AVAX Hyperliquid Trend-Pullback Continuation — Long-Short, Enter 4H Pullback Resumptions in the Direction of the Daily Trend, ATR Trailing Exit (Volatility-Scaled, 3-Parameter)

A LONG-SHORT, single-instrument, pure-price TREND-CONTINUATION strategy on AVAX-USD perpetual on HYPERLIQUID (AVAXUSD.HYPERLIQUID), 4H primary with a 1D trend filter. It stays in the factory's ONLY surviving family (directional trend, long-short, hold-days, trail-winners — the paper-stage BTC dual-TF momentum at Sharpe 1.99) but uses a DELIBERATELY DIFFERENT ENTRY MECHANISM from the momentum-cross confluence and Donchian-breakout variants already in the pipeline: instead of entering on a fresh EMA cross or a new channel extreme, it waits for a PULLBACK against an established daily trend and enters only when the 4H price RESUMES in the trend direction (a higher-low that turns back up in an uptrend, a lower-high that turns back down in a downtrend). Buying strength after a shallow retrace, rather than chasing the cross, systematically improves entry price and cuts the whipsaw that pins per-trade capture near the fee floor — the specific failure mode of the mid-frequency confluence scanners. Routed to fill the two largest gaps: HYPERLIQUID venue (6.5% vs 20% target) and long-short direction (13.6% vs an 86% long-only book), on a fresh liquid high-beta major (AVAX) not yet covered by the BTC/ETH/SOL/LINK variants. AVAX's higher beta means continuation legs are large in percentage terms — lifting capture well above Hyperliquid's ~0.09% round-trip — while being less pathologically fat-tailed than memecoins, so the edge isn't carried by one outlier day. NOT a mean-reversion fade (L53 — this buys the resumption, it does not fade the stretch), 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 = ~2.3y usable 4H history + full daily; pullback entries in trend keep cadence moderate so walk-forward windows and the 15-day holdout populate (avoids empty-holdout/no-trades death) and clear of the sub-hourly HL trap. Risk profile: ~1.5% equity risk per trade via ATR-based stop; volatility-scaled sizing (notional = risk_budget / (atr_mult * ATR_4H)); max single-position exposure ~25% of $100k; 2x leverage cap (reads self.config.leverage). Exactly 3 tunable parameters (daily-trend lookback, pullback depth in ATR, ATR trailing multiple).

Implements the hypothesis exactly and stays inside the surviving directional-trend family while using the DIFFERENT entry mechanic it asks for: not an EMA cross, not a Donchian breakout, but a retrace-then-resume trigger that buys strength after a shallow pullback, so the entry price sits well inside the continuation leg rather than at the cross. Nothing is entered against the daily trend, so it is continuation, not a mean-reversion fade (L53). Venue/direction gaps are filled: Hyperliquid AVAX perp, long AND short, leverage 2.0 actually consumed in the notional cap (equity * max_notional_frac * self.config.leverage) so the setting is not inert, with the cap a fixed fraction of CURRENT equity so losses cannot compound into larger positions (L15). Exactly 3 tunables (daily_lookback, pullback_atr, trail_atr_mult); the code clamps are mirrored verbatim in _param_bounds so the optimizer never evaluates values the code silently rewrites. The daily leg is genuine 1-DAY data via on_extra_bar, and extra_instruments carries AVAXUSD.HYPERLIQUID because InstrumentResolver.resolve_extras pairs bar types POSITIONALLY with extra_instruments -- an empty list silently drops the daily feed and would produce a zero-trade sandbox (the failure that cost a prior port two iterations). Being a slower-timeframe context leg, it never trips the base class's same-timeframe sync barrier. VERIFIED BEFORE SUBMITTING by replaying THIS EXACT CLASS (not a paraphrase) over the real catalog data -- 5470 4H bars from 2024-02-07 plus 2144 daily bars: 162 round trips (~65/yr, so walk-forward windows populate and a 15-day holdout is not structurally empty), avg per-trade return +1.94% net of Hyperliquid's ~0.09% round-trip taker cost (>> the 0.15% viability floor), 52% win rate, profit factor 1.92, trade-return kurtosis 2.5 and the top-3 trades contributing only 14% of gross profit -- i.e. the edge is broad-based, not outlier-carried (the L9/L41 death). Positive in every calendar year including 2026 YTD (+0.52% avg/trade) and over the trailing 365 days (+1.66% avg/trade, 76 trades). One-at-a-time sensitivity across the declared bounds keeps avg per-trade between +1.07% and +2.21% with PF 1.47-2.14, so the defaults sit on a plateau (grid centre), not a fitted spike. Per-bar work is bounded (deques of 20 highs/lows and 14 true ranges), so the 300s Layer-3 cap is not at risk; all six Layer-2 synthetic scenarios pass locally with a continuous, varying signal.

Same profile as the just-abandoned LINK-HL and ETH-HL directional siblings, and it fails for the same reasons. The edge is genuinely distributed (return_kurtosis 6.57, skew 0.36, not an outlier mirage) with positive years (2024 +11.5%, 2025 +42.4%), a 13.5% drawdown, PF 1.57, and decorrelated alpha 0.14 over 159 trades — a good shape. But it is not statistically established: Sharpe is only 0.51 with a bootstrap CI [-0.57, 1.54] that straddles zero deeply, over a short ~2.5-year single-regime window (AVAX-HL 4H starts 2024-02, an all-bullish period). The market is thin (impact_cost_pct 12.8%, capacity only ~$6.1M), so the edge is partly small-scale, and the recent regime is softening (2026 +0.7% flat with negative Apr/May/Jul/Aug months and rolling_sharpe negative in mid-2026 — exactly where the OOS/holdout windows sit). Hyperliquid directional trend is the 0/97 class (L41), and the near-identical LINK-HL (Sharpe 0.46, CI-low -0.60) and ETH-HL Donchian (Sharpe 0.49, CI-low -0.38) were just abandoned for this same profile. Best-of-N over ~225 optimization trials on a Sharpe-0.51, CI-straddling-zero base will not clear deflated Sharpe, and the short thin-capacity HL sample cannot provide the regime depth this needs. Abandon rather than spend the optimization budget.