EthHyperliquidOneHourDailyVwapExcursionMeanReversionLongShort

ETH HYPERLIQUID Perp 1H Daily-Anchored VWAP Excursion LONG-SHORT Mean-Reversion (Single Asset, 3-Parameter, Volatility-Scaled Stop, End-of-Day Time Exit)

A symmetric LONG-AND-SHORT intraday mean-reversion strategy on ETHUSD.HYPERLIQUID perpetual using 1-HOUR bars, transplanting the daily-anchored-VWAP excursion mechanism that the analyst flagged as 'genuinely good' (the ETH-1H Binance sibling was promoted) onto an under-represented venue and adding a short side. Mechanism: on each 1H bar, maintain a session-anchored VWAP that resets at 00:00 UTC, plus a rolling intraday dispersion band (std-dev of price-minus-VWAP since the daily anchor, i.e. a VWAP z-score). When price extends FAR ABOVE VWAP (z >= entry_z) the strategy SHORTS the excursion; when price extends FAR BELOW VWAP (z <= -entry_z) it goes LONG — in both cases betting on reversion back toward the volume-weighted fair value within the same session. This is fundamentally distinct from the abandoned 'short-at-recent-high' carry baskets (which short multi-DAY highs and accumulate adverse funding for weeks): here the holding period is intraday, the anchor is a continuously-recomputed fair value that travels with price, and the position is flat by session end — so it captures microstructure overshoot reversion, not a multi-day directional bet against the trend. It deliberately fills three under-represented buckets simultaneously: HYPERLIQUID venue (4.9% vs 20% target), long+short direction (12.6% vs the >45% neutral/short room), and short intraday horizon (2.2% vs 10% target). Targets 1-3 trades/day, ~0.6-1.2% gross reversion per trade vs ~0.09% HL round-trip cost. EXPLICIT DATA REQUIREMENT: ETHUSD.HYPERLIQUID 1H must have >= 12 months of history for a valid 3-window walk-forward + 15-day holdout; ETH has traded on Hyperliquid since 2023 (~2.5-3yr) and on-demand collection (3yr for HL) fetches it — this avoids the 38-day SOL-15m data-insufficiency that killed the prior HL transplant.

Iteration 3 targets the geometry blocker (net-losing base: PF 0.87, avg_loss $170 >> avg_win $113). Root cause per feedback: the prior stop was in ATR (1H true-range) units while the profit target was in z (intraday dispersion-std) units, so a stopped-out loser (~2 ATR) could be ~1.5x a captured winner. Smallest structural fix, taking the feedback's third suggested option: make the stop scale to the SAME z-band as the entry. I record the entry z-level (self._entry_z) and replace the primary stop with a z-band stop — a short entered at z0 stops at z0+stop_z_band, a long at z0-stop_z_band. Because the winner span is |entry_z|+exit_overshoot_z sigma and stop_z_band is clamped strictly below that (default 1.3 vs winner 2.4 sigma), a loser is by construction smaller than a winner (~0.54x here). At the observed 56.7% win rate that flips gross expectancy positive (0.567*2.4 - 0.433*1.3 = +0.80 sigma) before fees, resolving the avg_loss>avg_win asymmetry that the overshoot exit alone could not. The ATR stop is retained only as a wide catastrophe backstop (default mult widened 2.0->3.5) so the z-band governs normal exits. All entry/signal scaffolding that already passed Layers 1-3 is unchanged (calculate_signal still returns the continuous z, entry gates/time exit untouched), so no earlier layer regresses. On data blocker (1): venue stays HYPERLIQUID as the hypothesis requires; on-demand HL collection fetches ~3yr for ETHUSD, so the 140-day window should backfill to the full history needed for the 3-window walk-forward + 15-day holdout on the next backtest fetch.

Failed deflated Sharpe and every forward gate on optimization attempt 1. The base backtest LOSES money (Sharpe -0.999, PF 0.90, negative expectancy) and the z-band stop redesign FAILED its stated purpose — avg_loss ($179) still exceeds avg_win ($116). The walk-forward is is_overfitted=TRUE with avg IS Sharpe 9.91 collapsing to avg OOS Sharpe EXACTLY 0.0 across all three windows [0.0, 0.0, 0.0]; the holdout 'pass' is spurious (ratio=0, holdout_sharpe 11.6 divided by a 0.0 WF-OOS baseline). It fails multiple-testing deflation decisively: DSR=0.0225 (vs 0.95), the optimized Sharpe 4.66 is far below the 225-trial expected-max luck bar of 13.39 (is_significant=false, PBO 0.6164, sharpe_ci_low -2.08 straddling 0), and PSR 0.858 vs DSR 0.0225 is the textbook selection-inflation trap. The optimized config is degenerate — entry_z pushed to 2.66 produces only 19 trades over 140 days at 2.42% exposure, so the flattering Sharpe 4.66 is a 19-sample best-of-225 artifact — and entry_z is a confirmed cliff parameter whose entire sensitivity grid is negative except one extreme corner. Not iterate (despite attempt 1 of 2): the binding failure is all-zero out-of-sample on a losing base with a cliff entry parameter and no robust region to tune toward — a second best-of-225 sweep re-selects another overfit, sub-significant config. Not revise_hypothesis: per the overfit->abandon rule this is high-IS/zero-OOS overfitting on a money-losing base, not a proven mechanism stranded on a dead target (the promoted Binance ETH-1H sibling remains the configuration where this VWAP-excursion mechanism clears the bar; adding a short side on HYPERLIQUID does not inherit its promotability). FAILURE PATTERN: transplanting the promoted ETH-1H daily-anchored-VWAP-excursion mechanism onto HYPERLIQUID and bolting on a symmetric short side does NOT inherit promotability — the long+short variant loses money at base (PF 0.90, avg_loss>avg_win), its walk-forward OOS is identically zero, its holdout ratio is 0 against a zero baseline, and its best-of-225 optimized Sharpe (4.66, DSR 0.0225) sits far below the 13.39 luck bar on a 19-trade degenerate config. A satisfied data requirement and a novel z-band symmetric stop do not rescue a variant that fails deflated Sharpe with an empty out-of-sample.

ETHUSD.HYPERLIQUID 1H daily-anchored VWAP excursion mean-reversion, symmetric long+short. Resets a volume-weighted session VWAP at 00:00 UTC and measures the z-score of price-minus-VWAP over the day's elapsed bars. Shorts when z>=entry_z (price far above fair value) and longs when z<=-entry_z (far below), betting on intraday reversion. Exits on the earliest of: a VWAP-overshoot target (z reverts through and past fair value by exit_overshoot_z), a z-band stop measured in the SAME dispersion-std units as the entry (loser span = stop_z_band sigma, clamped strictly below the winner span so a stopped-out loser is structurally smaller than a captured winner), a wide ATR catastrophe backstop, or a forced end-of-day flat at 23:00 UTC before the anchor reset. Leverage 1.0, OHLCV-only, HYPERLIQUID venue.

Mechanism correctly implements the hypothesis: 177 trades over 140 days (~1.3/day, in the 1-3/day target), symmetric 77 long / 100 short, intraday holding (avg 6h25m), flat by session end — genuine VWAP-excursion reversion, not a multi-day carry bet.

Clean entry diagnostics: 177 signaled -> 177 submitted, zero drops for size/notional/cash.

Adequate sample (177 trades) for a 3-window walk-forward, and the mechanism is proven on the promoted ETH-1H Binance sibling.

Failure mode is near-breakeven and tunable (PF 0.905, win rate 58%, short-side win rate 60%) — not degenerate or frozen.

Base is net negative: Sharpe -1.0, PF 0.905, negative expectancy, both 2025 and 2026 negative.

The specifically-engineered z-band stop FAILED its stated purpose: avg_loss 179 > avg_win 116 (largest_loss -746 vs largest_win +453). The wide 3.5xATR catastrophe backstop and forced end-of-day time exit are producing the oversized losers, defeating the z-band design.

Only 140 days of HL data materialized vs the >=12 months the hypothesis explicitly required — WF windows + 15-day holdout will be thin.

Data requirement was met (140 days, 5317 bars, 1620 funding events) — the SOL-15m data-insufficiency that killed the prior HL transplant did not recur

Mechanism fires as designed (177 base trades, balanced 77 long / 100 short, ~6h holding, flat by session end) — the code implements the intraday overshoot-reversion hypothesis, not a multi-day carry bet

Base backtest LOSES money: Sharpe -0.999, total_return -1.29%, profit_factor 0.90, negative expectancy

The z-band stop redesign FAILED at its stated purpose — avg_loss ($179) still exceeds avg_win ($116) despite 58% win rate, so losers remain structurally larger than winners (the exact PF<1 pathology it was meant to fix)

Walk-forward is_overfitted=TRUE: avg IS Sharpe 9.91 collapses to avg OOS Sharpe 0.0 in ALL THREE windows [0.0, 0.0, 0.0] — no out-of-sample edge whatsoever

Holdout 'pass' is spurious: ratio=0 because holdout_sharpe 11.6 is divided by a 0.0 WF-OOS baseline — cannot validate a strategy with zero OOS edge

Fails deflated Sharpe decisively: DSR=0.0225, optimized Sharpe 4.66 far below the 225-trial expected-max luck bar of 13.39, PBO=0.6164, sharpe_ci_low=-2.08 straddling 0, is_significant=false

Optimized config is degenerate: entry_z=2.66 yields only 19 trades over 140 days at 2.42% exposure — the Sharpe 4.66 is a 19-sample best-of-225 artifact

entry_z is a confirmed cliff parameter; the entire sensitivity grid is negative except the extreme entry_z=2.4 corner — no robust region to tune toward

Heavy fee drag: commission 13.2% of gross, turnover 63x, plus 8.9% market-impact cost on the optimized run

Do NOT optimize yet — two blockers. (1) DATA: only 140 days of ETHUSD.HYPERLIQUID-1H loaded, but the hypothesis explicitly requires >=12 months for a valid 3-window walk-forward + 15-day holdout and warned about the prior 38-day HL transplant failure. Verify on-demand collection actually fetches the full ~2.5-3yr HL ETH history (check the catalog / ensure_bars_available for ETHUSD.HYPERLIQUID-1H); if only 140 days truly exist, this transplant cannot be validated on HL and should be reframed. (2) GEOMETRY: the base is net-losing (PF 0.87, expectancy -$9.24) with win rate 56.7% but avg_loss $170 >> avg_win $113. The overshoot-exit you added to fix the PF 0.97 asymmetry regressed it to 0.87 — the stop is too wide relative to the target. Before optimization, re-examine the stop/target relationship: either tighten stop_atr_mult, widen exit_overshoot_z, or reconsider whether the ATR stop should scale to the same z-band as the entry so a stopped-out loser isn't ~1.5x a captured winner. Confirm the reversion edge is positive on a full-history window first.

This strategy took the daily-anchored VWAP-excursion mechanism from a promoted ETH-1H Binance sibling and transplanted it onto the under-represented HYPERLIQUID venue with an added short side, aiming for intraday overshoot reversion that is flat by session end. It cleared the data hurdle that killed a prior HL transplant (140 days, 177 balanced trades, ~6h holding) and passed the backtest-review gate with an 'optimize' verdict despite a losing base — but the specifically engineered z-band stop failed its purpose, with average losses still larger than average wins (PF 0.905, Sharpe -0.999). Optimization confirmed the worst case: in-sample Sharpe of 9.91 collapsed to exactly 0.0 out-of-sample in all three walk-forward windows, deflated Sharpe was 0.0225, and the flattering optimized Sharpe rested on a degenerate 19-trade config. After 3 iterations the analyst abandoned it as high-in-sample, zero-out-of-sample overfitting on a money-losing base with a cliff-like entry parameter and no robust region to tune toward.

Transplanting a promoted mechanism onto a new venue and bolting on a symmetric short side does not inherit promotability — the long+short HYPERLIQUID variant lost money at base and showed zero out-of-sample edge, so a satisfied data requirement and a novel z-band stop can't rescue a mechanism with no forward signal.

The Strategy Analyst issued an abandon verdict at the post-optimization ANALYZING stage: the walk-forward was flagged is_overfitted=TRUE with average in-sample Sharpe 9.91 collapsing to exactly 0.0 out-of-sample across all three windows, and it failed deflated Sharpe (0.0225 vs 0.95), PBO (0.6164), and significance decisively.

A symmetric long+short intraday mean-reversion strategy on ETHUSD.HYPERLIQUID 1H perps that shorts price excursions far above and buys excursions far below a daily-anchored session VWAP (z-score gated), betting on reversion to volume-weighted fair value while staying flat by session end.

The base backtest lost money over 140 days and 177 trades: Sharpe -0.999, total return -1.29%, profit factor 0.905, with a 58% win rate undone by average losses ($179) exceeding average wins ($116). Optimization did not rescue it — the best-of-225 config produced only 19 trades at 2.42% exposure with a flattering but unreliable Sharpe of 4.66 (deflated Sharpe 0.0225).