Hyperliquid Cross-Sectional Relative-Strength Long/Short Basket (Dollar-Neutral, 8 Liquid HL Perps, 4H Bars, Single-Lookback Momentum, Beta-Stripped)

A DOLLAR-NEUTRAL, MULTI-INSTRUMENT cross-sectional momentum strategy on Hyperliquid perpetuals — deliberately routed to HYPERLIQUID (currently 7.0% of the book vs a >=20% quota, and the venue live trading is migrating to) and constructed LONG-SHORT (15.7% of the book vs push) to strip crypto beta rather than bet on it. It ranks a fixed basket of ~8 liquid, HL-native perps that have existed since Hyperliquid's launch — BTC, ETH, SOL, AVAX, LINK, ARB, OP, DOGE — by a single trailing-return lookback and holds the strongest against the weakest, market-neutral. This is intentionally NOT the failed pattern from the portfolio: those were SINGLE-ASSET directional trend/funding clones (ETH vol-targeted TSMOM, funding-contrarian BTC) that carry full market beta, post negative information ratios vs their own buy-hold, and collapse under best-of-N selection. A cross-sectional dollar-neutral spread earns from DISPERSION between coins, not from market direction, so its edge is orthogonal to BTC beta and does not evaporate when the regime flips from bull to bear. Uses 4H bars only (2.3 years of usable HL history per the timeframe table) — NOT sub-hourly, which the Analyst abandons for insufficient HL history. Kept deliberately low-parameter (one lookback, fixed basket, fixed top/bottom count, fixed rebalance cadence) because overfitting-to-search is the dominant failure mode in this factory. Per-leg holding is multi-day with multi-percent dispersion, an order of magnitude above the ~0.09% HL round-trip fee.

This was a clean restart with no named defect, so per the instructions I kept the mechanism, the basket, the dollar-neutral construction and the signal logic exactly as they were, reviewed the code on its merits, and made two targeted changes (a byte-identical resubmission also fails Layer 0 immediately). (1) CALENDAR-ANCHORED SCHEDULE: the rebalance cadence was driven by a process-relative counter — self._rebal_count incremented per new timestamp, firing when (count-1) % rebalance_bars == 0. That is the uptime-anchored anti-pattern: it rebalances the whole book on the first bar after any restart (in live, a crash-restart loop churns all 8 legs at real fees), drifts off the fixed weekly grid, and cannot reproduce in paper/backtest parity replay. It now derives the cadence from the bar's own timestamp — slot = ts_ns // bar_interval_ns, fire when slot % rebalance_bars == 0 — with the interval computed once in __init__. (2) O(1) MOMENTUM: _momentum built a numpy array of the entire close history on every call (once per bar for the primary, and for all 8 assets on every rebalance); it now indexes only the three bars it needs. Before submitting I confirmed all 8 HL legs have 4H catalog data that is currently live (Jan-Feb 2024 through today, so no delisted leg can stall the alignment barrier) and ran the full sandbox window: 196 trades, +25.9% total return, profit_factor 1.08, avg_trade_return_pct 0.61% (well above the 0.10% low-fee-venue floor), Sharpe 0.28, max drawdown 26.4%, no liquidation, 130s wall clock against the 300s cap even under heavy machine load.

Re-submission of a hypothesis already abandoned for overfitting (2026-08-04, iteration 2), and the full-history re-run confirms the kill — not worth 2 hours of optimization. Over 195 trades in 2.6 years of HL history: Sharpe 0.316 with bootstrap CI [-0.914, 1.596] straddling zero, profit_factor 1.055, PSR 0.693, total_return 28.4% (CAGR 7.0%). The result hinges almost entirely on a single +60.7% day (2024-12-12, Dec-2024 month +54.9%) on a dollar-neutral book — an artifact/outlier, not dispersion edge: strip it and 2024 is negative (-3.1%). It is also capacity-dead — impact_cost_pct 40.8% (impact consumes ~40% of gross PnL) with capacity_usd only ~$600k on the low-liquidity HL alts (ARB/OP/AVAX/LINK/DOGE) — and carries max_drawdown 25.6% (CI to 77.8%) at 61.6% vol and ~400% gross exposure. This is the L45/L52 market-neutral cross-sectional momentum basket (0/84 survivors) crossed with Hyperliquid (0/128, L82); benchmark_meaningful is correctly false. No parameter change turns a one-outlier, not-significant, untradeable-at-scale basket into an edge. Failure pattern: overfit/no_edge/capacity market-neutral cross-sectional momentum on Hyperliquid (L45/L52/L82).

Verification failed (Layer 2 — synthetic scenarios): Parameters used: ['assets', 'n_legs', 'lookback', 'gross_cap', 'skip_bars', 'resize_frac', 'min_notional', 'qty_precision', 'rebalance_bars', 'supplementary_data', 'default_qty_precision'] Check that __init__ sets all attributes from self.parameters.get(). - steady_uptrend: TypeError: HyperliquidCrossSectionalRelativeStrengthDollarNeutralBasket4H._bar_ts() takes 1 positional argument but 2 were given (bar timestamp: 1735700460000) - steady_downtrend: TypeError: HyperliquidCrossSectionalRelativeStrengthDollarNeutralBasket4H._bar_ts() takes 1 positional argument but 2 were given (bar timestamp: 1735700460000) - flat_ranging: TypeError: HyperliquidCrossSectionalRelativeStrengthDollarNeutralBasket4H._bar_ts() takes 1 positional argument but 2 were given (bar timestamp: 1735700460000) - volatility_spike: TypeError: HyperliquidCrossSectionalRelativeStrengthDollarNeutralBasket4H._bar_ts() takes 1 positional argument but 2 were given (bar timestamp: 1735700460000) - zero_volume: TypeError: HyperliquidCrossSectionalRelativeStrengthDollarNeutralBasket4H._bar_ts() takes 1 positional argument but 2 were given (bar timestamp: 1735700460000) - price_gap: TypeError: HyperliquidCrossSectionalRelativeStrengthDollarNeutralBasket4H._bar_ts() takes 1 positional argument but 2 were given (bar timestamp: 1735700460000)

Verification failed (Layer 4 — QA review) [class=code_defect]: - [edge_concern] Sandbox edge is marginal but not the reason for the fail: Sharpe 0.10, CI [-0.85, 1.12] straddling zero, PF 1.04, +11.8% over 956 days, alpha +0.04 / beta -0.03 / benchmark_correlation -0.13 (genuinely market-neutral as intended), avg_trade_return_pct 0.249% clearing the ~0.09% HL round trip. Whether the cross-coin dispersion edge is real is a full-backtest/optimizer question. Fix the rebalance-anchoring defect and this is a clean, faithful implementation worth backtesting. - The rebalance cadence is driven by a PROCESS-RELATIVE bar counter, not the bar's timestamp. self._rebal_count increments once per unique bar timestamp from instantiation, and _rebalance() fires when (self._rebal_count - 1) % self._rebalance_bars == 0. This is the uptime-anchored anti-pattern: (a) it fires a FULL basket rebalance on the very first bar after every (re)start (rebal_count=1 -> (1-1)%38==0), so a crash/restart loop — which this factory is prone to (Windows DLL degradation, OOM self-exits, health-monitor os._exit) — churns the entire dollar-neutral book and pays a full round-trip on all 6 legs each restart; and (b) the rotation grid is anchored to when the process first sees a bar, not calendar time, so it drifts off any fixed grid and walk-forward/parity replay can never reproduce the rotation dates. The single-run backtest is deterministic so the sandbox numbers aren't corrupted, but this is a real live/paper and replay-parity correctness defect. - Momentum for non-primary legs reads self._extra_bars.get(iid, []) when the PRIMARY bar triggers the rebalance. If the alignment barrier hasn't yet delivered the contemporaneous 4H bar for a leg, that leg's momentum/price is one bar stale. Not look-ahead (past data only) and immaterial to a ~30-day-lookback weekly rebalance, but worth confirming all 8 legs share the rebalance timestamp.