0-for-many in the outcome store: ETH ER-gated TSMOM (overfit), SOL ER-gated breakout (overfit), BTC ER-gated momentum (overfit), BTC ER-gated TSMOM (fee_edge), ER-filtered majors basket (overfit). Same mechanism, same overfit death cross-asset.

Backtest gate: - metrics unreliable: 3 trades / 2437 data days is too little evidence (need >=5 trades over >=10 days) — increase signal frequency or widen the window

Benjamini-Hochberg at q=0.10 over 231 programme candidates keeps 3. A candidate that does not survive here is not distinguishable from the programme's own noise, however good its individual statistics look.