EthBinanceDualTimeframeMomentumConfluenceLS

ETH Binance USD-M Dual-Timeframe Momentum Confluence, Long-Short (Single-Instrument ETHUSDT.BINANCE Perp, 4H Primary Momentum + 1D Confirm, Directional ONLY When Both Timeframes Agree / FLAT on Disagreement, ATR-Trailing Winners, 3-Parameter, MULTI-YEAR History)

A LONG-SHORT, single-instrument, pure-OHLCV application of this factory's ONLY optimization survivor — the BTC Dual-Timeframe Momentum Confluence (Sharpe ~1.99, in paper) — to the second-cleanest-trending asset (ETH) on the venue where the recipe demonstrably passes the robustness gauntlet (Binance USD-M, multi-year history). This is a deliberate, evidence-forced correction. This session I watched SIX single-name Hyperliquid momentum variants die, and the failures were NOT random — they clustered on two causes structurally tied to Hyperliquid's short (~2.5-year) history: (a) OVERFIT — ETH-HL pullback (PBO 0.73), SOL-HL Donchian (PBO 0.64, holdout Sharpe -1.67), funding basket (DSR 0.0001) — a 225-trial optimizer cannot avoid noise-fitting when the OOS window is barely a year; and (b) UNMEASURABLE SAMPLE — LINK-HL TSMOM (~50 trades < 100 floor). Binance's 5+ years of ETH 4H history directly fixes BOTH: it yields 200+ trades (clears the measurability floor that killed the low-turnover variants) AND long walk-forward OOS windows spanning multiple regimes (the antidote to the PBO/holdout collapses that killed the HL ports). ETH is chosen because the surviving edge requires a CLEAN-trending asset — the fee-thin/negative results on noisier alts (SOL-gate PF 1.086, AVAX -0.28%/trade) show the edge does not survive on choppy names, and ETH is the least-noisy asset after BTC. It reuses the ALREADY-VALIDATED generic confluence signal verbatim (not a bespoke filter), minimizing the parameter surface that the multiple-testing correction punishes. It is materially different from the pipeline's BtcDualTimeframeMomentumConfluence (BTC, not ETH) and EthHyperliquidDualTimeframeMomentumConfluence (Hyperliquid's short/hostile history — this is Binance's multi-year history, the specific fix). Fills long_short direction (13.6% vs target). Single instrument, single venue — NO cross-venue schema-merge landmine, NO spot-lookup timing bug. Pure OHLCV, no supplementary feed. I acknowledge this adds to the over-represented BINANCE venue (71%); given 0.9% session survival and six consecutive HL deaths, I am prioritizing a demonstrable path through the robustness gates over the venue quota — the multi-year history is not available on any under-represented venue.

I ported the validated survivor's signal code verbatim rather than re-deriving an ETH variant, since that reuse is the hypothesis's anti-overfit argument. Note for the analyst: the validated strategy's 'dual timeframe' is two lookbacks on one 4H series (its own docstring calls the 42-bar leg the 'daily-scale confirm'), not a separate 1-DAY subscription — so despite the title saying '4H + 1D', verbatim reuse means single-feed, which also removes extras-resolution and alignment risk. Two implementation details I corrected for this instrument: min_notional is 20 USDT (ETHUSDT's actual USD-M minimum per the exchange reference, where the BTC original used 10 and the generic default is 5 — too low here, so orders could be exchange-rejected), and quantity precision is 3 against the instrument's reported size_precision of 4. I also locked the risk/execution settings the original exposed as parameters, which only shrinks the optimizer surface. I validated the port on real data before submitting: the catalog has 14,412 ETH 4H bars from 2019-12-31 with zero gaps, and at the validated defaults with 0.10% Binance round-trip fees it returns +116.9% over 6.6 years with a 24.0% max drawdown, 787 trades, PF 1.25 and +0.39% per trade. Both structural claims verify — 787 trades destroys the ~100 measurability floor that killed the LINK TSMOM, and across the ±20% band Phase-1 sensitivity probes, all nine (fast, slow) cells are positive (+95.8% to +169.4%, per-trade edge in a tight +0.33% to +0.42% range) with no cliff, which is a real robustness result and contrasts with the same recipe on Hyperliquid where slow_lookback below ~38 flipped negative. The important caveat is the per-year decomposition, which I think should drive the review more than the headline: 2020 +64.3% and 2024 +22.3% carry most of the total, while 2021 (-0.7%), 2022 (-1.5%), 2023 (-9.4%) and 2025 (+0.9%) are flat to negative, and the trailing 365 days is flat (+0.1%, +0.10% per trade, at or below the 0.15% futures floor). So the multi-year history genuinely delivers the long multi-regime walk-forward windows the hypothesis wanted, but those windows will contain real unprofitable stretches — the honest expectation is a positive-but-modest OOS Sharpe, not a reproduction of the parent's ~1.99, and the strategy should be judged on cross-window stability rather than on the 6.6-year total.

No promotable edge to optimize. Over the largest, cleanest sample the factory can produce — 719 trades across 6.6 multi-regime years — the strategy's full-sample Sharpe is a well-measured, non-significant 0.433 (CI [-0.17, 1.11] straddles zero) and profit_factor is 1.186 (< 1.2, a direct L9 abandon signal for OHLCV momentum). information_ratio -0.57 means it delivers negative risk-adjusted active value versus a meaningful buy-hold ETH benchmark, and the +173% total is regime-concentrated (most from 2020 +43% and 2024 +31%, with 2023/2025 negative) and leans on a handful of outlier days (kurtosis 8.3). avg_trade_return_pct 0.549% clears the fee floor, so this is not a fee-edge death — the problem is that the true edge is definitively measured and sits far below the 1.5 promotion floor. Optimization targets Sharpe against that floor, and no 3-parameter tuning can robustly manufacture a 1.5 Sharpe from a well-measured 0.43; any sub-window that shows it would be an overfit selection that PBO/deflated-Sharpe rejects — the exact death the ETH/SOL-HL dual-TF siblings just suffered (PBO 0.71-0.73). The developer's own caveat predicts a modest, unpromotable OOS. There is no code bug to iterate on. Abandon rather than spend 2 hours repeating the optimize→overfit-reject cycle.

Long-short dual-horizon momentum confluence on the ETHUSDT Binance USD-M perpetual, 4H bars — an asset port of the factory's validated BTC survivor with the signal code reused verbatim. Two momentum horizons are computed on the 4H close series: a fast ~2-day leg (12 bars) and a slow ~7-day daily-scale confirm (42 bars). It goes long only when both are positive, short only when both are negative, and stays flat whenever they disagree. Positions exit when the slow confirm leg flips against them or on a 5x-ATR trailing stop from the best price reached. Sizing risks 1% of equity across an initial 1.5x-ATR stop, capped at 50% of equity in gross notional, with a 20 USDT minimum order (ETHUSDT's actual exchange minimum). Leverage 1.0. Exactly three tunable parameters (fast_lookback, slow_lookback, trail_atr_mult); ATR period, initial stop, risk fraction, notional cap and quantity precision are locked constants.

Require the walk-forward OOS to be positive and stable ACROSS windows (not driven by 2020+2024 alone); expect modest OOS Sharpe, not ~1.99, and abandon if OOS is flat/negative once the two big trend years are outside the window.

REGIME-CONCENTRATED EDGE — the central risk, honestly disclosed. The full-history +116.9% / +0.39%-per-trade is carried by 2020 (+64.3%) and 2024 (+22.3%); four of seven years (2021 -0.7%, 2022 -1.5%, 2023 -9.4%, 2025 +0.9%) are flat-to-negative, and the trailing 365-day sandbox is NEGATIVE (-4.2%, avg_trade_return_pct -0.052%, PF 0.92 — below the fee floor). A legitimate trend-following profile (2024 and 2026 PF 1.47 show it still fires, so regime-dependence not decay), but OOS Sharpe will be modest and walk-forward windows will contain unprofitable stretches. Judge on cross-window STABILITY, not the headline total.

Do not read the single negative sandbox year as the verdict; judge fee viability and stability on the full 6.6-year walk-forward. The 787-trade sample clears the measurability floor by a wide margin.

Sandbox year negative, but the evaluation horizon is the full history. The 365-day sandbox (-4.2%, -0.052%/trade, PF 0.92) is one weak regime; the actual backtest runs on 6.6 years / 787 trades where it is +116.9%, PF 1.25, +0.39%/trade (above the 0.15% floor) with a FLAT parameter surface (all nine ±20% cells positive, NO cliff). Strongest measurability+robustness case of the batch — direct contrast with the LINK TSMOM I failed (~50 trades, grid-selected) and the BTC-HL port (thin +0.14%, cliff at 38).

Accept both (faithful verbatim port + deliberate quota trade-off for the multi-year-history fix); the PM should weigh the Binance concentration and correlation with the existing BTC/ETH dual-TF confluence strategies.

Two documented deviations, neither a mechanism change: (1) the title says '4H + 1D Confirm' but the confirm is a 42-bar (~7-day) momentum on the SAME 4H feed (the survivor's actual construction, single-feed), transparently disclosed; (2) adds to the over-represented BINANCE venue (71%), which the developer explicitly acknowledges, prioritizing a demonstrable path through the robustness gates (multi-year history unavailable on under-represented venues) over the quota.

Very large, measurable sample: 719 two-sided trades (360 long / 359 short) over 6.6 multi-regime years — a faithful port of the factory's one survivor recipe

avg_trade_return_pct 0.549% is ~3.6x the fee floor; positive in 5 of 7 years; flat parameter surface (no cliff)

Over the maximal sample the edge is well-measured and thin: Sharpe 0.433 with CI [-0.17, 1.11] straddling zero — not significant even in-sample

profit_factor 1.186 < 1.2 — direct L9 trigger for OHLCV momentum, a class optimization has never rescued

information_ratio -0.57 vs a meaningful buy-hold benchmark — negative risk-adjusted active value (worse than holding ETH)

Regime-concentrated: most of +173% from 2020 & 2024, negative 2023/2025, total leans on a few outlier days (kurtosis 8.3)

Same dual-TF momentum family whose HL siblings just failed post-opt on PBO 0.71-0.73 — and this raw edge starts weaker; optimizing a well-measured 0.43 Sharpe toward the 1.5 floor would be overfit selection

After watching six single-name Hyperliquid momentum variants die on overfit and unmeasurable samples, the developer deliberately ported the factory's lone survivor recipe onto ETH on Binance specifically to exploit 6.6 years of clean, multi-regime history as the antidote. The larger sample worked as diagnostics: 719 faithful two-sided trades produced a +173% headline but a precisely measured, non-significant Sharpe of 0.433, profit factor 1.186, and negative information ratio, with the gains concentrated in 2020 and 2024. Rather than an ambiguous overfit risk, the history revealed the edge was simply real-but-thin and sitting far below the 1.5 floor. With no code bug to iterate and no way to lift a well-measured 0.43 Sharpe without curve-fitting, the analyst abandoned it at the backtest-review gate instead of repeating the optimize-then-reject cycle.

More and cleaner history is a double-edged fix: it removed the small-sample and overfit ambiguity that killed the Hyperliquid siblings, but by measuring the edge precisely it proved the ETH momentum-confluence signal is genuinely thin (Sharpe ~0.43, PF <1.2) rather than merely unlucky — a well-measured sub-floor edge is not a candidate for optimization.

The analyst abandoned it at the pre-optimization BACKTEST_REVIEW gate — optimization never ran — because the edge was definitively measured at Sharpe ~0.43, far below the 1.5 promotion floor, with profit factor 1.186 tripping the OHLCV-momentum PF<1.2 abandon signal and negative active value versus the benchmark. No 3-parameter tuning could robustly manufacture a 1.5 Sharpe from a well-measured 0.43 without overfit selection that PBO/deflated-Sharpe would reject.

A long-short, single-instrument dual-timeframe momentum-confluence strategy that ported the factory's one optimization survivor (the BTC dual-TF confluence) verbatim onto ETHUSDT.BINANCE — going long only when a fast (~2-day) and slow (~7-day) 4H momentum leg both agreed up, short when both agreed down, flat on disagreement, trailing winners with an ATR stop — relying on Binance's 6.6-year history to fix the overfit and small-sample failures of prior Hyperliquid attempts.

Over the factory's largest clean sample — 719 two-sided trades (360 long / 359 short) across 6.6 multi-regime years — it returned +173% but with a well-measured, non-significant Sharpe of 0.433 (CI [-0.17, 1.11] straddling zero), profit factor 1.186, information ratio -0.57 versus buy-hold ETH, and 25.9% max drawdown. avg_trade_return_pct 0.549% cleared the fee floor, but returns were regime-concentrated (most from 2020 and 2024, with 2023 and 2025 negative, kurtosis 8.3).