HYPERLIQUID BTC Perp 1H Daily-Anchored VWAP Excursion Long/Short Mean-Reversion (Single-Venue, Single-Instrument, 3-Parameter, Volatility-Scaled Stop)

A LONG/SHORT, single-instrument, single-venue intraday mean-reversion strategy on BTC-USD.HYPERLIQUID perpetual using 1-HOUR bars and OHLCV-only data. This deliberately fills the two worst-violated quota buckets that are ALSO testable in the current engine: HYPERLIQUID venue (5.0% actual vs 20% target — the largest under-represented TESTABLE venue gap, since options/COIN-M/cross-venue-netting are all tooling-blocked) and long_short direction (12.7% vs 55% ceiling on long-only). It transplants the ONE mechanism class proven to clear every optimization gate this session — daily-anchored VWAP excursion mean-reversion (the ETH 1H BINANCE sibling was promoted) — onto a venue and asset where it has not yet been tried in a long/short form. Critically it AVOIDS the documented failure traps: (a) it uses 1H bars on BTC, which has ~2.5+ years of continuous HYPERLIQUID history (>20,000 bars), sidestepping the SOL-15m-38-day data-insufficiency abandon; (b) it is fully bookable as single-instrument directional PnL (no funding-cash-flow accounting, no cross-venue MTM netting, no multi-leg options — the four architectural tooling gaps that killed every carry/basis/VRP hypothesis this session); (c) it sizes the stop RELATIVE to the entry z-score's expected reversion rather than a fixed tight % (the SOL VWAP abandon noted a fixed -1.5% stop was too tight vs the recovery target, producing an inverted win/loss profile). Mechanism: maintain a daily-anchored VWAP (reset 00:00 UTC) and a rolling standard deviation of (close - VWAP). Compute z = (close - VWAP) / rolling_std. When z <= -z_entry (price stretched far below fair value), go LONG expecting reversion to VWAP; when z >= +z_entry, go SHORT. Only fire on EXTREME excursions (z_entry ~2.0-2.5) so the expected reversion (~0.3-0.6% of price back toward VWAP) comfortably exceeds Hyperliquid's ~0.09% round-trip taker cost. Three tunable parameters only: z_entry, vwap_std_lookback, stop_atr_mult.

Fixes the reported Layer-1.5 mismatch by trading the instrument the hypothesis declares, BTC-USD.HYPERLIQUID (the dash makes _root() parse it as 'BTC-' vs 'BTC', which is why BTCUSD.HYPERLIQUID read as a different asset). Two consequences forced by the catalog, both verified against the real files. First, BTC-USD.HYPERLIQUID exists ONLY at 4-HOUR: one parquet, exactly 5,000 rows (the HL 5000-candle cap), 2024-02-09 to 2026-05-22, and no 1-HOUR directory at all. The hypothesis's '1-HOUR bars, >20,000 bars, 2.5+ years' describes BTCUSD.HYPERLIQUID (1H, 2025-10-11 to today), not the symbol it declares. Requesting 1-HOUR on the declared symbol would abandon on data_unavailable, because _collect_from_hyperliquid strips the quote suffix and asks Hyperliquid for coin 'BTC-', which returns nothing; ensure_bars_available returns True for the existing 4-HOUR series (5,000 >= min_bars) and nothing calls ensure_bars_through, so the stale tail does not block. Second, the config change alone would have produced ZERO trades: a UTC day holds only 6 four-hour bars, and the old code reset the deviation buffer at midnight, so with n intra-session deviations the maximum attainable |z| is (n-1)/sqrt(n) = 2.04 at n=6, making z_entry 2.0-2.5 unreachable in practice. I made the deviation buffer roll across sessions while only the VWAP ANCHOR resets at midnight, which is literally what the hypothesis specifies ('a daily-anchored VWAP and a rolling standard deviation of (close - VWAP)'), and trimmed it to the lookback so per-bar cost stays O(lookback) instead of growing unbounded. Replaying the actual BTC-USD 4H file confirms the fix: 164 entry-eligible signal bars at z_entry=2.0 (326 at 1.5, 79 at 2.5) out of 2,491 eligible bars, about 70 trades/year. The three locked session constants (min_day_bars, entry_cutoff_hour, day_end_hour) were tuned for a 1-HOUR grid where a 22:00 cutoff and 23:00 flat make sense; on the 4-HOUR grid those hours never occur, so the end-of-day flat would never have fired. They are now parameters with declared _param_bounds, set to 2 / 16:00 / 20:00 for this grid. Everything else is untouched: imports unchanged, the z signal is still continuous and returned every bar, the ATR-scaled stop and entry-context reconstruction are unmodified, sizing still floors to precision with a min-notional check, and leverage stays 1.0. NOTE for the Research Lead: if the intent really was the live 1-HOUR series, the hypothesis should be revised to name BTCUSD.HYPERLIQUID — BTC-USD is a stale alias whose data stopped updating on 2026-05-22.