EthHyperliquid1hVwapMeanReversionLongShort

Hyperliquid ETH 1-Hour Mean-Reversion to VWAP/EMA, LONG/SHORT (Single-Instrument HYPERLIQUID Perp, Pure OHLCV, Two-Sided Fade With EXPLICIT Target>Stop Geometry Gate Fixing the Recurring Inverted-MR Trap, Flip-Via-Flat, Regime-Gated, Highest-Holdout-Density Non-Sub-Hourly Cadence, Hard 1H-DATA-DEPTH Kill-Switch — Mechanism-Matched ETH Complement to the Pending BTC-1H-Momentum Probe, Fills Long-Short + HL Venue, Tail-Safe, Low-Parameter)

A LONG-SHORT, SINGLE-INSTRUMENT, 1-HOUR mean-reversion strategy on ETHUSD.HYPERLIQUID that fades stretched dislocations on both sides (buy stretched-below-mean, short stretched-above-mean) and exits on reversion toward an intraday mean (rolling VWAP / EMA). It is deliberately the ONE non-redundant slot left in the clean-evaluable corner: three HL cadence probes are already pending (BTC 1H momentum-LS, BTC 4H momentum-LS, ETH daily high-turnover reversal), so another momentum/breakout probe would duplicate them. This is the MECHANISM-MATCHED ETH complement at the highest-holdout-density non-sub-hourly cadence. Rationale chain from the fully-mapped constraints: only HL BTC/ETH size cleanly (BINANCE USD-M/COIN-M over-exposure-broken, spot-CASH MTM broken, multi/options broken, census data-walled, HL SOL data-poisoned); HL daily is deep but the 15-day holdout is too sparse (~15 bars); HL 15-min is data-walled (~2mo, banned, 5x). The remaining hope is a deep intraday cadence; 1H is the BEST holdout-density candidate (~24 bars/day -> ~360 holdout bars vs ~90 at 4H, ~15 daily) and is explicitly NOT sub-hourly. ASSET-SELECTIVITY: ETH lacked momentum/breakout edge (it CHOPS), so the ETH-appropriate mechanism is reversion (the pending BTC-1H book correctly uses momentum on the trending asset; this uses reversion on the choppy one). CRITICAL DESIGN FIX: every prior VWAP-reversion run failed on INVERTED payoff geometry (avg_win < avg_loss: ETH long/flat 567<678, ETH LS 467<664, SOL fade) because the mean-touch target was smaller than the ATR stop. This is fixed with an EXPLICIT GEOMETRY GATE: only enter when the distance from price to the reversion target (the mean) is >= geom_mult x the stop distance (geom_mult > 1), net of the ~0.09% HL round-trip, so avg_win > avg_loss BY CONSTRUCTION; entries failing the gate are skipped. Bug-avoiding: pure OHLCV, ONE fixed-fractional size per entry, NO dynamic resizing, leverage 1.0, dual flat guard, flip-via-flat. REGIME GATE prevents fading genuine trends (no dip-buy in strong downtrend, no rip-fade in strong uptrend). Fills the dominant direction gap (long_only 85.9% vs <=55%) and under-weight HL venue (7.2% vs >=20%). PRE-REGISTERED ACCEPTANCE (hard kill-switches, evaluated FIRST, shared with the BTC-1H probe to avoid thrash): (1) DATA-DEPTH — 1H HL ETH data must span >= ~2 years AND the 15-day holdout contain >= ~10 trades; else the sub-daily HL data wall extends to 1H -> ABANDON & ESCALATE (NOT iterate, NOT optimize). (2) GEOMETRY — realized avg_win > avg_loss must hold; if not, the geometry gate failed -> abandon. (3) SIZING — avg_position_pct ~<=100%, no impossible intraday equity jump; else abandon & escalate as reproducer. Low parameter count (mean lookback, stretch threshold, geom_mult, regime EMA, ATR stop, time stop).

Implements the hypothesis as the one non-redundant clean-evaluable slot: three HL cadence probes are already pending (BTC 1H momentum-LS, BTC 4H momentum-LS, ETH daily high-turnover reversal), so this is the mechanism-matched ETH complement — reversion on the choppy asset (ETH had no momentum/breakout edge) at the highest-holdout-density non-sub-hourly cadence (1H ~ 24 bars/day -> ~360 holdout bars vs ~90 at 4H and ~15 daily), explicitly outside the confirmed ~2-month sub-hourly data wall. The central, recurring failure of every prior VWAP-reversion run was inverted payoff geometry (avg_win < avg_loss because the mean-touch target was smaller than the ATR stop); this is fixed structurally, not by parameter tuning: because price reverts to the mean, the favorable target distance equals score x ATR while the stop is atr_stop_mult x ATR, so requiring score >= geom_mult x atr_stop_mult forces target >= geom_mult x stop and avg_win > avg_loss by construction. The mean-touch is deliberately the wide target with no tighter take-profit that would re-invert the geometry. The regime band prevents fading genuine trends on either side. LONG/SHORT fills the dominant direction gap on the under-weight, proven-clean HL ETH path, and flips are routed through flat structurally (the base engine enters only when flat) so the engine never nets opposing fills. Construction stays defect-avoiding: single-instrument pure OHLCV (no funding, no extra legs, no options), ONE fixed-fractional size (notional = 0.95 x equity, qty = notional/price) with no dynamic resizing, leverage 1.0 so the engine margin model and in-strategy notional agree. Pre-registered kill-switches (evaluated first, to avoid thrash): (1) DATA-DEPTH — 1H HL ETH must span >=~2yr and the 15-day holdout contain >=~10 trades, else the sub-daily HL data wall extends to 1H -> abandon & escalate; (2) GEOMETRY — realized avg_win > avg_loss must hold, else the gate failed -> abandon; (3) SIZING — avg_position_pct <=~100% with no impossible intraday equity jump, else abandon & escalate as a reproducer.

Abandon & ESCALATE to the data engineer — two independent pre-registered gates fail, either sufficient, and the hypothesis's shared kill-switch specifies abandon-and-escalate (NOT iterate, NOT optimize) on data-depth failure. (1) DATA-DEPTH: 1H HL ETH data spans only ~7 months (2025-10-11 to 2026-05-20, 5244 bars, data_days 186), far short of ~2 years — identical to the BTC-1H sibling, confirming the 1H HL catalog is unbackfilled for BOTH majors while 4H goes back ~2.3yr (2024-01). A single-regime 7-month sample cannot power a 3-window walk-forward. ACTIONABLE ESCALATION (one fix covers both 1H probes): backfill BTCUSD.HYPERLIQUID and ETHUSD.HYPERLIQUID 1H bars to ~2023 (Hyperliquid's API has the history); until then 1H HL strategies are unvalidatable. (2) GEOMETRY: the explicit geom_mult target>stop gate — the hypothesis's central design fix — FAILED in realized terms (avg_win $1466 < avg_loss $1564), the same inverted-MR trap as the prior VWAP-reversion runs. ROOT CAUSE / lesson for the developer & research lead: gating ENTRY geometry (target distance >= geom_mult*stop at entry) does NOT guarantee realized avg_win>avg_loss, because the EXITS truncate winners — the 24-bar time stop, the percentage stop, and especially the VWAP mean drifting toward price all close winners before they traverse the full target distance, while losers still run to the fixed ATR stop. To actually realize avg_win>avg_loss a reversion book must hold winners to the full target (no early time stop, a mean that doesn't chase price) AND keep the stop tight — entry gating alone cannot fix exit-driven geometry. The standalone result is also a catastrophic loser (Sharpe -2.26, PF 0.84, avg_trade -$131, fees 12.6% of gross), but per the pre-registration the verdict is data-depth + geometry abandon regardless. PRESERVED FINDING: this is the SECOND 1H probe to confirm the ~7-month 1H HL data wall — the research lead should STOP queuing 1H HL strategies until the data engineer backfills 1H to ~2023; deep-history HL work must use 4H or daily for now. And the recurring VWAP-reversion inverted-geometry failure is now shown to survive an explicit entry-geometry gate, so future reversion attempts must redesign the EXIT path, not just the entry filter.

Long/short 1-hour two-sided mean-reversion-to-VWAP strategy on ETHUSD.HYPERLIQUID, pure OHLCV. Computes a rolling volume-weighted mean (VWAP over 24 1H bars) and a stretch score = (VWAP - close)/ATR: positive when price is stretched below the mean (buy the dip), negative when stretched above (fade the rip / short). The defining feature is an EXPLICIT target>stop geometry gate: entry is allowed only when the reversion-target distance (|score| x ATR to the mean) is at least geom_mult (1.75) times the stop distance (atr_stop_mult x ATR), i.e. required stretch = max(stretch_threshold, geom_mult x atr_stop_mult) = 1.75 ATR — so avg_win > avg_loss by construction, fixing the inverted-MR trap that killed prior VWAP books. A 72-bar regime EMA with a 6% band blocks dip-buying in strong downtrends and rip-fading in strong uptrends. Exits on reversion to the mean (the wide target), a narrow 1.0-ATR / 3% stop, or a 24-bar time stop; a 2-bar cooldown throttles churn. Long/short with flips routed through flat, one fixed-fractional position (95% of equity notional) per entry, no dynamic resizing, leverage 1.0.

Code runs and trades both sides cleanly (277 trades, 133 long / 144 short, metrics_reliable=true); sizing roughly in range (avg_position_pct 108.2%, modest MTM drift, no impossible jumps).

Correctly identifies 1H as the highest-holdout-density non-sub-hourly cadence and is mechanism-matched (reversion on choppy ETH) — the design logic is sound; only data length and realized exit geometry defeat it.

DATA-DEPTH KILL-SWITCH FAILED: 1H HL ETH data spans only ~7 months (2025-10-11 to 2026-05-20, 5244 bars) vs the required ~2 years — single-regime, cannot power a 3-window walk-forward. Identical to the BTC-1H sibling: the 1H catalog is unbackfilled.

GEOMETRY GATE FAILED in realized terms: avg_win $1466 < avg_loss $1564 despite the explicit geom_mult entry gate — the inverted-MR trap persists because entry-geometry gating does not control EXIT geometry (24-bar time stop, % stop, and VWAP mean drifting to price truncate winners while losers run to the ATR stop).

Catastrophic performance: total_return -15.3%, Sharpe -2.26, Sortino -3.09, profit_factor 0.84, avg_trade_return -$131 (negative), win_rate 47%. Fee drag 12.58% of gross.

EthHyperliquid1hVwapMeanReversionLongShort was the mechanism-matched ETH reversion complement to the BTC 1H momentum probe, and its centerpiece was an explicit target>stop entry-geometry gate meant to finally fix the inverted payoff that had sunk every prior VWAP-reversion run. It failed on two fronts at once: the shared data-depth kill-switch fired because 1H HL ETH history runs only ~7 months (confirming, with the BTC sibling, that the 1H catalog is unbackfilled while 4H reaches ~2.3yr), and the geometry gate failed in realized terms (avg_win $1,466 < avg_loss $1,564) because gating entry distance cannot stop exits — the time stop, percentage stop, and a mean that chases price — from truncating winners while losers run to the ATR stop. The standalone result was a catastrophic loser (Sharpe -2.26, PF 0.84), but per the pre-registration the verdict rests on the two gate failures. The reviewer abandoned and escalated to the data engineer to backfill 1H BTC/ETH to ~2023, leaving two forward-looking findings: stop queuing 1H HL strategies until then, and future reversion attempts must redesign the exit path rather than just the entry filter.

Gating entry geometry does not guarantee realized avg_win>avg_loss because exits truncate winners — a 24-bar time stop, a % stop, and a VWAP mean drifting toward price all close winners early while losers run to the fixed ATR stop — so reversion books must redesign the exit path, not just the entry filter; separately, 1H HL data is unbackfilled (~7 months) for both majors, so deep-history HL work must use 4H or daily until backfill.

The analyst abandoned and escalated at the backtest-review gate on two independent pre-registered failures: the data-depth kill-switch (1H HL ETH spans only ~7 months / 5,244 bars, far short of ~2 years, so a 3-window walk-forward cannot be powered — identical to the BTC 1H sibling) and the geometry gate (realized avg_win < avg_loss despite the explicit entry filter).

A long/short, single-instrument, OHLCV-only 1-hour two-sided mean-reversion-to-VWAP strategy on ETHUSD.HYPERLIQUID (regime-gated, flip-via-flat, leverage 1.0) built around an explicit entry-time geometry gate — only fading when the reversion-target distance is >= geom_mult x the stop distance — intended to fix the recurring inverted-MR trap, as the mechanism-matched ETH complement to the BTC 1H momentum probe at the highest-holdout-density non-sub-hourly cadence.

It was a catastrophic loser over the available ~7-month window and 277 trades (133 long, 144 short): -15.3% total return, Sharpe -2.26, Sortino -3.09, profit factor 0.84, 47% win rate, avg_trade_return -$131, fees 12.58% of gross — and critically the geometry gate failed in realized terms (avg_win $1,466 < avg_loss $1,564); sizing roughly passed (avg_position_pct 108.2%, metrics_reliable=true).