EthHyperliquid15mVwapMeanReversionLongShort

Hyperliquid ETH Intraday 15-Minute Mean-Reversion to VWAP, LONG/SHORT (Single-Instrument HYPERLIQUID Perp, Pure OHLCV, Two-Sided Fade of Volatility-Stretched Dislocations, Flip-Via-Flat, Regime/Volatility-Gated, Fee-Gated, Fixed-Fractional Sizing — Mechanism Matched to ETH's Chop, 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 mean-reversion strategy on ETHUSD.HYPERLIQUID that FADES volatility-stretched intraday dislocations on BOTH sides: BUY when price is stretched far BELOW the intraday mean (VWAP / short EMA) and SHORT when stretched far ABOVE it, exiting on reversion toward the mean. It is the LONG-SHORT counterpart of the pending long/flat HL ETH VWAP-reversion book, and is the highest-value remaining EVALUABLE idea because it fills the THREE most under-represented reachable buckets at once: (1) DIRECTION — long_only is 85.8% vs the <=55% target (the single largest quota miss); only a long-SHORT book fixes it, and the existing ETH reversion book is long/flat (dip-only). (2) VENUE — HYPERLIQUID 6.8% vs >=20%. (3) HORIZON — short 1m-15m 3.7% vs >=10%. Mechanism is MATCHED to instrument character (the key asset-selectivity lesson): ETH has NO clean momentum/breakout edge on the HL window (daily breakout failed, 1.05x geometry, net-losing) i.e. it CHOPS — exactly what two-sided mean-reversion monetizes — whereas momentum books belong on BTC. It sits on a proven-clean instrument (HL ETH sized 64.5% on daily; HL BTC/ETH are the only clean HL instruments; HL SOL is data-poisoned and dropped; all BINANCE USD-M/COIN-M perp, spot-CASH, multi-instrument, options paths are engine-broken; census signals data-walled -> OHLCV only). Intraday cadence over ~3yr (~75k 15m bars) yields HUNDREDS of trades (both sides), defeating the HL daily data-sparsity wall that fails daily-cadence validation. It is DISTINCT from HLETHIntradayVwapReversionLongFlat (long/flat dip-only) by adding the short-the-rip side, and from HLBTCIntradayMomentumLongShort (momentum, BTC) by mechanism and asset. FEE DISCIPLINE central (HL ~0.09% RT, need >0.15% net/trade): each entry requires a LARGE stretch (>=~1.75 ATR / >~0.5-0.7% from the mean) so the reversion target (~0.4-0.6%) clears fees; entries throttled (cooldown) — NOT scalping. FLIP-VIA-FLAT: never reverse directly; exit to flat then open the opposite side next bar. REGIME/VOLATILITY GATING limits knife-catching: suppress dip-buys in strong intraday downtrends and rip-fades in strong intraday uptrends (fade only counter-moves against a neutral/range regime). Bug-avoiding: 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, intraday-LS path new): avg_position_pct ~<=100% and no physically impossible intraday equity jump (ETH at 1x ~+/-15-20%); if violated, abandon & escalate as the HL-ETH-intraday-LS reproducer. Low parameter count (mean lookback, stretch threshold, ATR, regime EMA, reversion target).

Implements the hypothesis by fading volatility-stretched intraday dislocations on BOTH sides of a VWAP mean, matching mechanism to instrument character: ETH has no clean momentum/breakout edge on the HL window (it chops), which is exactly what two-sided mean-reversion monetizes, whereas momentum books belong on BTC. It fills the three most under-represented reachable buckets simultaneously — DIRECTION (long_short, the single largest quota miss; this adds the short-the-rip side absent from the long/flat sibling), VENUE (under-weight HYPERLIQUID), and HORIZON (short 1m-15m) — on a proven-clean instrument (HL ETH). Intraday cadence over ~3yr (~75k 15m bars) yields hundreds of two-sided trades, defeating the HL daily data-sparsity wall that fails daily-cadence walk-forward/holdout validation. Fee discipline is enforced as a conjunction of an ATR-stretch gate (>=1.75 ATR) and a percentage-distance floor (>=0.6% from the mean) so the ~0.4-0.8% reversion target clears the ~0.09% HL round-trip, with a cooldown and coarse 15m cadence keeping this out of scalping territory. Directional regime gating (suppress dip-buys when the regime EMA is strongly falling, suppress rip-fades when strongly rising) limits knife-catching while still allowing trend-aligned fades. Construction stays defect-avoiding: single-instrument pure OHLCV (no funding, no extra legs, no options), LONG/SHORT with flips routed structurally through flat so the engine never nets opposing fills, 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. Pre-registered sizing probe: avg_position_pct should be <=~100% with no physically impossible intraday equity jump (ETH at 1x); a violation would indicate the over-exposure defect reaches the HL ETH intraday long-short path and should be abandoned/escalated as the HL-ETH-intraday-LS reproducer.

Abandon on two independent disqualifiers, either sufficient. (1) The hypothesis's OWN pre-registered DATA-DEPTH kill-switch fired: it assumed ~3yr / ~75k 15m bars, but the 15-min HL ETH data spans only 2026-03-16 to 2026-05-27 = 70 days / 6862 bars — the confirmed ~2-month HYPERLIQUID sub-hourly data wall. A single-regime ~2.5-month sample cannot support a 3-window walk-forward + 15-day holdout, so the strategy is structurally unvalidatable at this cadence (the hypothesis itself specified abandon-and-escalate if this gate failed; it did). (2) Independent of the data wall, the base strategy is a catastrophic net loser: total_return -18.6%, Sharpe -9.66, Sortino -10.5, profit_factor 0.46, avg_trade_return -$251 (negative), with the classic mean-reversion avg_win<avg_loss trap (win rate 56.5% but avg_win $382 vs avg_loss $1075, tail_ratio 0.22) and fees devouring the book (commission_pct_of_gross 37.74% — the 'fee-gated, not scalping' design failed). The sizing gate is also breached (avg_position_pct 129.9% > pre-registered 100%). Not worth 2 hours of optimization: there is no parameter region to validate over 2.5 months of single-regime data, and optimizing a -9.66-Sharpe / 0.46-PF loser whose costs alone are 37.7% of gross is hopeless. Not iterate: no code-level fix adds ~3 years of history or converts a negative-expectancy over-trading fade into edge — the construction is sound but the premise (deep 15m HL data + fee-clearing intraday ETH reversion) is false. PRESERVED FINDING for the research lead: the HYPERLIQUID sub-hourly data wall is confirmed at 15-min (~70 days), while 4H HL data is deep (~2.3yr, just confirmed on the BTC 4H review) — STOP proposing sub-hourly (1m-15m) HL strategies as data-walled and unvalidatable; if a shorter-than-daily HL cadence is wanted, use 4H, and pair it with a mechanism that actually clears the ~0.09% HL round-trip without 37%-of-gross fee drag.

Long/short intraday 15-minute two-sided mean-reversion-to-VWAP strategy on ETHUSD.HYPERLIQUID, pure OHLCV. Computes a rolling volume-weighted mean (VWAP over 48 bars) and a stretch score = (VWAP - close)/ATR. Buys a stretched dip (score >= 1.75 ATR and price >= 0.6% below the mean) unless the 96-bar regime EMA has fallen >= 0.5% over the last 8 bars (strong downtrend), and shorts a stretched rip (score <= -1.75 ATR and price >= 0.6% above the mean) unless the regime has risen >= 0.5% over 8 bars (strong uptrend) — i.e. it fades counter-moves only against a neutral/range regime. Both gates require a 4-bar post-exit cooldown so the reversion target clears HL's ~0.09% round-trip fee without churn. Exits on reversion to the mean (close crossing back through VWAP in the position's favor), a 0.8% take-profit, a 2.5% stop-loss, or a 16-bar time stop. Flips are routed through flat (exit fully, then open the opposite side on a later flat bar), one fixed-fractional position (95% of equity notional) per entry with no dynamic resizing, leverage 1.0.

Mechanism is implemented as designed (two-sided VWAP fade, flip-via-flat, regime/cooldown gating, 237 trades with balanced long 128 / short 109 legs) and metrics_reliable=true — the code runs and trades both sides.

Useful confirmatory finding: the HYPERLIQUID sub-hourly data wall is real and bites at 15-min — this localizes the viable HL cadence to 4H+ (just confirmed deep) for future work.

DATA-DEPTH KILL-SWITCH FIRED: the hypothesis assumed ~3yr/~75k 15m bars but the data spans only 2026-03-16 to 2026-05-27 = 70 days / 6862 bars — the ~2-month sub-hourly wall. A single-regime ~2.5-month sliver cannot support a 3-window walk-forward + 15-day holdout, so optimization is not validatable.

Catastrophic performance: total_return -18.6%, Sharpe -9.66, Sortino -10.5, profit_factor 0.46, avg_trade_return -$251 (NEGATIVE). Monthly -32.8% (Apr), -36.8% (May).

The avg_win<avg_loss mean-reversion trap is severe: win rate 56.5% but avg_win $382 vs avg_loss $1075 (tail_ratio 0.22, omega 0.14) — fades pay small and lose big when price keeps trending, exactly the failure mode the hypothesis claimed to avoid.

Fees dominate: commission_pct_of_gross 37.74% (the 'fee-gated, not scalping' claim failed); avg_position_pct 129.9% breaches the pre-registered <=100% sizing gate.

EthHyperliquid15mVwapMeanReversionLongShort was the two-sided counterpart of the earlier long/flat ETH reversion book, designed to fade ETH's chop on both sides and fill three quota gaps simultaneously, carrying a pre-registered data-depth kill-switch in case HL's 15m history proved shallow. It did: only 70 days / 6,862 bars existed against an assumed ~3 years, so the kill-switch fired exactly as specified — and independently, the strategy was a disaster, losing 18.6% with a -9.66 Sharpe and a 0.46 profit factor because, despite a 56.5% win rate, its fades paid small and lost big (avg_win $382 vs avg_loss $1,075) while fees consumed 37.7% of gross and sizing breached the 100% gate at 129.9%. The reviewer abandoned it pre-optimization, reasoning that no parameter sweep can add years of history or convert a negative-expectancy, fee-saturated over-trading fade into edge. The preserved, now firmly established finding is to stop proposing sub-hourly HL strategies as data-walled and unvalidatable, and to use 4H instead when a shorter-than-daily HL cadence is wanted.

A high win rate does not rescue a fade when avg_win is a third of avg_loss — the mean-reversion trap the hypothesis claimed to have fixed reappeared — and HL's sub-hourly (1m–15m) data wall (~2 months) makes 15-minute strategies structurally unvalidatable, so the viable shorter-than-daily HL cadence is 4H, paired with a mechanism that actually clears the ~0.09% round-trip.

The analyst abandoned it at the backtest-review gate, before optimization, on two independent disqualifiers: the hypothesis's own data-depth kill-switch fired (only 70 days / 6,862 bars existed, not the assumed ~3yr/75k — the confirmed ~2-month HL sub-hourly wall, making a walk-forward and holdout impossible), and, independently, the base strategy was a catastrophic negative-expectancy loser with fees devouring the book.

A long/short, single-instrument, OHLCV-only intraday 15-minute two-sided mean-reversion strategy on ETHUSD.HYPERLIQUID that fades volatility-stretched dislocations on both sides (buy stretched dips, short stretched rips) with regime/volatility and cooldown gating, flip-via-flat, leverage 1.0 — matched to ETH's chop and meant to fill the long-short, HL-venue, and short-horizon gaps at once.

It was a catastrophic net loser over 70 days / 237 trades (128 long, 109 short): -18.6% total return, Sharpe -9.66, Sortino -10.5, profit factor 0.46, avg_trade_return -$251 — and despite a 56.5% win rate it fell into the avg_win<avg_loss trap (avg_win $382 vs avg_loss $1,075, tail_ratio 0.22) with fees eating 37.74% of gross and avg_position_pct of 129.9% breaching the pre-registered sizing gate.