Hyperliquid BTC 4-Hour Mean-Reversion, LONG/SHORT (Single-Instrument HYPERLIQUID Perp, Pure OHLCV, Counter-Trend Fade of 4H Dislocations With EXPLICIT Target>Stop Geometry Gate, Mild Daily-Regime Filter, Flip-Via-Flat, Fixed-Fractional Sizing — Grounded in the 4H-Momentum-Breakeven Evidence That BTC REVERTS at 4H While It TRENDS Daily, Mechanism-Distinct/Low-Correlation to the Promoted Daily-Momentum Book, Fills Long-Short + HL Venue, Tail-Safe, Low-Parameter)

A LONG-SHORT, SINGLE-INSTRUMENT, 4-HOUR mean-reversion strategy on BTCUSD.HYPERLIQUID that fades stretched 4H dislocations from a 4H VWAP/EMA on both sides, exiting on reversion to the mean. The thesis is EVIDENCE-GROUNDED, not assumed: the HL BTC 4H MOMENTUM book came back BREAKEVEN (PF 0.99, Sharpe -0.05) while the HL BTC DAILY TS-momentum was strong enough to PROMOTE (Sharpe 0.68, PF 1.59). Momentum working at daily but NOT at 4H is direct evidence that at the 4H timescale BTC does not continue — it REVERTS. So a 4H counter-trend fade on BTC targets a genuinely different, low-correlation edge from the promoted daily-momentum book (counter-trend vs trend), satisfying the analyst's directive to 'pursue 4H on a DIFFERENT mechanism/instrument, not another BTC momentum clone.' It is on the CONFIRMED-CLEAN instrument (HL BTC sizes 90-97%; over-exposure localized to BINANCE + HL SOL) and the CONFIRMED-DEEP/evaluable cadence (HL 4H ~1.3yr, 178-trade class, 15-day holdout ~90 bars — the validatable sub-daily cadence, NOT the 2-month sub-hourly wall, NOT the daily holdout-sparsity wall). Paired with the pending HL ETH 4H reversion, it tests whether 4H reversion is a venue-wide edge on the two clean majors (per-instrument validation). It fills the dominant direction gap (long_only 85.9% vs <=55%) and the under-weight HL venue (7.4% vs >=20%). CRITICAL GEOMETRY FIX (carried from what works): every prior VWAP-reversion failed on INVERTED geometry (avg_win < avg_loss); this uses an EXPLICIT gate — only enter when distance(price, mean target) >= geom_mult x stop_distance (geom_mult > 1), net of the ~0.09% HL round-trip — so avg_win > avg_loss BY CONSTRUCTION. MILD REGIME FILTER: skip counter-trend fades against a VERY strong opposing daily trend (don't short-fade in a strong daily uptrend, don't buy-dip in a strong daily downtrend) to avoid the knife-catch tail, while still allowing reversion in neutral/normal regimes (the dominant 4H state). Bug-avoiding: pure OHLCV, ONE fixed-fractional size per entry, NO dynamic resizing, leverage 1.0, dual flat guard, flip-via-flat. PRE-REGISTERED ACCEPTANCE: (1) GEOMETRY — realized avg_win > avg_loss, else the gate failed -> abandon. (2) EDGE-vs-REDUNDANCY — must show low correlation / distinct return profile vs the promoted daily-momentum book (if it merely re-expresses daily momentum, abandon as redundant). (3) SIZING — avg_position_pct ~95-100%; mild benign winning-leg notional drift acceptable (4H BTC momentum showed 120.9% with plausible returns), but >~150% with impossible single-period returns -> escalate. 4H depth already confirmed (no data-wall kill-switch needed). Low parameter count (mean lookback, stretch threshold, geom_mult, regime EMA, ATR stop, time stop).

Implements the hypothesis as an evidence-grounded counter-trend complement: HL BTC momentum was breakeven at 4H (PF 0.99) but promoted at daily (Sharpe 0.68), which is direct evidence BTC reverts at the 4H timescale, so a 4H fade targets a genuinely different, low-correlation edge from the promoted daily-momentum book (satisfying the analyst's 'different mechanism, not another BTC momentum clone' directive). It runs on the confirmed-clean instrument (HL BTC sizes 90-97%; over-exposure localized to BINANCE + HL SOL) and the confirmed-deep, evaluable cadence (HL 4H ~1.3yr, ~90-bar 15-day holdout — avoiding both the sub-hourly 2-month wall and the daily holdout-sparsity wall), filling the dominant long-only direction gap and the under-weight HL venue. The CRITICAL geometry fix from prior inverted-geometry VWAP failures is made structural: the explicit gate target_dist >= geom_mult * stop_dist is checked before any entry, so every trade taken has its reversion target larger than its stop and avg_win > avg_loss by construction (reinforced by a min_target_pct floor that clears fees). The mild, asymmetric regime filter cuts the knife-catch tail by disabling only the side facing a very strong opposing trend while preserving two-sided reversion in the dominant neutral 4H state. Construction stays defect-avoiding: single-instrument pure OHLCV (no funding, no extra legs, no options), LONG/SHORT with flips routed through flat structurally (the base engine evaluates an entry only when flat, so opposing fills never net), ONE fixed-fractional size (notional = 0.95 * equity, qty = notional/price) with no dynamic resizing, leverage 1.0 so the engine margin model and in-strategy notional agree (avg_position_pct ~95-100%). Pre-registered acceptance: (1) realized avg_win > avg_loss or the gate failed -> abandon; (2) must show low correlation / distinct return profile vs the promoted daily-momentum book or abandon as redundant; (3) sizing avg_position_pct ~95-100% with benign drift acceptable, but >~150% with impossible single-period returns -> escalate.

Abandon — over a healthy, fully-validatable sample (219 trades, ~1.3yr deep 4H data, clean sizing) the strategy is a net loser with negative per-trade expectancy (total_return -32.6%, Sharpe -0.83, profit_factor 0.80, avg_trade_return -$173, max_drawdown 54.8%), which is a robust no-edge result, not noise. The three pre-registered gates pass only technically/hollowly and do not rescue it: (1) GEOMETRY passes at parity — avg_win $1561 vs avg_loss $1553 (1.005x) — so the book needs ~50% win rate but gets 44.3% and therefore loses; the explicit geom_mult ENTRY gate again fails to deliver realized avg_win>avg_loss because the EXITS (18-bar time stop, percentage stop, and the VWAP mean drifting toward price) truncate winners back to roughly the stop distance — the same exit-geometry flaw that sank the prior VWAP-reversion books, unfixable by an entry filter. (2) The CENTRAL THESIS is directly contradicted by the result: 'momentum was breakeven at 4H, therefore BTC reverts at 4H' does NOT follow — reversion is also a loser (PF 0.80), so both momentum (PF 0.99) and reversion (PF 0.80) lose at 4H BTC, meaning the 4H timescale simply has no tradable autocorrelation edge net of the ~0.09% round-trip; optimization cannot manufacture an edge that is absent in both directions. The recent sub-period is positive (2025 +24.4%, 2026 +3.9%) and the cadence is genuinely validatable, which makes this the most borderline review of the batch, but the full-period negative expectancy (driven by 2024 -52%), the parity geometry, and the contradicted premise make 2 hours of optimization poor EV — an optimizer would curve-fit the 2025-2026 regime and likely die on DSR/PBO post-optimization. Not iterate: there is no code-level bug (the geometry fix is present and the construction is clean); the failure is structural (reversion-exit truncation + no autocorrelation edge at 4H BTC), not a parameter the developer can change. PRESERVED FINDINGS for the research lead: (a) 4H BTC is autocorrelation-efficient net of fees — BOTH trend (PF 0.99) and counter-trend (PF 0.80) fail, so stop proposing single-asset 4H BTC directional-autocorrelation books (momentum OR reversion); a 4H edge on BTC would need a different signal class (e.g. volatility, funding, cross-sectional), not faster/slower trend/reversion; (b) the VWAP-reversion family's inverted/parity geometry is NOT fixable by an entry-side geom_mult gate — it requires an EXIT redesign that lets winners run to the full mean target without time-stop/mean-drift truncation while keeping the stop tight; until that is solved, further VWAP-reversion clones should be deprioritized.