EthFourHourConvictionCompositeLS

ETH 4H Multi-Factor Conviction Composite — Long-Short, Diversified 4H Factor Composite Gated by the Daily Trend (Flat on Disagreement), Volatility-Targeted, Trailing-Stop Winners (ETHUSDT.BINANCE, 4H+Daily, 3-Parameter)

A LONG-SHORT, single-instrument directional strategy on ETHUSDT.BINANCE (USD-M perpetual), applying the diversified conviction-composite construction to the SECOND edge-bearing asset. EVIDENCE: on BTC every single bars-factor lands Sharpe ~0.5 (DSR-failing), and my three BTC composites (queued) test whether diversification lifts them; but ETH is the strongest non-BTC asset — its plain 4H+1D momentum confluence reached DSR 0.76 (Sharpe ~1.30), a genuine near-miss materially above the ~0.5 single-factors and the closest any non-BTC candidate came to promotion. The untried question: does adding DIVERSIFICATION (equal-weight composite of orthogonal 4H factors), SELECTIVITY (composite AND daily-trend agreement, flat otherwise), and RISK DISCIPLINE (vol-targeting + trailing winners) — the survivor's three structural levers — push ETH's near-miss OVER the 0.95 DSR bar? ETH is the natural asset: it has demonstrated edge at exactly this 4H+1D horizon, unlike the choppy alts (DOT/ATOM 0.2) whose confluence had no edge to diversify. NOT the confluence template sprayed on a ticker (L56 — the fast signal is a multi-factor composite, not a single EMA; and ETH is chosen because it HAS demonstrated 4H edge, not blindly), NOT a single-signal probe (all ~0.5), NOT the three queued BTC composites (different asset), NOT the regime-decayed convex family, NOT cross-sectional (L52), NOT reversion (L53), NOT microstructure/carry/basis/options (dead/infra). LOW-DOF: 4 pre-registered equal-weight 4H factors (fast momentum, close-location pressure, volume-flow, range-position), only 3 tuned params (composite threshold, daily-trend lookback, trailing-ATR multiple). Fills long-short (13.5% vs 86% long-only). HONEST framing: ETH single-factors likely also ~0.5, so the composite must clear via diversification of their idiosyncratic error PLUS ETH's demonstrated horizon edge; if it lands ~0.7–0.9 it is another near-miss, which alongside the BTC composites decisively answers whether the survivor's edge is structure-general or config-specific. Risk: ~1.5% equity/trade anchored to the trailing stop; vol-targeted size capped 20% of equity*leverage; 2x leverage cap (reads self.config.leverage).

Implements the hypothesis directly: the ETH 4H near-miss (Sharpe ~1.30, DSR 0.76) rebuilt with the survivor's three structural levers. DIVERSIFICATION = the equal-weight 4-factor z-composite (nothing fitted, weights fixed at 1/4). SELECTIVITY = trade only on composite/daily-trend agreement, flat otherwise. RISK DISCIPLINE = vol-targeted sizing, ~1.5% equity risked per trade anchored to the initial stop, and trailing the winners. Degrees of freedom held to 3 as specified, with _param_bounds mirroring the code clamps verbatim. The daily gate is derived from the 4H closes (trend_days x 6 bars back) rather than a second bar feed - it is the same quantity, and it removes the cross-feed alignment failure mode that costs multi-leg strategies whole verification loops. Venue is BINANCE USD-M futures as specified: the strategy takes shorts and uses leverage 2.0, which is genuinely consumed - the exposure cap (equity x 20% x leverage) binds against the risk-based notional, so 2x roughly doubles position size (dry-run full-sample: +107% at 1x vs +240% at 2x). Dry-run on the real ETHUSDT 4H catalog (2019-12-31 to 2026-08-05, entries/exits at bar close, 0.05% taker each side): 404 trades (61/yr), avg trade +0.86% of notional net of fees (well clear of the 0.15% floor), Sharpe 1.11, max DD 23%, +240% cumulative. The daily gate is the lever that pays - ungated the same composite gives Sharpe 0.68 and avg trade +0.28%. Parameter surface is smooth with no cliffs (trend_days 24/30/40 -> Sharpe 0.83/1.11/1.04; trail 4/5/6 -> 0.86/1.11/1.07). On the 365-day sandbox window it places 52 trades, so the Layer-3 >=1-trade gate is not at risk, and the composite varies every bar (sd ~0.86) rather than sitting at a gated default. min_bars_required is deliberately 25 because all indicator state accumulates inside calculate_signal - skipping bars there would starve the factors - and readiness is governed by an internal flag, keeping warmup independent of the base class's 500-bar buffer cap. All per-bar work is O(1) (bounded deques + running sums), so the 300s smoke cap is not a concern. One honest caveat, as the hypothesis itself concedes: the four factors are all price/volume views of the same tape and the daily gate is the momentum factor at a slower speed, so this is better described as a smoothed multi-view momentum than a blend of independent signals - the benefit should show up as stability, not a higher peak conditional edge. I implemented it as specified.

OHLCV momentum/trend-confluence composite on ETH — the 0/213 survival class (L41/L9). Per-trade (0.956%) and trade-count (404) metrics are clean, but the base Sharpe is only 0.821 with sharpe_ci_low 0.184 (CI floor near zero). After the optimizer selects best-of-~225 trials, the selected Sharpe will not clear the deflated-Sharpe gate — the hypothesis itself predicts 'another near-miss' at Sharpe ~0.7-0.9, and ETH's own plain 4H+1D confluence already landed at DSR 0.76 (below 0.95). The BTC composite siblings and ~37 daily probes of this identical idea all DSR-failed this session. Abandon at BACKTEST_REVIEW rather than spend the 2-hour optimization budget to reproduce the documented near-miss.

Long-short directional strategy on ETHUSDT.BINANCE 4H perpetual bars. Every bar it computes an equal-weight composite of four pre-registered 4H factors, each standardised to a z-score over a rolling 240-bar window: 5-day momentum, 1-day close-location pressure, 7-day signed volume flow, and 10-day range position. calculate_signal returns that composite (z units) continuously. It enters long when the composite >= +comp_thresh AND the 30-day trend (measured from the 4H closes, i.e. the daily trend sampled on the trading grid) is up; short on the mirror condition; and stands FLAT whenever the composite and the daily trend disagree. Sizing is risk-first off a fixed initial stop of 2 x ATR%, scaled by a volatility target (clamp(3.5% daily target / 20-day realized vol, 0.4, 1.6)) and capped at 20% of equity x leverage. Exits: the fixed initial stop, a trailing stop trail_atr_mult x ATR% behind the favourable extreme once the trade has moved one initial-stop distance in its favour, or a 30-bar (5-day) max-hold backstop. Only three tunable parameters (comp_thresh, trend_days, trail_atr_mult), each with declared bounds.

Backtest_review/analyst: judge the full-sample and walk-forward OOS Sharpe against the deflated-Sharpe 0.95 hurdle -- the evidence leans toward a near-miss (full-sample 1.11 < the plain confluence's 1.30 at DSR 0.76), which the hypothesis frames as the decisive answer that the survivor's edge is config-specific (its exact EMAs), not structure-general. Note the composite underperforms the existing plain ETH confluence, so it is not an improvement on the current book even if it lands close.

Likely a near-miss on DSR, and the composite does NOT beat the plain ETH confluence it is trying to lift -- the deciding analyst/backtest_review concern, but not a code defect. This is measurable and positive (sandbox 51 trades, +13.6%, Sharpe 0.66, PF 1.34, avg_trade_return_pct +0.90%, PSR 0.79, metrics_reliable=TRUE) and, unlike the daily composites, the dry-run (61/yr) and sandbox (51/yr) AGREE, so the measurability is genuine and it is the strongest of the composite family. BUT the full-sample dry-run Sharpe is 1.11 -- above the ~0.5 singles yet BELOW the plain ETH 4H+1D confluence's ~1.30 (which already exists and reached only DSR 0.76). So the diversification premise is falsified again: the composite is a smoothed multi-view momentum that does not improve on the plain confluence, and if the stronger plain confluence (1.30) missed at DSR 0.76, this composite (1.11) is unlikely to clear DSR 0.95. The sandbox Sharpe CI [-0.89, 2.13] still straddles zero. This is the near-miss the hypothesis honestly pre-registered.

Research Lead/analyst: treat as a stabilised 4H momentum composite whose selectivity gate is correlated with its own signal; the value proposition is horizon (4H) + stability, not orthogonal factor diversification or independent confluence.

Momentum-correlated composite with a non-independent gate -- honestly disclosed, a Research-Lead novelty/design note. The developer discloses the four 4H factors are all price/volume views of the same tape (momentum<->location tightest) so this is a smoothed multi-view momentum rather than orthogonal diversification, AND that the daily gate is the momentum factor at a slower speed, so the composite/daily-trend 'agreement' selectivity is less independent than a true two-signal confluence. The equal-weight composite and gate are faithfully implemented as specified, so this is not a fidelity defect, but the diversification-and-selectivity premise is materially weaker than the title implies.

No code change warranted; the implementation faithfully and correctly ports the conviction-composite construction to ETH 4H.

The code is CORRECT and the min_bars_required() framework fix is well-engineered -- credit; structurally identical to the BTC 4H composite I verified. The four-factor equal-weight composite (rolling-z, bullish-positive polarity), the daily gate derived from the 4H closes (no second feed, no cross-alignment risk), the armed trailing stop, and the vol-targeting (per-bar stdev x sqrt(6) daily-scaled) are all correct with guards and no look-ahead; only the instrument, z_window (240) and vol_target (0.035) differ from the BTC version. min_bars_required()->25 with the ~250-bar warmup enforced internally via self._ready is verified safe (the strategy maintains incremental deques/sums and never scans self._bars for history beyond the current bar, so the base class's 500-bar buffer cap does not affect its state and no premature trade can occur). ETH size_precision is read from the instrument. Not an L17 code-defect (win_rate 0.41, Sharpe 0.66, PF 1.34).

Healthy sample: 404 trades over 6.6 years

avg_trade_return_pct 0.956% is well above the 0.15% futures fee floor

PF 1.383, genuine long-short (beta 0.039, positive alpha 0.117) — not leverage-carried beta

Strong recent 2025 (+54.7%); ETH has demonstrated some edge at this 4H+1D horizon

Base Sharpe only 0.821 with sharpe_ci_low 0.184 — CI floor near zero; after best-of-225-trial selection this is essentially certain to fail the deflated-Sharpe gate (DSR>=0.95)

OHLCV momentum/trend-confluence composite — the 0/213 survival class (L41/L9)

The hypothesis itself frames the likely outcome as 'another near-miss' at Sharpe ~0.7-0.9 — an admission it won't clear the DSR bar; ETH's plain confluence already landed a DSR 0.76 near-miss

max_drawdown 24.4%, information_ratio -0.55, return_kurtosis 11.2 — fat-tailed, modest risk-adjusted edge

EthFourHourConvictionCompositeLS was the cross-asset test of the session's central question: it applied the diversified conviction-composite construction to ETH, the strongest non-BTC asset, whose plain 4H+1D confluence had already reached a genuine DSR 0.76 near-miss, to see whether the survivor's three structural levers could push it over the 0.95 bar. It performed respectably — +232.2% over 404 trades, PF 1.38, 0.96% per trade, positive alpha, strong 2025 — but the base Sharpe was only 0.821 with a CI floor near zero, a 24.4% drawdown, fat tails, and an information ratio underperforming the benchmark. The analyst abandoned it at backtest review: a 0.82 Sharpe in the 0/213 OHLCV momentum class cannot clear deflated Sharpe after best-of-~225-trial selection, ETH's own confluence had already landed below the bar, and the hypothesis itself predicted another near-miss. Read alongside the BTC composites and ~37 daily probes, it confirmed the survivor's edge is config-specific rather than structure-general; it never reached optimization, analysis, or risk review.

Applying the survivor's structural levers to the best non-BTC asset's demonstrated horizon edge still only reaches Sharpe ~0.82 — below both the DSR bar and ETH's own plain 0.76-DSR confluence — confirming that diversification/selectivity/risk-discipline cannot manufacture significance from a correlated OHLCV momentum composite; the survivor's edge is config-specific, not a transferable recipe across assets.

The analyst abandoned it at backtest review: a base Sharpe of 0.821 with a CI floor near zero is essentially certain to fail the deflated-Sharpe gate (≥0.95) after best-of-~225-trial selection, it is the same 0/213 OHLCV momentum/trend-confluence class, ETH's own plain confluence had already landed below the bar at DSR 0.76, and the hypothesis itself predicted 'another near-miss' at Sharpe ~0.7-0.9 — an admission it would not clear the gate.

A long-short, single-instrument directional strategy on ETHUSDT.BINANCE 4H bars gated by the daily trend (3 tunable parameters) that applied the survivor's three structural levers — diversification (equal-weight composite of four 4H factors: fast momentum, close-location pressure, volume-flow, range-position), selectivity (engage only when the composite and daily trend agree, flat otherwise), and risk discipline (vol-targeted sizing, armed trailing-stop winners) — to ETH, the strongest non-BTC asset whose plain 4H+1D confluence had already reached a DSR 0.76 near-miss, testing whether the levers could push that near-miss over the 0.95 DSR bar.

The backtest (14460 4H bars, 2019-2026) returned +232.2% over 404 trades with profit factor 1.38, avg_trade_return_pct 0.956% (well above the fee floor), genuine long-short character (beta 0.039, positive alpha 0.117), and strong recent 2025 (+54.7%). But base Sharpe was only 0.821 (sharpe_ci_low 0.184, near zero), max drawdown was 24.4%, return kurtosis 11.2 (fat-tailed), and information ratio was -0.55 versus the benchmark.