EthHyperliquidDailyShortHorizonReversalLongShort

Hyperliquid ETH Daily Short-Horizon (2-3 Day) Reversal, LONG/SHORT (Single-Instrument HYPERLIQUID Perp, Pure OHLCV, DEEP Daily Data, High-Turnover 1-2 Day Holds to Defeat the 15-Day-Holdout Sparsity Wall, Flip-Via-Flat, Corrected Payoff Geometry, Regime-Gated — Reversion Matched to ETH's Chop, Fills Long-Short + HL Venue, Pre-Registered Holdout-Density + Sizing Kill-Switch, Tail-Safe, Low-Parameter)

A LONG-SHORT, SINGLE-INSTRUMENT, DAILY short-horizon REVERSAL strategy on ETHUSD.HYPERLIQUID that buys after a sharp 2-3 day DOWN move and shorts after a sharp 2-3 day UP move, holding ~1-2 days for the snap-back. It is engineered against the exact, fully-mapped constraint set: every BINANCE venue is engine-broken (USD-M/COIN-M over-exposure, spot-CASH MTM, multi aggregation, options settlement); census signals are data-walled (OHLCV only); HL SOL is data-poisoned; HL sub-hourly (15m) has only ~2 months of data (banned until backfill); and HL DAILY is DEEP (~3yr, 2023+) BUT low-turnover daily trend/breakout books die on the 15-day holdout (32-day holds -> 0-1 holdout trades, auto-reject). This design threads all of that: (1) it uses the CONFIRMED-DEEP HL daily data (NOT the 15m wall); (2) it is HIGH-TURNOVER (1-2 day holds) so even a 15-day holdout contains several trades and each walk-forward OOS window is populated — directly defeating the sparsity wall that killed the daily breakout; (3) it is REVERSION, not trend, so it is outside the banned/failed HL daily-TREND family and is mechanism-matched to ETH's demonstrated CHOP (ETH had no daily trend/breakout edge, 1.05x geometry — the exact condition short-horizon reversal exploits); (4) it is on a PROVEN-CLEAN instrument (HL ETH sized 64.5%; over-exposure localized to BINANCE + HL SOL); (5) LONG/SHORT fills the dominant direction gap (long_only 85.9% vs <=55%) on HL (under-weight venue 7.1% vs >=20%); (6) payoff geometry is CORRECTED from the prior inverted-MR trap (the 15m ETH VWAP book had avg_win 567 < avg_loss 678) — here the reversion target is set WIDER than the stop, net of the ~0.09% HL round-trip, so avg_win > avg_loss; (7) a REGIME GATE prevents fading genuine multi-week trends on either side. It is DISTINCT from the pending HLETHIntradayVwapReversionLongShort (that is 15m, data-walled; this is DAILY, deep data) and from the HL daily-trend books (reversal, not trend). 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 (two kill-switches, to avoid thrash): (1) HOLDOUT DENSITY — the 15-day holdout AND each walk-forward OOS window must contain >= ~4 trades; if the turnover is too low to populate them, this is the daily-sparsity wall again -> ABANDON & escalate (do NOT iterate). (2) SIZING — avg_position_pct ~<=100% and no physically impossible daily equity jump; if violated, abandon & escalate as a reproducer. Low parameter count (reversal lookback, stretch threshold, regime EMA, ATR stop, reversion target/time stop).

Implements the hypothesis by threading the fully-mapped constraint set: it uses the CONFIRMED-DEEP HL daily data (not the ~2-month 15m wall), on the proven-clean HL ETH instrument (sized 64.5%; over-exposure localized to BINANCE venues and HL SOL), and is mechanism-matched to ETH's demonstrated chop — reversion, not the trend/breakout family that had no daily edge on ETH (1.05x geometry). The defining design choice is HIGH TURNOVER: short 1-3 day holds (time stop + reversion-decay exit) ensure even a 15-day holdout and each walk-forward OOS window contain several trades, directly defeating the daily-sparsity wall that auto-rejected the low-turnover daily breakout. Payoff geometry is corrected from the prior inverted-MR trap: entry requires a >=2.0-ATR stretch while the hard stop is only 1.5 ATR, so the favorable reversion distance mechanically exceeds the adverse stop distance (avg_win > avg_loss), reinforced by a wide 6% TP vs narrower 4% stop, all comfortably clearing the ~0.09% HL round-trip. LONG/SHORT fills the dominant direction gap on the under-weight HL venue, and flips are routed through flat structurally (the base engine evaluates an entry only when flat) so the engine never nets opposing fills. The regime band prevents fading genuine multi-week trends on either side while still allowing two-sided fading inside the range. Construction stays defect-avoiding: single-instrument pure OHLCV (no funding, no extra legs, no options), 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. Pre-registered kill-switches: (1) if the 15-day holdout / OOS windows still cannot reach ~4 trades, the daily-sparsity wall persists -> abandon & escalate (do not iterate); (2) if avg_position_pct exceeds ~100% or a physically impossible daily equity jump appears, the over-exposure defect reaches HL ETH daily -> abandon & escalate as a reproducer.

Abandon — the hypothesis's OWN pre-registered HOLDOUT-DENSITY kill-switch fires, and it is not rescuable by tuning. Turnover is far too low: only 43 trades over the whole sample (entry_signaled 43 from 2070 bars ≈ 2% of bars; exposure_pct 4.97%), so the 15-day holdout contains ~0-1 trades and the walk-forward OOS windows are under-populated — the 'high-turnover 1-2 day holds' premise failed because short HOLDS do not create turnover when ENTRIES are rare. The rarity comes from a too-selective gate (stretch_threshold 2.0 ATR ANDed with the regime band), so the optimizer would face the same daily-sparsity wall that auto-rejects on holdout density (the hypothesis pre-registered abandon-and-escalate, NOT iterate, in exactly this case). Critically, this is not a fixable just-loosen-the-threshold problem: the 'corrected' payoff geometry failed AGAIN — avg_win $2492 < avg_loss $3307 with win_rate 37.2% (expectancy -$1149) — i.e. the reversal has no edge even at its MOST selective 2-ATR entries, the highest-quality ones; lowering the threshold to manufacture holdout density would add lower-quality, less-reversion-prone fades and worsen the edge, not create one. The standalone result is also negligible (total_return +13.4% over ~3 active years, CAGR 1.75%, Sharpe 0.77 with sharpe_ci_low -2.85 straddling 0). Not worth 2 hours of optimization: under-populated holdout/OOS windows + a negative-expectancy fade at the selective entries = the daily-sparsity wall with no edge underneath. Not iterate: the hypothesis pre-registered abandon on this exact density failure, and there is no edge to tune toward. PRESERVED FINDINGS for the research lead: (a) the daily-sparsity holdout wall is NOT defeated merely by shortening HOLD time — it requires a mechanism whose ENTRY signal fires frequently (many entries/month), which a 2-ATR-stretch + regime-gated daily fade does not; (b) the short-horizon daily reversal shows no avg_win>avg_loss edge on ETH HL even at selective entries, so the 'reversion matches ETH chop' thesis is not validating at daily cadence; (c) DATA FLAG to the data engineer: ETHUSD.HYPERLIQUID daily data starts 2020-08-19 with ~18 zero-PnL trades pre-2023, before HL existed — likely flat/backfilled pre-launch bars that should be trimmed or verified, as they pollute any pre-2023 walk-forward window.

Long/short daily short-horizon reversal strategy on ETHUSD.HYPERLIQUID using deep daily data, pure OHLCV. Computes a reversal score = (close 3 bars ago - close)/ATR: positive after a sharp 2-3 day DOWN move (fade long), negative after a sharp UP move (fade short). Enters long when score >= +2.0 ATR and the regime permits, short when score <= -2.0 ATR and the regime permits. A regime gate using a 30-day EMA blocks longs when price is more than 10% below the long mean (genuine downtrend) and shorts when more than 10% above it (genuine uptrend); inside the band both sides trade. Exits when the stretch decays back toward the mean (reversion captured), on a wide 6% take-profit, a narrower 1.5-ATR / 4% stop, or a 3-bar time stop. Holds are short (1-3 days) to keep turnover high so the 15-day holdout and walk-forward OOS windows are populated. Long/short with flips routed through flat, one fixed-fractional position (95% of equity notional) per entry, no dynamic resizing, leverage 1.0.

Correctly uses the confirmed-DEEP HL daily data (not the 15m wall) and is mechanism-distinct from the failed HL daily-TREND family (this is reversal). Sizing is clean: avg_position_pct 56.2%, metrics_reliable=true, no impossible daily equity jumps — the HL ETH daily LS path sizes correctly.

Holds are genuinely short (avg 2d 8h), as designed.

PRE-REGISTERED HOLDOUT-DENSITY KILL-SWITCH FIRES: only 43 trades over the whole sample (entry_signaled 43 from 2070 bars ≈ 2% of bars), exposure_pct 4.97%. The 15-day holdout would contain ~0-1 trades and walk-forward OOS windows are under-populated — the 'high-turnover' premise failed because short HOLDS do not create turnover when ENTRIES are rare (stretch_threshold 2.0 ATR + regime gate are too selective).

The 'corrected' payoff geometry failed again: avg_win $2492 < avg_loss $3307 with win_rate 37.2% — a losing reversal even at the most selective 2-ATR entries. Expectancy -$1149 by the avg_win/avg_loss math (metrics internally contradictory — PF 1.34 vs negative expectancy — a tell that 43 trades is too thin to trust).

Negligible edge: total_return +13.4% over ~3 active years, CAGR 1.75%, Sharpe 0.77 with sharpe_ci_low -2.85 (straddles 0).

Suspect pre-2023 data: start_time 2020-08-19 on an instrument that launched ~2023; ~18 of 43 trades (2020-2022) show exactly 0.0 return (flat/backfilled bars) — data-quality issue to flag.

EthHyperliquidDailyShortHorizonReversalLongShort tried to thread the factory's full constraint map: use Hyperliquid's confirmed-deep daily data but hold only 1-2 days so the 15-day holdout would be populated, while fading ETH's chop on both sides. The flaw was conceptual — short holds don't create turnover when entries are rare — and the too-selective 2-ATR + regime gate fired just 43 times across 2,070 bars, so the holdout-density kill-switch tripped exactly as pre-registered. Worse, the payoff geometry failed a third time (avg_win $2,492 < avg_loss $3,307, expectancy -$1,149) at the highest-quality entries, leaving no edge to tune toward, and the standalone result was negligible (CAGR 1.75%). The reviewer abandoned it pre-optimization with three preserved findings: the holdout wall needs frequent entries (not just short holds), daily reversal shows no edge on ETH-HL, and ETHUSD.HYPERLIQUID daily data carries ~18 backfilled pre-2023 zero-PnL bars to flag to the data engineer.

Shortening hold time does not defeat the daily-sparsity holdout wall — turnover requires a signal whose ENTRY fires frequently, which a 2-ATR-stretch + regime-gated fade does not — and short-horizon daily reversal shows no avg_win>avg_loss edge on ETH-HL even at selective entries, so 'reversion matches ETH's chop' is not validating at daily cadence.

The analyst abandoned it at the backtest-review gate, before optimization, because the hypothesis's own pre-registered holdout-density kill-switch fired — only 43 entries in 2,070 bars (~2%) means the 15-day holdout holds ~0-1 trades and OOS windows are under-populated — and the reversal showed no avg_win>avg_loss edge even at its most selective entries, so loosening the gate to add density would only worsen the edge.

A long/short, single-instrument, OHLCV-only daily short-horizon reversal strategy on ETHUSD.HYPERLIQUID that buys after sharp 2-3 day down moves and shorts after sharp up moves (regime-gated, flip-via-flat, leverage 1.0, holds ~1-2 days), engineered to use HL's deep daily data while keeping holds short so the 15-day holdout and walk-forward windows would be populated — defeating the daily-sparsity wall that killed the daily-trend books.

It produced a negligible, thin result: +13.4% total return over ~3 active years (CAGR 1.75%, Sharpe 0.77, sharpe_ci_low -2.85), only 43 trades (21 long, 22 short) at 37.2% win rate and 4.97% exposure, with avg_win $2,492 below avg_loss $3,307 (expectancy -$1,149) despite the intended payoff fix; sizing was clean (avg_position_pct 56.2%, metrics_reliable=true).