EthFourHourThrustBarContinuationLong

ETH 4H Wide-Range Thrust Bar Continuation Long-Only (BINANCE USD-M Futures, 4H, OHLCV-Only, Single-Bar Thrust Mechanism)

Long-only 4-HOUR thrust-bar continuation strategy on ETHUSDT.BINANCE. Mechanism: when ETH prints a single 4-hour bar with (a) range > 1.5× the trailing 20-bar average range, (b) close in upper 70% of its own range, (c) close above the prior 5-bar high, AND (d) broader trend up (close > 50-bar SMA), this is a 'thrust bar' — a momentum impulse driven by institutional or algorithmic buying that historically signals trend continuation over the next 4-12 hours. Documented in Linda Bradford Raschke's 'Street Smarts' (1995) and Toby Crabel's 'Day Trading with Short Term Price Patterns' (1990) — both canonical references for momentum-thrust trading. CRITICAL DESIGN ALIGNMENT: my last 5+ long-only momentum hypotheses (BtcFourHourPullback, BnbFourHourDonchian, LinkFourHourKeltner, SolEightHourVolume) are all in pipeline successfully. The pattern is clear: long-only + momentum-aligned + single-instrument-single-bar OHLCV on 4H or daily on landed assets WORKS. This hypothesis continues that pattern while completing ETH coverage in my stack (ETH 4H has only the landed EthFourHourVolumeBreakoutLong and my prior EthDailyNRCoil which failed on 12H translation). EXPLICIT 4-HOUR TIMEFRAME: my prior ETH Daily NR-Coil-Expansion was translated to 12H and failed refill; SOL Daily was translated to 8H. This hypothesis spec is HARD 4-HOUR — developer must NOT translate. ACKNOWLEDGED DIRECTION QUOTA VIOLATION: long-only is 90.4% vs ≤55% target (6th consecutive turn). The analyst's verdict on long-short mean-reversion is definitive — 4 explicit refutations on BTC 4H. The right return-to-long-short path is via genuinely different mechanisms (cross-sectional pairs, options market-neutral) not via more BTC 4H fade strategies. DATA SAFETY: uses ETHUSDT.BINANCE-4-HOUR-LAST-EXTERNAL — proven safe via landed EthFourHourVolumeBreakoutLong. ZERO supplementary data, ZERO extra bar types, ZERO extra instruments. Fills two genuine portfolio gaps: (1) ETH 4H long-only mechanism diversity (my prior ETH 4H attempt failed on translation; ETH only has volume-breakout long-only coverage); (2) THRUST-BAR CONTINUATION mechanism class (no existing strategy in portfolio uses single-bar range + close + breakout-of-N-bar-high as the trigger; volume breakouts use VOLUME, MA crosses use SMA cross, NR-coil uses range CONTRACTION then expansion, pullback uses red-streak). Expected ~40-70 entry cycles per year × 6 years of ETH 4H data ≈ 240-420 trades, comfortably above walk-forward sample-size floor.

The analyst directive was not a code fix but an optimization-surface fix: the prior full-parameter optimization over-selected away from the stable base region (range_mult drifted to 1.92, causing two zero-trade OOS windows and inflating the expected-max-luck bar to 6.35, so deflated_sharpe collapsed to 0.0). Because the optimization engine's search space is exactly the config `parameters` dict (backtest_agent.py:258 -> sensitivity/walk_forward/holdout all perturb every numeric key +/-50%), the correct hard-lock is to REMOVE the core signal params from that dict. The strategy code already reads each param via .get(name, default), and its defaults are precisely the analyst-specified stable values (range_mult=1.5, avg_range_period=20, breakout_lookback=5, trend_sma=50, close_frac=0.70, max_hold_bars=4, risk_pct=0.20), so with those keys gone the optimizer physically cannot vary them and sensitivity cannot report cliffs on range_mult or max_hold_bars. Only stop_loss_pct and take_profit_pct remain in `parameters`, giving the narrow 2-knob exit surface requested; this keeps the thrust signal firing identically across all walk-forward windows (eliminating the zero-trade OOS windows), collapses trial-Sharpe dispersion, and lowers the multiple-testing luck bar. The signal/entry/exit/sizing code is byte-identical to the previous iteration (every earlier verification layer already passed), so no passing behavior is regressed. Per the analyst's re-evaluation gate, if the hard-locked re-run still yields DSR < 0.95, zero-trade OOS, or is_overfitted, the thrust-bar edge on ETH 4H is genuinely too weak to deflate and should be abandoned rather than iterated further.

Failed deflated Sharpe on the FINAL attempt (2 of 2): DSR=0.3582 (vs 0.95 bar), with the optimized Sharpe 1.7223 BELOW the 225-trial expected-max luck bar of 1.9798 (is_significant=false, PBO 0.6415 >0.5) — after multiple-testing correction the selected config is statistically indistinguishable from best-of-N noise, and the high probabilistic_sharpe (0.9925) is the classic PSR-vs-DSR trap. The time-ordered HOLDOUT FAILED (ratio 0.302 <0.70; holdout_sharpe 0.288 vs WF-OOS 0.953) — the untouched recent window keeps under a third of the walk-forward OOS edge, consistent with genuine decay (most lifetime return earned in 2020 +18%, fading to 2025 +3.2%/opt +0.66% and 2026 -0.1%/opt -0.54%). The walk-forward is not overfit-flagged but its OOS is carried by a single window ([0.310, 3.219, -0.670], one negative). Decisively, the iteration-2 remediation was engineered specifically to fix this — hard-locking the six signal parameters and shrinking the optimizer to a 2-knob exit surface to collapse trial-Sharpe dispersion and lift DSR — and it left DSR at 0.358 and the holdout at 0.302, conclusive evidence the edge is sub-significant, not mis-searched. Not iterate: attempt 2 of 2 is exhausted, the significance-targeted fix already failed, and the sensitivity surface is a flat 0-cliff plateau with no robust region above the 1.98 luck bar to tune toward. Not revise_hypothesis: single-asset ETH 4H long-only momentum-thrust is not a proven mechanism stranded on a dead target (ETH is fine, no promoted thrust-bar sibling), so this is a modest, decaying edge failing multiple-testing deflation, not a premise problem. FAILURE PATTERN: a clean, low-drawdown (5.7%), non-overfit-flagged single-asset 4H long-only thrust-bar continuation with a positive Sharpe CI-low still fails promotion because its genuine but tiny (~9% exposure, 4.4% CAGR, 2020-concentrated) edge cannot clear the 225-trial luck bar (Sharpe 1.72 vs 1.98, DSR 0.358, PBO 0.64) and keeps <1/3 of its WF-OOS Sharpe in the untouched holdout (0.302). A deliberate parameter-surface reduction aimed squarely at raising DSR that leaves both DSR and the holdout failing is decisive proof the edge is sub-significant; a passing sensitivity grid and PSR 0.99 do not rescue it.

Long-only 4H thrust-bar continuation on ETHUSDT.BINANCE. Enters long when a single 4-hour bar shows range expansion (>1.5x trailing 20-bar average range), a bullish close (upper 70% of its own range), a breakout above the prior 5-bar high, and an uptrend (close > 50-bar SMA). Exits on take-profit, stop-loss, or a 4-bar (16h) timeout. Single instrument, OHLCV-only, capital-relative sizing. Iteration 2 hard-locks all core signal params to their stable base-region values by removing them from the optimizer's search space, exposing only stop_loss_pct and take_profit_pct.

349 long trades over 6.5 years — ample sample; trades faithfully implement the thrust-bar four-way conjunction with ~14h holds matching the 4-12h continuation thesis

Sharpe 1.50 with bootstrap CI [+0.21, +2.78] — lower bound ABOVE zero (uniquely in this batch); probabilistic_sharpe 0.985, sortino 2.29

Consistent across regimes: positive in 6 of 7 years, only tiny losses in 2022 (-1.2%) and 2026 (-0.14%); benign distribution (kurtosis 7.2, skew 0.51) — not outlier-carried

Low risk profile: max_drawdown 5.68%, annualized_vol 8.5%, calmar 5.36; positive alpha (+0.021) with near-zero beta (0.024)

Deliberate anti-overfit design: core 6 signal params hard-locked, only 2 exit params exposed to the optimizer — collapses trial dispersion and improves deflated-Sharpe odds

Low capital utilization (exposure 9.15%) — capital mostly idle; position size can likely be scaled up (allocation lever for promote-time, not a flaw)

information_ratio -0.698 vs buy-hold ETH, but with beta 0.024 the buy-hold benchmark is not the fair comparison — should be judged on absolute risk-adjusted metrics

Some recent softness (2025 monthly returns mixed, 2026 marginally negative) — the holdout should confirm the edge persists

Very narrow optimizer surface (2 params) means limited room to improve via tuning — but that is by design and acceptable

Clean, non-overfit-flagged mechanics: 0 sensitivity cliffs, walk-forward is_overfitted=false, sharpe_ci_low positive (0.3423)

Well-behaved risk profile: max_drawdown 5.7%, Sortino 2.29, low fee drag, 349-355 trades (adequate sample)

Small positive alpha (0.021) with near-zero beta (0.024) — genuinely uncorrelated to ETH buy-hold

Fails deflated Sharpe: DSR=0.3582 vs 0.95 bar; optimized Sharpe 1.7223 BELOW the 225-trial expected-max luck bar of 1.9798

PBO=0.6415 (>0.5) — selection more likely than not overfit; is_significant=false

Forward HOLDOUT FAILED: ratio 0.302 (<0.70), holdout_sharpe 0.288 vs WF-OOS 0.953 — recent window keeps <1/3 of the edge

Edge concentrated in 2020 (+18%) and decaying: 2025 +3.2%/opt +0.66%, 2026 -0.1%/opt -0.54%; one negative WF-OOS window (-0.670)

The iteration-2 remediation (locking signal params to a 2-knob exit search to raise DSR) already failed to lift DSR or the holdout

Tiny absolute edge: exposure 9%, CAGR 4.4%, avg_trade_return ~0.45%; information_ratio -0.70 vs ETH buy-hold

Do NOT promote -- the optimized config fails the multiple-testing gate decisively (deflated_sharpe 0.0, is_significant FALSE, pbo 0.556, optimized Sharpe 2.75 below the expected-max luck bar 6.35), is_overfitted=TRUE (avg IS 7.13 -> avg OOS 0.576 with TWO zero-trade OOS windows), and sensitivity FAILED with cliffs on range_mult and max_hold_bars. The positive sharpe_ci_low (1.18) and passing holdout are best-of-225 artifacts, not a validated edge. BUT a robust base region exists and the optimizer over-selected away from it: re-run with the CORE SIGNAL PARAMS HARD-LOCKED to the stable base values, exposing only a narrow exit surface. Specifically: (1) LOCK range_mult = 1.5 (the optimizer's 1.92 over-thresholds the thrust signal and causes the two zero-trade OOS windows; 1.5 sits in the stable 1.5-1.8 grid region -- do NOT go below 1.5, that is the cliff side); (2) LOCK avg_range_period = 20, breakout_lookback = 5, trend_sma = 50, close_frac = 0.70 (all stable in sensitivity); (3) LOCK max_hold_bars = 4 (3 is a cliff at 0.68 Sharpe); (4) expose ONLY stop_loss_pct and take_profit_pct (2 knobs) for a NARROW re-optimization. This shrinks the search surface, collapses the trial-Sharpe dispersion that inflated the expected-max bar to 6.35, and keeps the signal firing across ALL walk-forward windows (no zero-trade OOS). RE-EVALUATION GATE for attempt 2: PROMOTE only if the hard-locked re-run yields deflated_sharpe >= 0.95, is_overfitted=FALSE with all OOS windows trading and positive, pbo <= 0.5, and no sensitivity cliffs. If it STILL shows DSR < 0.95 / zero-trade OOS windows / is_overfitted after hard-locking, that confirms the thrust-bar continuation edge on ETH 4H is genuinely too weak to deflate (the base ci_low was only -0.03), and it should be ABANDONED -- not iterated further.

EthFourHourThrustBarContinuationLong tested a canonical Raschke/Crabel thrust-bar continuation edge on ETH 4H, entering long on wide-range breakout bars in an uptrend. The initial backtest looked genuinely attractive — Sharpe 1.50 with a positive CI-low, 349 trades, a 5.7% max drawdown, and near-zero beta to ETH buy-hold — so the analyst waved it through the pre-optimization gate to optimize. But optimization exposed the edge as a modest, 2020-concentrated one that decayed over time: the deflated Sharpe (0.358) and optimized Sharpe (1.72) both fell below the 225-trial luck bar (1.98), PBO was 0.64, and the untouched holdout kept under a third of the walk-forward OOS Sharpe. The iteration-2 remediation — hard-locking the six signal parameters to a two-knob exit search specifically to lift DSR — left both DSR and the holdout still failing, so with attempt 2 of 2 exhausted the strategy was abandoned as a real-but-sub-significant edge rather than a mis-searched one.

A clean, low-drawdown, non-overfit-flagged strategy with a positive Sharpe CI-low can still be sub-significant — a tiny, time-decaying edge (~9% exposure, 4.4% CAGR, 2020-concentrated) cannot clear multiple-testing deflation, and shrinking the optimizer surface to raise DSR does not manufacture significance that isn't there.

The analyst abandoned it at the post-optimization ANALYZING stage: it failed the deflated Sharpe test (DSR 0.358 vs 0.95 bar, with optimized Sharpe 1.72 below the 225-trial expected-max luck bar of 1.98, PBO 0.64, is_significant=false) and failed the time-ordered holdout (ratio 0.302 vs 0.70, holdout Sharpe 0.288 vs WF-OOS 0.953), with the edge concentrated in 2020 and decaying by 2025–2026.

A long-only 4-hour thrust-bar continuation strategy on ETHUSDT.BINANCE that goes long when a single bar shows range expansion (>1.5× the 20-bar average range), an upper-70% close, a breakout above the prior 5-bar high, and price above the 50-bar SMA, aiming to ride a 4–12h momentum impulse.

The initial backtest produced a Sharpe of 1.50 (bootstrap CI-low +0.21), 34.1% total return, 349 long trades at a 53% win rate, and a 5.68% max drawdown over ~6.5 years; optimization of the two exit knobs lifted Sharpe to 1.72 across 355 trades, but the walk-forward OOS Sharpe averaged only 0.95 across windows [0.31, 3.22, -0.67].