LINK Hyperliquid Dual-Timeframe Momentum Confluence — Long-Short 4H Trend Aligned with the Daily Trend, Flat on Disagreement, ATR Trailing Exit (Volatility-Scaled, 3-Parameter)

A LONG-SHORT, single-instrument, pure-price directional MOMENTUM-CONFLUENCE trend follower on LINK-USD perpetual on HYPERLIQUID (LINKUSD.HYPERLIQUID), 4H primary bars with a 1D trend confirmation. This deliberately reproduces the factory's ONLY surviving recipe — the paper-stage 'BTC Dual-Timeframe Momentum Confluence' (Sharpe 1.99), a long-short 4H+1D momentum confluence that goes flat on disagreement and trails winners — and applies it to (a) the most under-represented actionable venue (HYPERLIQUID 6.5% vs 20% target), (b) the under-represented long-short direction bucket (13.7% vs an 86% long-only book), and (c) a NEW liquid major (LINK) not yet covered by the BTC (Binance), ETH-HL (Donchian), or SOL-HL variants. LINK is chosen over BTC because its trend moves are larger in percentage terms (higher beta) — which lifts per-trade capture well above Hyperliquid's ~0.09% round-trip — but it is materially LESS fat-tailed than memecoins (DOGE/WIF), so the backtest edge is not carried by one or two outlier days (the outlier-dependence abandon that killed the BNB/XRP daily trend followers). It is NOT a mean-reversion fade (L53), NOT a squeeze breakout (L54), NOT a cointegration/basis pair (dead), NOT a cross-sectional rank basket (L52), NOT a non-price-feed gate (L46), NOT options/COIN-M (L50/L51). 4H+1D on Hyperliquid gives ~2.3y of usable 4H history plus full daily history — enough to populate walk-forward windows and the 15-day holdout (trade cadence ~80–140/yr keeps the holdout non-empty) and clear of the sub-hourly HL history trap. Risk profile: ~1.5% equity risk per trade via ATR-based stop distance; volatility-scaled sizing (notional = risk_budget / (atr_mult * ATR_4H)); max single-position exposure ~25% of $100k; leverage capped at 2x (reads self.config.leverage). Exactly 3 tunable parameters (fast/slow EMA pair as a single ratio, daily-trend lookback, ATR trailing multiple) to resist the 305 overfit deaths.

Implements the hypothesis exactly: the paper-stage BTC dual-timeframe confluence recipe (fast trend must agree with a slower higher-timeframe trend, flat on disagreement, ATR trail runs winners) ported to LINK on Hyperliquid -- under-represented venue, under-represented long-short bucket, new liquid major. Genuine dual-timeframe: the daily leg is a real LINKUSD.HYPERLIQUID-1-DAY series consumed via on_extra_bar (extra_instruments carries LINKUSD.HYPERLIQUID so InstrumentResolver.resolve_extras, which pairs bar types POSITIONALLY, does not silently drop the daily feed -- the exact failure that produced a zero-trade sandbox on a prior port). Daily is a slower-timeframe context leg, so the base class's same-timeframe sync barrier never defers the 4H signal. Exactly 3 tunables as specified (fast/slow EMA pair collapsed into one ratio, daily lookback, ATR trail multiple); clamps in code are mirrored verbatim in _param_bounds so the optimizer never tests values the code rewrites. Sizing is ATR/risk-based (1.5% equity risked at the stop distance) and consumes self.config.leverage in the notional cap, so leverage=2.0 is genuinely used rather than inert; the notional cap is independent of the equity path, avoiding the compounding-sizing blowups (L15). VERIFIED OFFLINE before submitting, on the actual catalog data (5549 4H bars from 2024-01-25 + 2178 daily bars): the coded rule fires ~78 round trips/yr (197 trades over 2.5y) -- comfortably above the zero-trade gate and dense enough that a 15-day holdout is not empty -- with avg per-trade return +1.48% net of Hyperliquid's ~0.09% round-trip taker cost, ~44% win rate (winners larger than losers, as the trail intends), trade-return kurtosis 5.8 and the top-3 trades contributing only ~17% of gross profit, so the edge is not outlier-carried. Critically, the ENTIRE 4x4x4 parameter grid (ema_ratio 2.0-4.0, daily_lookback 10-30, trail 2-5) is profitable both full-sample AND over the trailing 365 days (recent avg per trade +0.18% to +1.21%), so the chosen defaults sit on a plateau rather than a spike -- the defaults are the plateau centre, not the grid maximum. Layer-2 synthetic scenarios were run locally and all six pass with a continuous, varying signal. Per-bar work is O(1) (incremental EMAs, bounded TR deque), so the 300s sandbox cap is not at risk.

Promising shape, unestablished edge, wrong venue/data for it to survive. This is the correctly-wired port of the proven dual-TF mechanism and, unlike the abandoned Donchian/breadth mirages, its edge is genuinely distributed (return_kurtosis 7.68, skew 0.22), positive every year (2024 +17.9%, 2025 +26.7%, 2026 +6.4%), with a low 8.7% drawdown, PF 1.60, and decorrelated alpha 0.13 over 195 trades. But the edge is not statistically established: Sharpe is only 0.46 with a bootstrap CI [-0.60, 1.48] that straddles zero deeply — the widest straddle of any directional candidate this batch — over a short ~2.5-year single-regime window (LINK-HL 4H history only starts 2024-01, an all-bullish period). The market is also thin: impact_cost_pct 13.4% of gross PnL with capacity only ~$5.6M, so the edge is partly a small-scale artifact that higher-turnover optimized params could worsen. Hyperliquid directional trend is the 0/97 class (L41), and the near-identical ETH-HL Donchian (Sharpe 0.49, CI-low -0.38) was just abandoned for this same profile — this LINK version is weaker on significance. Best-of-N over ~225 optimization trials on a Sharpe-0.46, CI-straddling-zero base will not clear deflated Sharpe (every directional sibling this session failed it from higher base Sharpes), and OOS windows on a 2.5-year sample are short. Abandon rather than spend the optimization budget; the distributed profile would need a longer, deeper track record than thin HL LINK can provide.