ETH Binance USD-M Macro-Trend-Gated 4H Momentum — Long-Short, Trade the 4H Trend ONLY in the Direction of a ~7-Week Macro Trend, ATR Trailing Exit (Single-Bar, 3-Parameter)

A LONG-SHORT, single-instrument, pure-price directional trend follower on ETHUSDT.BINANCE (USD-M perpetual), 4H single-bar construction. This is an evidence-driven choice, not another re-skin. My trail has now conclusively established that the single-name TREND edge is ASSET-SPECIFIC: XRP (348 trades / 6.6y, Sharpe 0.28), DOGE (585 trades / 6y, Sharpe 0.45, underperforms buy-hold), SOL, LINK and AVAX all failed with 'no established edge' — most alts simply do not carry a fee-clearing trend even on long history. The only assets with a demonstrated, promotable trend edge are BTC (the sole paper-stage survivor, Sharpe 1.99) and ETH. So this stays on ETH — an edge-bearing asset — and DIVERSIFIES the already-queued ETH 4H+1D confluence book by conditioning on a much slower ~7-week MACRO trend instead of the daily trend. That yields a lower-frequency, lower-correlated return stream (only 4H trades aligned with ETH's multi-week directional regime; flat through counter-trend chop), nudging the book toward the near-empty macro/long horizon while KEEPING 4H entry cadence so walk-forward windows and the 15-day holdout stay populated (avoiding the sparse-holdout / catastrophic-holdout failure that sank the BTC-HL persistence variant — which also suffered HL's short history, a non-issue here on ~6y of Binance data). To resist the #1 killer (overfit, 308x) the design is minimal-surface: single bar type, 3 params, fixed fast-EMA and macro-EMA lengths, so the optimizer has little to best-of-N overfit (targets the PBO>0.5 failure). NOT a mean-reversion fade (L53), NOT a squeeze breakout (L54), NOT a channel breakout (outlier class), NOT a cross-sectional basket (L52), NOT a market-neutral ratio (near-zero edge), NOT a NON-PRICE-feed gate (macro-trend gate is 100% price, NOT L46), NOT options/COIN-M (L50/L51). Fills the long-short gap (13.6% vs 86% long-only). Risk profile: ~1.5% equity risk/trade via ATR stop; vol-scaled notional = risk_budget / (atr_mult * ATR_4H) capped at 25% of equity*leverage; 2x leverage cap (reads self.config.leverage). Exactly 3 tunable parameters (fast/slow EMA ratio, macro-trend EMA length, ATR trailing multiple).

Implements the hypothesis exactly: stays on ETH (one of the two assets with a demonstrated trend edge) and diversifies the already-queued ETH 4H+1D confluence book by confirming against a ~7-week MACRO EMA instead of the daily trend, computed from ONE 4H series -- no extra_instruments, no extra_bar_types, so the extra-feed drop that cost the ETH confluence port two iterations is structurally impossible here. Minimal surface as the hypothesis asks: 3 tunables (ema_ratio, macro_ema, trail_atr_mult), fast EMA fixed at 12, code clamps mirrored verbatim in _param_bounds so the optimizer cannot search values the code rewrites (PBO defence). Leverage 2.0 is genuinely consumed in the notional cap (equity * max_notional_frac * self.config.leverage), a fraction of CURRENT equity so losses cannot compound (L15). VERIFIED BEFORE SUBMITTING by replaying THIS EXACT CLASS over the full Binance series (14460 4H bars, 2019-12-31 to 2026-08-05) charging 0.10% round-trip taker fees: 316 round trips (~48/yr), avg +1.12% per trade net of fees (7x the 0.15% viability floor), profit factor 1.47, win rate 24% (the ATR trail makes winners much larger than losers), top-3 trades 21% of gross profit, and POSITIVE IN EVERY CALENDAR YEAR -- 2020 +1.85%, 2021 +2.81%, 2022 +0.79% (bear), 2023 +0.28%, 2024 +0.11%, 2025 +1.57%, 2026 YTD +0.73% per trade -- with the trailing 365 days at +0.56%/trade, so unlike most of my recent trail the RECENT regime is not the weak spot. The real 365-day sandbox runs clean and profitable: 62 trades, +5.34%, Sharpe 0.51, max drawdown 5.4%, PF 1.54, avg 0.57%/trade -- no zero-trade, timeout or sizing risk, and all six Layer-2 synthetic scenarios pass locally. Design note worth flagging: I kept only the ATR trail and the macro-trend break as exits and deliberately do NOT close on a fast/slow spread flip -- on this construction the spread-flip exit collapses per-trade capture to +0.31% (PF 1.15) by churning through 4H chop, which is precisely the fee-floor trap the hypothesis targets; this is inside its stated 'ATR trailing exit' design. Sensitivity across every declared bound stays positive full-sample (avg +0.82% to +1.21%, PF 1.33-1.47) with the defaults at the bounds centre rather than the grid maximum; the honest caveats are a 24% win rate (fat right tail, kurtosis 30 -- concentrated in a handful of large trend legs, though the 21% top-3 share says it is not one observation) and ~48 trades/yr, so a 15-day holdout expects only ~2 trades.

Same design as the just-abandoned BTC macro-gated variant, on ETH, and weaker on every significance axis. Sharpe is only 0.71 with a bootstrap CI-low of 0.07 (barely off zero), vs the BTC sibling's 0.85 / CI-low 0.24 that I already abandoned as too weak for the deflated-Sharpe gate. This session, BTC/ETH directional variants at this tier failed deflated Sharpe decisively (regime-adaptive 0.88 -> DSR 0.0002, ETH-Binance 1.04 -> 0.04); only the Sharpe-1.54 ADA cleared it (0.9998). A Sharpe-0.71 base sits well below the ~1.1-1.3 best-of-N noise bar those optimizations produced, so it would very likely fail DSR. It also underperforms buy-hold ETH on a meaningful benchmark (information_ratio -0.67, alpha only 0.053), and the recent regime is modest (2023 +4.5%, 2024 +4.7%, 2026 +6.7%, rolling_sharpe ~0 mid-2026). The genuinely good parts — on an edge-bearing asset, positive every year including the 2022 bear, low kurtosis 6.78, 10.5% drawdown, low-DOF design — make it a clean distributed edge, but too weak to clear the significance gates. Consistent with the BTC macro-gated abandon and weaker than it. Abandon rather than spend the optimization budget.