EthEfficiencyGatedTimeSeriesMomentumLS

ETH USD-M Efficiency-Gated Time-Series Momentum, Long-Short (Single-Instrument BINANCE Perp, 4h Bars, Trade the N-Bar Trend ONLY When Kaufman Efficiency Ratio Confirms the Move Is SMOOTH/Directional — Skip Choppy Trends Where Whipsaw Kills Trend PnL — Volatility-Targeted Sizing + ATR Trailing Exit, 3-Parameter)

A LONG-SHORT, single-instrument directional strategy on ETHUSDT.BINANCE (USD-M perp, 4h bars) that is deliberately NOT the plain moving-average-crossover trend that keeps dying as 'other', and NOT any mean-reversion (the fade families are dead). The mechanism directly targets the ONE reason trend-following fails in crypto — whipsaw in choppy regimes — by gating every entry on the Kaufman Efficiency Ratio (ER = |close_t - close_{t-N}| / Σ|close_i - close_{i-1}| over the last N bars). ER≈1 means a clean directional move; ER≈0 means the same net move happened via noisy back-and-forth. We only take the N-bar trend when ER exceeds a threshold (the trend is genuinely EFFICIENT), and we sit flat when ER is low (choppy) — the exact regime where crossover trend bleeds. Direction = sign of the N-bar return: efficient up-move → LONG, efficient down-move → SHORT (shorts make returns regime-distributed and cap bear-market drawdown, unlike the long-only momentum baskets that all died on >50% drawdown / single-regime leverage). Position sizing is VOLATILITY-TARGETED (constant fractional risk via ATR) so no single trade or regime can produce the leverage-not-edge blowup that killed the vol-scaled 15-name basket. Pure OHLCV — no supplementary-data or cross-venue dependency, so no data_unavailable risk. This is NOT a rank/rotation basket (L12/L24 — single instrument, no cross-sectional sort), NOT carry (dead outside BTC), NOT options (L23), NOT liquidation-dependent (L3), NOT funding-directional (L13). 3 parameters: ER lookback N, ER entry threshold, ATR trail multiple.

Implements the hypothesis directly with exactly the 3 named core parameters (ER lookback N=20, ER entry threshold=0.30, ATR trail multiple=3.0) plus standard risk knobs. The efficiency gate is the mechanism: it trades trend only when the move is genuinely EFFICIENT and skips choppy trends, the exact regime that kills crossover trend-following in crypto. Long+short makes returns regime-distributed and caps bear-market drawdown (not a long-only momentum basket). Volatility-targeted ATR sizing gives constant fractional risk (L15 — per-trade risk = stop_distance x size ~= 1% of equity, and gross notional is capped independently at 1x equity), avoiding the compounding/leverage blowups that killed prior vol-scaled baskets. calculate_signal returns the RAW continuous ER*sign value every bar (varies bar-to-bar, natural units, not gated-to-zero); should_enter thresholds it against a matching-unit threshold, satisfying the continuous-signal guidance. Zero-trade risk (L21/L25) was explicitly checked on real ETH 4h data: at N=20 / threshold=0.30 the entry gate is in-signal on ~34% of bars (~744 of 2170 in a 365-day sandbox window), so >=1 trade is guaranteed with wide margin, and warmup (22 bars) is a negligible fraction of the window. All helper methods are instance methods (no @staticmethod) so they survive the Layer-2 verification proxy's MethodType re-binding. Single-instrument pure-OHLCV: NOT a rank/rotation basket (L12/L24), NOT carry (L1/L13), NOT options (L23), NOT liquidation-dependent (L3), NOT mean-reversion, and no data_unavailable risk. Venue is futures (BINANCE) because the strategy goes short; leverage stays 1.0 and is unreferenced in sizing so the leverage_set_but_unused gate is not triggered.

failed deflated Sharpe: DSR=0.416, expected-max=0.834 over 225 trials — the selected Sharpe (IS ~1.22) is indistinguishable from best-of-N noise, and the realized out-of-sample Sharpe is only 0.135. This is a decisive overfit, not a tuning problem: the walk-forward flags is_overfitted=true (IS avg 1.058 vs OOS avg 0.135) with OOS windows [0.074, 0.824, -0.492] (one negative — the edge lives in a single window, not a stable signal), and PBO 0.68 > 0.5 means the parameter selection is more likely than not overfit. Both are unwaivable hard gates, so promote is impossible. Even the selected config misses the promotion floors (Sharpe 0.768 < 1.5, win_rate 0.40 < 0.45, max_drawdown 16.1% > 10%, is_significant=false, sharpe_ci_low 0.106). Iteration cannot rescue it: the only non-cliff/robust parameter region sits at Sharpe ~0.5 with a CI that straddles zero — itself unpromotable — so there is no robust region for a second optimization pass to tune toward; it would only re-select noise from the same weak edge. This is the classic overfit signature (high IS, collapsed OOS), which per policy is abandon rather than revise, and the calibration confirms it: the near-identical Kaufman-ER majors basket (similarity 0.72), ETH Donchian trend, and BTC 200-MA trend followers were all abandoned as overfit. The ER-gated ETH trend edge is genuinely too weak to survive multiple-testing correction.

Long-short, single-instrument time-series momentum on ETHUSDT.BINANCE (USD-M perp, 4h bars), gated by the Kaufman Efficiency Ratio. Each bar it computes ER = |close_t - close_{t-N}| / sum|close_i - close_{i-1}| over the last N bars (ER~1 = a smooth directional move, ER~0 = the same net move achieved via noisy back-and-forth). The signal is ER * sign(N-bar return), a continuous value in [-1,1]. It enters the N-bar trend ONLY when |signal| exceeds an efficiency threshold (0.30) — LONG on an efficient up-move, SHORT on an efficient down-move — and sits flat when ER is low (the choppy regime where crossover trend bleeds to whipsaw). Sizing is volatility-targeted: constant fractional risk (1% of equity) divided by an ATR-based stop distance (atr_trail_mult x ATR), so larger volatility automatically shrinks the position and no single trade/regime can produce a leverage-not-edge blowup; a hard notional cap (1x equity) and no leverage bound gross exposure. Exit is an ATR trailing stop (trails the favourable extreme by atr_trail_mult x ATR) plus an efficient-reversal exit that flattens when the opposite efficient trend appears. Pure OHLCV — no supplementary or cross-venue data.

Analyst to judge on full history; the contained drawdown and positive-skew profile make this a reasonable optimize/iterate candidate rather than an obvious abandon, but the negative 1-year sandbox and zero-straddling Sharpe warrant caution.

Edge outcome for the analyst (not a code defect): the sandbox is modestly negative — total_return -8.7%, Sharpe -0.67 (CI [-2.27, +0.90], straddles zero), profit_factor 0.74, win_rate 0.36 over 59 trades, max_drawdown 19.5%. Unlike the session's blown-up strategies this has a HEALTHY trend-following profile (return_skew +0.52, avg_win $1173 > avg_loss $878, tail_ratio 1.14, contained 19.5% drawdown) — the efficiency gate and ATR vol-targeting are capping drawdown as designed; the shortfall is a sub-50% hit rate that didn't clear costs in this one-year window. A measurable edge question for BACKTEST_REVIEW, not recodable.

Confirm the full-history backtest yields >= ~100 trades before drawing firm conclusions.

59 trades over 364 days is borderline in isolation (below the ~100 single-window floor) but at ~59/year extrapolates comfortably above ~100 on the full multi-year backtest, so measurability is not a blocker here (unlike an inherently low-frequency signal).

avg_trade_return_pct 0.79% — well above the 0.15% futures fee floor; genuine edge net of costs

profit_factor 1.30 with positive skew 1.53 and tail_ratio 1.30 — healthy trend-following payoff structure (small losses, large wins)

max_drawdown 18.0% (well under 50%), calmar 4.44 — risk is controlled, no leverage-blowup signature

Regime-distributed returns: positive in 2020/2021/2023/2024/2025 and only -6.7% in the 2022 bear — the long-short design delivered the bear-market protection the long-only baskets lacked

Balanced book (170 long / 164 short), low beta 0.05, positive alpha 0.053 — differentiated, near-market-neutral return stream

Well-sampled (334 trades) so parameter sensitivity will be statistically meaningful

Sharpe 0.637 is modest and its CI low (-0.03) just grazes zero — the edge needs strengthening in optimization to clear promotion

information_ratio -0.66 vs buy-hold ETH is negative (though buy-hold is not the fair benchmark for a low-beta long-short)

2026 shows a sustained drawdown stretch (rolling Sharpe negative) — watch OOS robustness in the recent window during walk-forward

Clean risk profile: optimized max_drawdown 16.1%, no liquidation, positive skew 1.38, tail_ratio 1.34 — a genuine trend-following payoff shape

avg_trade_return_pct 1.42% clears the fee floor comfortably; capacity ~$42M is deployable

Sensitivity is smooth with zero cliffs — the code is sound and behaves as designed

Long-short design kept 2022 positive (+12.6% optimized), delivering the regime distribution the hypothesis promised

HARD FAIL: walk-forward is_overfitted=true — IS Sharpe 1.058 collapses to OOS 0.135, with OOS windows [0.074, 0.824, -0.492] (one negative, wildly inconsistent)

HARD FAIL: PBO 0.68 > 0.5 — parameter selection is more likely than not overfit

deflated_sharpe 0.416 << 0.95, is_significant=false, sharpe_ci_low 0.106 — the selected Sharpe is a best-of-225-trials artifact, not a validated edge

Even the selected config misses floors: Sharpe 0.768 < 1.5, win_rate 0.40 < 0.45, max_drawdown 16.1% > 10%

The only robust (non-cliff) parameter region sits at Sharpe ~0.5 — itself below promotion floors with a CI straddling zero; there is no promotable region to tune toward

Family calibration: near-identical ER/trend strategies (Kaufman-ER majors basket sim 0.72, ETH Donchian, BTC 200-MA) were ALL abandoned as overfit

This strategy was the strongest-looking member of the trend family this session: an efficiency-ratio gate that only traded smooth directional moves, plus vol-targeted sizing and a short side, delivered a genuinely healthy risk profile — 18% max drawdown, positive skew, a balanced near-market-neutral book, and the regime distribution (only -6.7% in 2022) that the long-only baskets never achieved. It cleared the pre-optimization gate with an 'optimize' verdict, but the 3-phase optimization exposed the same fatal flaw as its siblings: in-sample Sharpe of 1.058 collapsed to 0.135 out-of-sample with one negative window, PBO hit 0.68, and the deflated Sharpe of 0.416 marked the result as best-of-225-trials noise. Even the selected config missed every promotion floor, and the only robust parameter region sat at an unpromotable Sharpe ~0.5. The analyst abandoned it on its first iteration as a decisive overfit with no robust region to tune toward — the ER-gated ETH trend edge was real but simply too weak to survive multiple-testing correction.

An efficiency-ratio gate genuinely improved the trend payoff shape — controlled drawdown, positive skew, and real 2022 bear protection from the short side — but the ER-gated ETH trend edge was still too weak to survive multiple-testing correction, confirming (alongside the near-identical Kaufman-ER basket, ETH Donchian, and BTC 200-MA siblings) that this single-name trend class overfits rather than generalizes.

The analyst issued an abandon verdict on unwaivable overfit hard gates: walk-forward is_overfitted=true with OOS windows [0.074, 0.824, -0.492] (edge living in a single window), PBO 0.68 > 0.5, and a deflated Sharpe of 0.416 against an expected-max of 0.834 over 225 trials — meaning the selected Sharpe was best-of-N noise — while the only robust non-cliff parameter region sat at Sharpe ~0.5 with a CI straddling zero, leaving no promotable region to tune toward.

A long-short, single-instrument time-series momentum strategy on ETHUSDT.BINANCE (USD-M perp, 4h bars) that trades the N-bar trend only when the Kaufman Efficiency Ratio confirms the move is smooth/directional (sitting flat in choppy regimes), with volatility-targeted ATR sizing and an ATR trailing exit, and shorts to make returns regime-distributed.

The initial backtest was the healthiest of its family — +97% return over 334 balanced trades (170 long / 164 short), profit factor 1.30, avg_trade_return_pct 0.79%, positive skew 1.53, only 18.0% max drawdown, and just -6.7% in the 2022 bear — but a modest Sharpe of 0.637 (CI grazing zero); optimization lifted it to Sharpe 0.768 / +256% yet flagged is_overfitted=true (IS 1.058 → OOS 0.135), PBO 0.68, and deflated Sharpe 0.416.