EthHyperliquid15mVwapMeanReversionLongFlat

Hyperliquid ETH Intraday 15-Minute Mean-Reversion to VWAP, LONG/FLAT (Single-Instrument HYPERLIQUID Perp, Pure OHLCV, Volatility-Stretch-Gated Dip Buys With a Regime Filter, Fee-Gated, Fixed Single-Entry Sizing, No Flips — Mechanism Matched to ETH's Choppy Character Where Breakout Failed, Fills Under-Weight HL Venue + Short Horizon, Pre-Registered Sizing Probe, Tail-Safe, Low-Parameter)

A LONG-ONLY (LONG/FLAT, never short), SINGLE-INSTRUMENT, INTRADAY 15-MINUTE MEAN-REVERSION strategy on ETHUSD.HYPERLIQUID that buys volatility-stretched intraday dips and exits on reversion toward an intraday mean (VWAP / short EMA). This is deliberately the FIRST non-trend mechanism in the current pending set (which is all trend/breakout/momentum), chosen by matching mechanism to instrument character: the analyst's asset-selectivity finding showed ETH has NO daily breakout/trend edge (weak 1.05x reward:risk, choppy, net-losing) while HL BTC does — so porting another trend/breakout book to ETH would likely repeat that failure, whereas ETH's choppiness is exactly what an intraday mean-reversion edge monetizes. Constraint context that forces this corner of the design space: BINANCE USD-M AND COIN-M perp are both confirmed over-exposure-broken (stop probing them); spot-CASH MTM, multi-instrument aggregation, options settlement are broken; census signals (liq/OI/taker/account-ratio) are data-walled -> OHLCV only; HL is the only clean-sizing venue but NOT uniformly (HL SOL 386% and HL long-short over-exposed; only HL BTC 90% and HL ETH 64.5% are proven-clean), and HL DAILY cadence fails validation on a data-sparsity wall (~3yr history -> 0-1 OOS trades). The analyst's explicit prescription: if HL is used, only INTRADAY strategies accumulate enough OOS trades. A 15-minute ETH strategy over ~3yr (~75k bars) generates hundreds of trades, making a valid 3-window walk-forward + 15-day holdout possible. It fills the TWO most under-represented STILL-REACHABLE buckets: HYPERLIQUID venue (6.7% vs >=20%) and SHORT horizon (1m-15m, 3.6% vs >=10%). FEE DISCIPLINE is central (HL ~0.09% RT, need >0.15% net/trade): entries require a LARGE stretch (>=~1.75 ATR / >~0.5-0.7% below the intraday mean) so the reversion target (~0.4-0.6%) clears fees, and are throttled (no churn on small wiggles, cooldown after exits) — explicitly NOT tick/sub-bar scalping. A REGIME FILTER prevents catching falling knives in sustained downtrends. Construction stays bug-avoiding: pure OHLCV, ONE fixed-fractional size per entry, NO dynamic resizing, NO flips (long/flat), leverage 1.0, dual flat guard. PRE-REGISTERED ACCEPTANCE / SIZING PROBE (retained, HL not assumed uniformly clean): avg_position_pct ~<=100% and no physically impossible intraday equity jump; if violated, the over-exposure defect reaches HL ETH intraday -> abandon & escalate. Low parameter count (mean lookback, stretch threshold, ATR, regime EMA, reversion target).

Implements the hypothesis by matching mechanism to instrument character: the analyst's asset-selectivity finding showed ETH has no daily trend/breakout edge (choppy, net-losing) but its choppiness is exactly what an intraday mean-reversion edge monetizes, so this is the first non-trend book in the pending set rather than another trend port that would likely repeat the ETH breakout failure. Venue/horizon are chosen to be both quota-filling and validation-feasible: HYPERLIQUID is under-weight and is the clean-sizing venue, and HL daily fails on data sparsity (~3yr -> 0-1 OOS trades), so a 15-minute cadence (~75k bars over ~3yr) accumulates the hundreds of trades needed for a valid 3-window walk-forward + 15-day holdout; it simultaneously fills the under-weight SHORT horizon bucket. Fee discipline is central and enforced as a conjunction: an ATR-stretch gate (>=1.75 ATR) AND a percentage-dip floor (>=0.6%) guarantee the reversion target (~0.4-0.8%) clears the ~0.09% HL round-trip with margin, while a post-exit cooldown and a deliberately coarse 15m cadence keep this out of tick/sub-bar scalping. The regime EMA (non-falling requirement) prevents knife-catching in sustained downtrends. Construction stays defect-avoiding: single-instrument pure OHLCV (no funding subscription, no extra legs, no options), LONG/FLAT with no shorts or flips, 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 HL ETH intraday and should be abandoned/escalated.

Three disqualifying failures, before any optimization, none fixable by parameter tuning. (1) PREMISE FALSIFIED — SAME DATA WALL AS THE BTC INTRADAY SIBLING: the hypothesis bet that 15m cadence over ~3yr (~75k bars) would yield hundreds of trades for a valid 3-window walk-forward + 15-day holdout, defeating the HL daily data-sparsity wall. But the run has only ~2.4 months of HL ETH 15m data (2026-03-16 to 2026-05-27, 6862 bars), so the promised validation is impossible — the intraday cadence hit the SAME wall at finer resolution (symptom observed; recommend verifying HL 15m catalog depth before retrying this family). (2) DEGENERATE SAMPLE: only 8 trades (entry_signaled 8), data_days 6, metrics_reliable=FALSE — the 'hundreds of trades' that justified the entire intraday-HL approach did not appear (data wall plus over-restrictive 1.75-ATR + 0.6%-dip + regime + cooldown gating). (3) STRONGLY NEGATIVE EDGE: profit_factor 0.17, Sharpe -9.66, avg_win $251 vs avg_loss $1494 (winners clipped at TP/VWAP while losers run to the 2.5% stop), avg_trade_return -$621, commission_pct_of_gross 68% (fees dominate despite the fee-discipline design). The only thing that worked is the sizing probe (avg_position_pct 95.5%, clean), which does not rescue the run. Not optimize: optimization cannot manufacture multi-year 15m data (so the promised walk-forward/holdout is impossible regardless of parameters), cannot validate 8 metrics_unreliable trades, and cannot convert a PF-0.17, fees-eat-68%-of-gross signal into an edge. The mechanism-to-asset reasoning (ETH chops -> mean-reversion) was sound, but it has no clean+deep-data home: HL has only ~2.4 months of 15m history and the deep-history venues (BINANCE USD-M/COIN-M) are blocked by the engine over-exposure defect. FAILURE PATTERN: HL intraday 15m strategies (both BTC momentum-LS and ETH mean-reversion-LF) were proposed specifically to escape the HL daily ~3yr sparsity wall via thousands of bars, but the HL 15m catalog itself is only ~2-2.5 months, so they generate too few trades (8-108) with metrics_reliable false/marginal and the promised walk-forward is impossible — the data wall defeats HL strategies at BOTH daily and intraday cadence. The validated edge needs a deep-history venue the engine can size correctly, which currently does not exist.

Long/flat intraday 15-minute 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 dip when the stretch is large (>= 1.75 ATR) AND the price is at least 0.6% below the mean (so the reversion target clears HL's ~0.09% round-trip fee) AND the 96-bar regime EMA is not falling (avoids catching knives in downtrends) AND a 4-bar post-exit cooldown has elapsed (throttles churn). Exits on reversion (close back at/above VWAP), a 0.8% take-profit, a 2.5% stop-loss, or a 16-bar time stop. Long/flat only, never shorts, one fixed-fractional position (95% of equity notional) per entry with no dynamic resizing, leverage 1.0.

Sizing probe passed: avg_position_pct 95.5% (<=100%), no physically impossible intraday equity jumps — confirms HL ETH long/flat sizing is clean, consistent with the sibling evidence.

Thoughtful mechanism-to-asset matching: chose intraday mean-reversion (not another trend/breakout port) because ETH showed no daily trend edge; sound, bug-avoiding construction (long/flat, fee-gated, regime filter, cooldown, leverage 1.0).

PREMISE FALSIFIED (same as BTC intraday sibling): HL ETH 15m data is only ~2.4 months (span 2026-03-16 to 2026-05-27, 6862 bars), NOT ~3yr/75k bars — the promised valid 3-window walk-forward + 15-day holdout is impossible.

Degenerate sample: only 8 trades (entry_signaled 8), data_days 6, metrics_reliable=FALSE — the 'hundreds of trades' that justified the intraday-HL approach did not materialize (data wall + over-restrictive gating).

Strongly negative edge: profit_factor 0.17, Sharpe -9.66 (CI -24.6 to +4.4), avg_win 251 vs avg_loss 1494 (losers ~6x wins), avg_trade_return -$621.

Fee discipline failed: commission_pct_of_gross 68% — fees dominate gross despite the fee-gated design.

EthHyperliquid15mVwapMeanReversionLongFlat was a deliberate change of mechanism — intraday mean-reversion matched to ETH's choppy, non-trending character — and was meant to defeat the Hyperliquid data-sparsity wall by trading at 15-minute cadence over a presumed ~3 years. Instead it hit the same wall its BTC sibling did: only ~2.4 months of HL ETH 15m data existed, yielding a degenerate 8 trades with metrics_reliable=false and a strongly negative signal (PF 0.17, Sharpe -9.66, losers ~6× winners, fees 68% of gross). The reviewer abandoned it pre-optimization, noting that no parameter sweep can manufacture multi-year data, validate 8 unreliable trades, or turn a fees-dominated losing signal into an edge. The sizing probe passed (95.5%, clean) and the asset-matching logic was sound, but the conclusion was structural: the HL data wall defeats both daily and intraday cadence, and the validated edge has no clean, deep-data home at present.

HL intraday 15m strategies cannot escape the venue's short-history wall because the 15m catalog itself is only ~2–2.5 months — too little to generate the 'hundreds of trades' the approach assumed — so a sound mechanism-to-asset match still fails without a deep-history, correctly-sizing venue.

The analyst abandoned it at the backtest-review gate, before optimization, for three unfixable reasons: the core premise was falsified — only ~2.4 months / 6,862 bars of HL ETH 15m data existed (not ~3yr/75k), making the promised walk-forward and holdout impossible — the sample was degenerate (8 trades, metrics_reliable=false), and the edge was strongly negative with fees dominating gross.

A long/flat, single-instrument, OHLCV-only intraday 15-minute mean-reversion-to-VWAP strategy on ETHUSD.HYPERLIQUID that buys volatility-stretched dips (>=1.75 ATR / >=0.6%) with a regime filter and exits on reversion — the first non-trend mechanism in the pending set, chosen because ETH chops rather than trends, and intended to clear the HL data-sparsity wall via intraday cadence.

It produced a degenerate, strongly negative result: only 8 trades over data_days 6 with metrics_reliable=false, total return -3.2%, Sharpe -9.66, profit factor 0.17, avg_win $251 vs avg_loss $1,494 (losers ~6x wins), avg_trade_return -$621, and commission_pct_of_gross of 68%; the sizing probe passed cleanly (avg_position_pct 95.5%).