ETH/BTC Ratio Trend — Market-Neutral Long-Short of the ETH-vs-BTC Relative-Strength Trend on Hyperliquid (Long ETH-Perp / Short BTC-Perp When the Ratio Trends Up, Reverse When It Trends Down; Dollar-Neutral, 8H Signal + Daily Confirm, 3-Parameter)

A MARKET-NEUTRAL, two-leg, pure-price TREND-follower on the ETH/BTC relative-strength ratio, traded on Hyperliquid perps (primary ETHUSD.HYPERLIQUID, hedge BTCUSD.HYPERLIQUID). This is a deliberate FAMILY change away from the single-name directional trend-followers I have been proposing (BTC/ETH/SOL/LINK/AVAX), which keep dying for one repeated reason at review: they capture crypto BETA, not alpha, and underperform simply holding the coin (the ETH-HL Donchian was just abandoned with alpha 0.038 / information_ratio -0.41 vs buy-hold ETH; the BTC-Donchian died as an outlier-mirage). Making the book DOLLAR-NEUTRAL (equal notional long ETH / short BTC) structurally removes the market beta, so this strategy CANNOT be dismissed as captured buy-hold return — its PnL is the ETH-minus-BTC relative trend, an orthogonal alpha source the portfolio has no exposure to. Crucially this is the TREND of the ratio (relative-strength persistence: when ETH is out-performing BTC it tends to keep out-performing over weeks — the alt-season vs BTC-dominance regimes), NOT the mean-REVERSION of a cointegration residual (the ETH–SOL cointegration pair and 'ETH/BTC ratio reversion' both died — this trades the OPPOSITE sign, so it is not that dead family). It is NOT a cross-sectional rank-rotation basket (L52 is about ranking ≥3 names; this is a single, fixed, economically-motivated spread), NOT a squeeze breakout (L54), NOT a non-price-feed gate (L46), NOT options/COIN-M (L50/L51), NOT a delivery-future basis (L47). It also fills multiple under-represented buckets at once: HYPERLIQUID venue (6.6% vs 20%), pairs scope (13.7%), long-short/market-neutral direction (13.6% vs 86% long-only). 8H signal on Hyperliquid has ~4.6y usable history and daily is full — enough to populate walk-forward and the 15-day holdout. Risk profile: gross exposure ~50% of $100k equity (≈25% per leg), dollar-neutral so residual complex-beta ≈0.1; a hard gross cap prevents the cross-leg margin blow-ups that destroyed prior pairs; ~1.5% equity risk per trade via a ratio-based stop; leverage ≤2x (reads self.config.leverage). Exactly 3 tunable parameters (fast/slow ratio-EMA pair as one ratio, daily-ratio-trend lookback, ratio ATR stop/trailing multiple).

Implements the hypothesis exactly: the traded object is the ETH/BTC ratio TREND (relative-strength persistence), expressed as equal-notional long/short perp legs, so market beta is structurally removed and the PnL cannot be re-described as captured buy-and-hold -- the specific reason the single-name HL trend followers were abandoned. The sign is deliberately the OPPOSITE of the dead ratio-reversion/cointegration family: a positive ratio trend BUYS more ETH-vs-BTC. It fills the HYPERLIQUID venue, pairs-scope and market-neutral-direction buckets at once. Exactly 3 tunables (ema_ratio collapsing the fast/slow pair, daily_lookback, trail_atr_mult); code clamps mirror _param_bounds verbatim so the optimizer never tests values the code rewrites. MULTI-LEG PLUMBING, done deliberately: extra_instruments and extra_bar_types are paired POSITIONALLY by InstrumentResolver.resolve_extras, so all three extra feeds carry BOTH entries (BTC 8H hedge leg, ETH 1D, BTC 1D) -- omitting the instrument entry silently drops the feed and produces a zero-trade sandbox. The 8H BTC leg is the same timeframe as the primary, so the base class's cross-leg barrier defers the signal until the contemporaneous BTC bar arrives, and calculate_signal ADDITIONALLY refuses to compute or trade unless the stored BTC bar carries the primary bar's exact timestamp, so a one-bar-stale hedge leg can never manufacture a phantom ratio move (the classic pairs failure). Exits close both legs (overridden _submit_exit closes the hedge position too) and on_stop already flattens extras, so the book never carries a naked leg. VENUE/LEVERAGE NOTE: leverage is left at 1.0 rather than 2.0 -- a dollar-neutral book at ~50% gross exposure needs no borrowing, and setting an unused leverage is a verification reject; the code still READS self.config.leverage in the gross cap, so raising it scales gross exposure. VERIFIED BEFORE SUBMITTING by replaying THIS EXACT CLASS over real catalog data (3713 aligned 8H ETH/BTC bar pairs from 2023-01-01 plus 2179 daily pairs), charging Hyperliquid taker fees on ALL FOUR legs (2 legs x 2 sides = 0.18% round trip): 229 pair round-trips (~68/yr, so walk-forward windows and the 15-day holdout populate), avg +0.57% per trade on the per-leg notional (~+0.29% of the 2-leg gross) NET of those fees, 46% win rate, profit factor 1.67, top-3 trades only 16.6% of gross profit. Positive in 2024 (+0.67%/trade), 2025 (+1.14%) and 2026 YTD (+0.39%), ~flat in 2023 (-0.04%), and +0.81%/trade over the trailing 365 days -- the edge is not one regime or one outlier. One-at-a-time sensitivity across the declared bounds stays positive everywhere (avg +0.33% to +0.74%, PF 1.37-1.78), so the defaults are a plateau centre, not a fitted spike. Per-bar work is O(1) (incremental EMAs, bounded deques, dict-paired daily bars), so the 300s Layer-3 cap is safe; all six Layer-2 synthetic scenarios pass locally. Data caveat: the HL 8H catalog currently ends 2026-05-22 while the 1D feeds are current, so the sandbox will trigger on-demand collection to top up the 8H legs before the run.

Sound market-neutral construction around a near-zero edge that market impact destroys. The book is genuinely dollar-neutral (beta 0.002) and trades the ratio TREND (not the dead reversion), but there is no established edge: Sharpe 0.16 with a bootstrap CI [-0.70, 0.93] straddling zero deeply, profit_factor 1.07 (gross edge inside costs), and avg_trade_return_pct 0.28% barely above the ~0.18% two-leg Hyperliquid round trip. The decisive blocker is capacity/impact: impact_cost_pct 33.2% (well over the 25% line) eats a third of gross PnL and capacity is only ~$907K — the thin HL ETH/BTC ratio market makes this edge real only at toy scale, and even there it is marginal (total +16.1% over 3.4 years carried almost entirely by 2025 +13.7%, with 2023/2024/2026 flat, IR -0.56, alpha 0.033). This is not a tuning problem: a Sharpe-0.16, PF-1.07 result with a CI straddling zero has nothing to optimize toward, and the impact/capacity limit is structural to the instrument, not a parameter. Best-of-N over ~225 trials would fail deflated Sharpe, and any 'winning' config still trades at ~$907K capacity with 33% impact drag. benchmark_meaningful is false so judged on absolute metrics, which show no deployable edge. Abandon rather than spend the optimization budget.