Hyperliquid BTC 4-Hour Momentum Trend-Continuation, LONG/SHORT (Single-Instrument HYPERLIQUID Perp, Pure OHLCV, Flip-Via-Flat, Fixed-Fractional Sizing, Positive Payoff Geometry — Tests the Untested 4H Middle-Cadence That Could Satisfy BOTH HL's Deep-History AND Holdout-Trade-Density Requirements, Pre-Registered DATA-DEPTH Kill-Switch + Sizing Probe, Fills Long-Short + HL Venue, Tail-Safe, Low-Parameter)

A LONG-SHORT, SINGLE-INSTRUMENT, 4-HOUR momentum / trend-continuation strategy on BTCUSD.HYPERLIQUID. This deliberately targets the ONE cadence not yet ruled out, derived from the exact failure geometry of the prior HL runs: HL is the only clean-sizing venue (BTC/ETH ~90%/64.5%; BINANCE USD-M/COIN-M over-exposure-broken, spot-CASH/multi/options broken, census data-walled), but HL DAILY (deep ~3yr) is too trade-sparse for the 15-day holdout (~15 daily bars -> 0-1 holdout trades, which auto-rejected the daily breakout), while HL 15-MIN has enough bars/day but only ~2 MONTHS of history (2026-03-16+, both majors) -> also unvalidatable. Daily and 15-min each fail a DIFFERENT validation requirement. 4-HOUR is the untested middle ground designed to satisfy BOTH: it is NOT sub-hourly (so NOT inside the confirmed ~2-month sub-hourly wall), and a 15-day holdout spans ~90 4H bars (vs ~15 daily) giving enough holdout trades, while a momentum book holding ~1-3 days generates a few hundred trades over the sample for a valid 3-window walk-forward. The binding UNKNOWN is whether 4H HL BTC history is deep (like daily, ~2023+) or shallow (like 15-min, ~2 months); I pre-register a hard kill-switch on it. Mechanism is asset-matched: BTC trends (the HL BTC daily breakout was clean & positive, PF 1.76), so momentum/continuation fits BTC (whereas mean-reversion fits choppy ETH). LONG/SHORT fills the dominant direction gap (long_only 85.9% vs <=55%) on the PROVEN-clean HL BTC long-short path (TS-mom LS sized 97%); flip-via-flat so the engine never nets opposing fills. Payoff geometry is explicitly CORRECTED from the prior 15-min reversion flaw (avg_win < avg_loss trap): the trend-continuation target/trail must exceed the stop distance net of the ~0.09% HL round-trip so avg_win > avg_loss. Bug-avoiding: pure OHLCV, ONE fixed-fractional size per entry, NO dynamic resizing, leverage 1.0, dual flat guard. PRE-REGISTERED ACCEPTANCE (two gates, to avoid thrashing): (1) DATA-DEPTH — verify 4H HL BTC data spans >= ~2 years AND the 15-day holdout contains >= ~5 trades; if NOT, this confirms the HL sub-daily data wall extends to 4H -> ABANDON & ESCALATE to the data engineer (do NOT iterate). (2) SIZING — avg_position_pct ~<=100% and no physically impossible 4H/daily equity jump; if violated, abandon & escalate as a reproducer. Low parameter count (momentum lookback, trend filter, ATR stop/trail).

Implements the hypothesis by targeting the one HL cadence not yet ruled out. HL daily is deep (~3yr) but too sparse for the 15-day holdout (~15 bars -> 0-1 trades), and HL 15-min has enough bars/day but only ~2 months of history; 4-HOUR is the middle ground that can satisfy BOTH requirements (a 15-day holdout spans ~90 4H bars for enough holdout trades, while a ~1-3 day-hold momentum book yields a few hundred trades over the sample for a valid 3-window walk-forward) AND sits above the confirmed ~2-month sub-hourly data wall. Mechanism is asset-matched: BTC trends cleanly (the HL BTC daily breakout was positive, PF 1.76), so momentum/continuation fits BTC, and LONG/SHORT fills the dominant direction quota gap on the proven-clean HL BTC long-short path (TS-mom LS sized ~97%). Payoff geometry is explicitly corrected from the prior 15m reversion avg_win<avg_loss trap: there is no early fixed take-profit; an ATR trailing stop (3.5x) lets the trend's right tail run while a tighter frozen initial stop (2.5x) caps losses, so avg_win exceeds avg_loss net of the ~0.09% HL round-trip. Construction is defect-avoiding: single-instrument pure OHLCV (no funding subscription, no extra legs, no options), flip-via-flat achieved structurally (should_exit closes to flat and the base engine only opens an entry when flat, so opposing fills are never netted), ONE fixed-fractional size (notional = 0.95 * equity, qty = notional/price) with no dynamic resizing, and leverage left at 1.0 so the engine margin model and in-strategy notional agree. Two pre-registered downstream gates: (1) DATA-DEPTH — if 4H HL BTC data does not span >= ~2 years AND the 15-day holdout contains >= ~5 trades, this confirms the HL sub-daily data wall extends to 4H and the strategy should be abandoned & escalated to the data engineer (do NOT iterate); (2) SIZING — avg_position_pct should be <=~100% with no physically impossible 4H/daily equity jump, else abandon & escalate as an HL over-exposure reproducer.

Abandon for NO EDGE, not for a data wall — the hypothesis's own pre-registered kill-switches both PASSED, so the infrastructure question was answered favorably and need not be re-run. DATA-DEPTH gate: 4H HL BTC data is deep (5057 bars ≈ 2.3 years back to 2024-01, 267 trades, 15-day holdout would span ~90 bars / several trades) — the HYPERLIQUID sub-daily data wall does NOT extend to 4H; this is a useful, reusable finding for FUTURE HL strategies. SIZING gate: avg_position_pct 106% (exposure_fraction 0.95 + MTM, not the over-exposure defect), no impossible equity jumps, metrics_reliable=true, balanced long/short legs — the HL 4H long-short flip-via-flat path sizes cleanly. However, the strategy is a confirmed net loser at base parameters over a large, statistically meaningful 267-trade sample: total_return -19.05%, Sharpe -0.54, profit_factor 0.887, avg_trade_return -$137.5 (≈ -0.14%/trade, NEGATIVE, below trading costs and below the analyst abandon threshold). The payoff geometry was ALREADY corrected (avg_win 2.19x avg_loss) and it still loses because win rate (28.8%) is below the 31.3% geometry breakeven — i.e. the vol-normalized momentum signal has no edge on 4H BTC, and optimization can only re-threshold a vanilla t-stat, not manufacture edge. The single positive period is 2024 bull beta (+18.9%; 2025 -43.6%, 2026 -5.7%), so optimization would curve-fit the 2024 window and fail the 2025-2026 holdout — the identical OOS/holdout/DSR collapse the four prior HL daily/4H trend siblings (BTC/ETH/SOL Donchian, SOL TS-momentum) suffered this session. Not worth 2 hours of optimization: a robustly negative-expectancy base case from a family that uniformly fails post-optimization generalization. Not iterate: there is no code-level bug to fix — the construction (pure OHLCV, flip-via-flat, corrected payoff geometry, clean sizing) is sound; the premise (momentum edge on 4H BTC) simply has no edge. PRESERVED FINDING for the research lead: HL 4H is a validatable, clean-sizing cadence — point a mechanism with a REAL edge at it rather than another vanilla single-asset trend/momentum signal.