Hyperliquid BTC Intraday 15-Minute Momentum Trend-Continuation, LONG/SHORT (Single-Instrument HYPERLIQUID Perp, Pure OHLCV, Flip-Via-Flat, Fee-Gated Selective Entries Holding Hours, Fixed-Fractional Sizing — Intraday LS on the One Fully-Proven-Clean Combo HL BTC [LS sizes 97%], Defeats the HL Daily Data-Sparsity Wall, Fills ALL THREE Top Gaps: Long-Short + HL Venue + Short Horizon, Pre-Registered Sizing Probe, Tail-Safe, Low-Parameter)

A LONG-SHORT, SINGLE-INSTRUMENT, INTRADAY 15-MINUTE momentum / trend-continuation strategy on BTCUSD.HYPERLIQUID that goes LONG on confirmed intraday up-momentum and SHORT on confirmed intraday down-momentum, flipping ONLY via flat. This is the highest-value remaining EVALUABLE idea because it uniquely fills the THREE most under-represented reachable buckets simultaneously: (1) DIRECTION — long_only is 85.8% vs the <=55% target (the single largest quota miss); only a long-SHORT book can fix it, and long/flat books (the two pending intraday HL books) cannot. (2) VENUE — HYPERLIQUID 6.8% vs >=20%. (3) HORIZON — short 1m-15m 3.7% vs >=10%. Crucially it sits on the ONE combination now PROVEN clean on every axis: HL BTC long-short sizes correctly at avg_position_pct 97% (the HL BTC TS-mom LS sibling, sent to optimize), so the LS / flip path is NOT the over-exposure bug — that was BINANCE USD-M/COIN-M (engine over-exposure) and HL SOL (early-data artifact, now dropped), while HL BTC/ETH are clean. It is DISTINCT from both the daily HL BTC TS-mom LS (which risks the HL daily data-sparsity wall: ~3yr history -> ~0-1 OOS trades for daily cadence) and the long/flat HL BTC intraday breakout (no short side): intraday cadence over ~3yr (~75k 15m bars) yields HUNDREDS of trades, making a valid 3-window walk-forward + 15-day holdout possible (defeating the wall that killed the HL daily books), and the LS construction adds intraday DOWN-trend capture. Constraint context: BINANCE USD-M + COIN-M perp over-exposure-broken (stop probing them); spot-CASH MTM, multi-instrument aggregation, options settlement broken; census signals data-walled -> OHLCV only; HL is the only clean venue and HL BTC the safest instrument. FEE DISCIPLINE is central (HL ~0.09% RT, need >0.15% net/trade): entries are SELECTIVE momentum continuations gated by range/volume expansion, holding HOURS (not minutes), targeting >0.3-0.4% gross legs — explicitly NOT scalping. FLIP-VIA-FLAT: on a momentum sign reversal, exit fully to flat one bar, then open the opposite side, so the engine never nets opposing fills (cleanest LS construction). Bug-avoiding otherwise: pure OHLCV, ONE fixed-fractional size per entry, NO dynamic resizing, leverage 1.0, dual flat guard. PRE-REGISTERED ACCEPTANCE / SIZING PROBE (retained; HL not assumed uniformly clean and this is the intraday-LS path): avg_position_pct ~<=100% and no physically impossible intraday equity jump (BTC at 1x ~+/-15-20%); if violated, abandon & escalate as the HL-BTC-intraday-LS reproducer. Low parameter count (momentum lookback, expansion filter, ATR stop, trend filter).

Implements the hypothesis as an intraday two-sided momentum book on the single combination proven clean on every axis (HL BTC long-short sizes ~97%), filling the three largest reachable quota gaps simultaneously: long_short direction, the HYPERLIQUID venue, and the short intraday horizon. The 15-minute cadence over ~3yr yields hundreds of trades, defeating the daily HL data-sparsity wall that produces too few OOS trades for a valid walk-forward + holdout. Defect-avoidance: (1) single-instrument, pure OHLCV (no funding subscription, no extra legs) so multi-instrument aggregation and funding defects cannot apply; (2) flip-via-flat is structural — should_exit closes to flat on a reversal and the base engine evaluates an entry only when position is None, so opposing fills are never netted; (3) ONE fixed-fractional size (notional = 0.95*equity, qty = notional/price), no dynamic resizing, leverage left at 1.0 so engine margin model and in-strategy notional agree. Fee discipline is central: the expansion gate makes entries selective momentum continuations targeting >0.3-0.4% gross legs, comfortably above HL's ~0.09% round-trip, explicitly avoiding scalping; the vol-normalized score keeps the signal continuous, conviction-bearing, and threshold-stable across volatility regimes. Pre-registered sizing probe: avg_position_pct should be <=~100% and no physically impossible intraday equity jump (BTC at 1x ~+/-15-20%); a violation would indicate the over-exposure defect reaches the HL BTC intraday LS path and should be abandoned/escalated as the reproducer.

Two independent decisive failures, before any optimization. (1) PREMISE FALSIFIED — DATA WALL NOT DEFEATED: the hypothesis bet that intraday 15m cadence over ~3yr (~75k bars) would yield hundreds of trades for a valid 3-window walk-forward + 15-day holdout, defeating the HL daily data-sparsity wall. But the run has only ~57 days of HL BTC 15m data (2026-03-16 to 2026-05-13, 4914 bars), not 3 years. A walk-forward/holdout over 57 days is meaningless, so the entire rationale collapses — the intraday cadence hit the SAME sparsity wall at finer resolution (symptom observed in-run; recommend verifying HL 15m catalog depth before retrying this family). (2) STRONGLY NEGATIVE EDGE: total_return -14.0%, Sharpe -10.0 (CI -14.7 to -6.4, entirely below 0), Sortino -11.9, profit_factor 0.42, win_rate 20.4%, avg_trade_return -$217 — and the fee-discipline claim failed outright (commission_pct_of_gross 51.68%, +$9,544 impact), so fees consumed half the gross despite holding hours. The only thing that worked is the sizing probe (avg_position_pct 101%, clean, metrics_reliable=true), which does not rescue the run. Not optimize: optimization cannot manufacture multi-year 15m data (so the promised validation is impossible regardless of parameters) and cannot convert a PF-0.42, fees-eat-52%-of-gross signal into an edge. The intraday-LS-momentum family has no clean+deep-data home right now — HL has only ~2 months of 15m history and the deep-history venues (BINANCE USD-M/COIN-M) are blocked by the engine over-exposure defect. FAILURE PATTERN: moving the HL trend edge to intraday 15m to escape the daily ~3yr sparsity wall does NOT work — the HL 15m catalog itself is only ~57 days, so the promised valid walk-forward/holdout is impossible, and the momentum-continuation signal on BTC 15m is a heavy net loser (Sharpe -10, PF 0.42) whose fees consume half the gross. Both the daily and intraday HL trend books are defeated by HL's short data history; the validated trend edge needs a deep-history venue that the engine can size correctly, which currently does not exist.