EthSustainedMultiBarDriveContinuationLS1H

ETH Intraday Sustained Multi-Bar Drive Continuation, Long-Short (Binance USD-M, 1H, Persistence-Gated Momentum Burst on the Proven Intraday-Continuation Asset, Clamped Drive Threshold, Tight Initial Stop, Trailing-Stop Winners, Discrete Capital-Capped)

A LONG-SHORT, single-instrument, pure-OHLCV intraday CONTINUATION strategy on ETHUSDT.BINANCE USD-M perpetual using 1-HOUR bars, whose trigger is a SUSTAINED MULTI-BAR DRIVE (persistence) — a third distinct continuation trigger on ETH, the factory's single best and only multi-trigger-validated intraday-continuation asset (single-impulse ETH-1H promoted ~1.40; opening-range-breakout ETH-1H positive +80.8% and in optimization). It enters in the direction of a confirmed momentum burst (N consecutive same-direction 1H closes whose cumulative move exceeds a volatility threshold) and rides the continuation. The recent map is now decisive: the magnitude-outlier impulse trigger TRANSFERS across high-vol L1s (ETH and SOL both work), but other triggers are ASSET-SPECIFIC (ETH-ORB +80.8% vs SOL-ORB catastrophic -94% — SOL's high-beta intraday chop whipsaws time-anchored breaks), and BTC carries NO intraday-continuation edge at all (fee-bound-zero across every trigger/venue). This hypothesis targets ETH precisely because ETH — unlike BTC — HAS a strong, promoted intraday-continuation base edge, so the persistence-gated drive (which failed on BTC ONLY because BTC had no edge to extract, not because the trigger is bad) has a genuinely better prior here. The drive is a distinct, additive trade population from the single-outlier-bar impulse and the time-anchored ORB triggers already validated on ETH. It AVOIDS every dead/blocked path: NOT BTC (dead all triggers/venues — stop), NOT a SOL/alt port of an asset-specific trigger (SOL-ORB disconfirmed; AVAX/DOT impulse already pending), NOT a funding/basis/carry/cross-venue strategy (entire family fee-dominated/event-sparse, confirmed across BTC/ETH/SOL), NOT reversion (dead), NOT daily-basket refinements (correlated with promoted AbsoluteMomentum), NOT options (engine), NOT sub-1H (fee-dead), NOT HL intraday (data wall). It carries the SOL-sibling hardening to resist overfitting: drive_atr_mult CLAMPED, drive length and exhaustion filter FROZEN, <=4 tunables, flat-between-signals. Single-venue, deep ETH 1H history (validatable), engine-safe (discrete enter-once/exit-once, single position, hard-capped, no continuous rebalance/vol-scaling).

Iteration 2 addresses the two analyst fixes without touching the (already Layer-3-passing) signal/exit logic. (1) Impact/over-sizing: risk_frac cut from 0.01 to 0.004, and a new hard notional cap (max_notional_frac=0.20 of deployable capital) replaces the old equity*leverage cap. Previously notional could equal/exceed equity (avg_position_pct ~125%, impact_cost_pct ~94.5%, capacity ~$112k); capping at ~20% of equity drastically reduces modeled market impact so the ~$200/trade gross-of-impact edge can be evaluated net, and prevents any single position from exceeding equity. (2) Realized-P&L / open-at-end: the trailing+initial ATR stop in should_exit is unchanged and confirmed to close positions on stop hits; the prior +380% headline was one un-exited MTM position at horizon end (end_unrealized_pct=362.6). Smaller size shrinks that MTM distortion and lets the strategy be judged on realized returns/Sharpe. Leverage stays 1.0 and is not relied upon for sizing (cap uses leverage*frac but leverage=1), so no leverage_set_but_unused issue. Tunables remain <=4 (drive_atr_mult clamped, stop_atr_mult, trail_atr_mult, risk_frac); structural constants frozen.

failed deflated Sharpe: DSR=0.0, expected-max=6.22 over 225 trials — selected Sharpe indistinguishable from best-of-N noise. Walk-forward is_overfitted=TRUE (IS 6.40 → OOS 0.00 across all windows) and holdout FAILED (Sharpe -0.383 on 3 trades). The optimizer curve-fit to just 15 trades in 6 years (drive_len=1, atr_period=3), a fabricated in-sample edge with no generalizable parameter region. This is the OHLCV single-instrument directional-continuation family (0/213 survival, L9/L41/L63); the underlying ETH intraday-drive premise showed no risk-adjusted edge before optimization (Sharpe 0.307, IR -0.67, CI straddling zero) and only degrades under deflation. Overfit + failed holdout on the same mechanism whose BTC sibling was risk_reject (-100%) and ETH-4H sibling was overfit — reframing the same clone onto another target would just overfit again. No parameter change can conjure an edge the full 225-trial sweep failed to find.

Long-short intraday continuation on ETHUSDT.BINANCE USD-M perpetual (1H bars). Trigger is a sustained multi-bar drive: drive_len consecutive same-direction 1H closes whose cumulative close-to-close move (in ATR units) exceeds a clamped threshold and is below an exhaustion ceiling. Enters in the drive's direction, protects with a tight initial ATR stop, and rides continuation with a wider trailing ATR stop. Flat between signals, single position, enter-once/exit-once.

Mechanism matches the hypothesis: long-short continuation with balanced 814 long / 822 short trades, ~17h holds, all signals cleanly submitted (no size/min-notional drops).

Healthy trade count (1636 over ~5.4yr, ~300/yr) — enough sample for meaningful parameter sensitivity without being fee-churning tick frequency.

Net-positive after real fees AND folded market impact (total_return 48.2%, max_dd 14.7%, recovery_factor 2.81); the asset has a promoted intraday-continuation sibling, so the prior is reasonable.

A clear optimization lever exists: drive_atr_mult is at 1.6 within a clamped 1.0-3.0 band; raising it cuts trade count and directly attacks both the thin PF and the high impact drag.

Thin edge: profit_factor 1.089 and Sharpe 0.307 with the Sharpe CI straddling 0 (-0.355 to 0.97) — base is barely distinguishable from no-skill.

impact_cost_pct 27.9% — modeled market impact eats ~28% of gross PnL (above the 25% flag) even with max_notional_frac=0.2; the strategy is sized aggressively relative to the edge.

Regime decay: profits concentrated in 2020-2022; 2023 ~flat, 2025 -8.5%. The recent (holdout-relevant) regime is weak, raising the risk the walk-forward/holdout gates reject it.

win_rate 36% with PF only 1.089 means the result leans on a few large winners (largest_win 8314 vs avg_win 901) — fragile to the trailing-stop tail.

Clean, engine-safe implementation (discrete enter-once/exit-once, notional-capped, ≤4 tunables, impact controlled at 7.7-28%)

Optimized per-trade edge 0.725% clears the fee floor, so this is not a fee-viability death

Sensitivity phase showed no cliff parameters

Walk-forward is_overfitted=TRUE: IS Sharpe 6.40 collapses to OOS 0.00 across all three windows

Holdout FAILED: Sharpe -0.383 on only 3 trades — negative forward edge on unseen data

Deflated Sharpe 0.0, is_significant=false, PSR 0.70, Sharpe CI [-0.48, 0.66] straddles zero — selected Sharpe indistinguishable from best-of-N noise over 225 trials

Optimizer curve-fit to 15 trades / 6 years (drive_len=1, atr_period=3) — a fabricated edge, not a robust region

Initial backtest already marginal (Sharpe 0.307, PF 1.089, information_ratio -0.67); a single-instrument pure-OHLCV directional continuation clone (0/213 family survival)

Iteration 2 with sibling precedent: BTC variant risk_reject (-100%), ETH-4H variant overfit

Do NOT optimize yet — the tested configuration is over-sized and the profitability is a measurement artifact. Two specific fixes before re-backtesting: 1. CUT POSITION SIZE / FIX THE IMPACT-DESTROYED CAPACITY. impact_cost_pct=94.5% means market impact consumes ~94% of gross PnL; capacity_usd is only $112k and avg_position_pct=125% (notional exceeds equity). The risk-based sizing (risk_frac=0.01 over a tight 1.5-ATR stop) is producing untradeable size on ETH-1H. Drastically reduce risk_frac (e.g. 0.0025-0.005) and/or add a hard notional cap well below equity so modeled impact falls to a small fraction of gross (target impact_cost_pct < ~15-20%). The gross-of-impact edge is ~$200/trade — if impact is brought down, a real net edge may survive. Report impact_cost_pct and capacity_usd after the change. 2. EXIT THE FINAL POSITION / EVALUATE ON REALIZED P&L. end_unrealized_pct=362.6 means the +380% headline is almost entirely one un-exited open position at backtest end (realized return is only ~18% over 6 years; the +47% on 2026-06-19 is the open mark). Verify the trailing/initial-stop exit logic actually closes positions and that the strategy is flat at horizon end, then judge on REALIZED returns and the realized Sharpe (currently -0.34) — not the MTM headline. After both fixes, re-backtest. If a genuine net positive risk-adjusted edge emerges (realized Sharpe > 0.5, net PF > 1.2, impact a small share of gross, capacity not trivially tiny), proceed to optimization. If the realized Sharpe stays negative / PF ~1.0 once impact is controlled, the continuation signal has no net edge here and it should be abandoned. Also flag the regime dependence (2025 -68%) for the optimizer to stress via walk-forward OOS.

This strategy proposed a sustained multi-bar drive as a third continuation trigger on ETH 1H perps, betting that ETH's promoted intraday-continuation pedigree gave the persistence mechanism a better prior than its failed BTC sibling. The base backtest was net positive over 1,636 trades (48.2% return) but thin — Sharpe 0.307 with a CI straddling zero — so it advanced to the full 3-phase optimization. There the walk-forward was not overfit and the best config showed a 2.79 Sharpe, but the analyst abandoned it: the deflated Sharpe of 0.029 placed it below the best-of-225 luck bar, the holdout retained only ~16% of the edge, and sensitivity revealed four cliffs on the core trigger knobs. Decisively, the optimizer collapsed the frozen drive_len from 3 to 1, reverting toward the already-promoted single-impulse trigger and proving the 'drive' premise added nothing — a selection-noise artifact rather than a deployable, distinct edge.

A 'distinct additive trigger' layered onto an asset with a real promoted edge can still fail multiple-testing deflation — a non-overfit walk-forward and positive selection-CI are masked by a deflated Sharpe below the luck bar, a failed holdout, and parameter cliffs, especially when the optimizer collapses the defining parameter and proves the new mechanism adds nothing beyond the existing one.

It passed the backtest-review gate ('optimize') but the post-optimization analyst issued an 'abandon' verdict: the optimized Sharpe of 2.79 fell below the best-of-225 luck bar of 4.05 (deflated Sharpe 0.029, not significant), the holdout failed (ratio 0.157), sensitivity showed 4 cliffs on the trigger-defining parameters, and the optimizer collapsed the frozen drive_len from 3 to 1 — abandoning the 'sustained drive' premise and reverting toward the already-promoted single-impulse trigger.

A long-short, single-instrument intraday continuation strategy on ETHUSDT.BINANCE 1H perps whose trigger was a sustained multi-bar drive — N consecutive same-direction closes whose ATR-normalized cumulative move cleared a clamped threshold — entering in the drive's direction with a tight initial stop and trailing stop, pitched as a third distinct continuation trigger on ETH, the factory's validated intraday-continuation asset.

The base backtest fired 1,636 trades (814 long / 822 short) and was net positive after fees and impact (total return 48.2%, max drawdown 14.7%, recovery factor 2.81) but with a thin edge (Sharpe 0.307 with CI straddling 0, profit factor 1.089, 36% win rate, impact 27.9% of gross). Optimization's walk-forward was not overfit (avg OOS Sharpe 2.28) and the best config reached Sharpe 2.79, but its deflated Sharpe was only 0.0291 and the holdout kept just ~16% of the edge (ratio 0.157).