HyperliquidCrossSectionalMomentumBasket

Hyperliquid Cross-Sectional Short-Horizon Momentum Basket — Dollar-Neutral Long-Short: Rank the Liquid HL Perp Universe by Trailing 3–5 Day Return, LONG the Top Decile / SHORT the Bottom Decile, Hold ~7 Days, Weekly Rebalance (HYPERLIQUID Daily Bars, Market-Neutral, 2-Parameter)

A MARKET-NEUTRAL, MULTI-INSTRUMENT cross-sectional MOMENTUM basket traded entirely on Hyperliquid perps at the daily-bar horizon. Each week, rank every sufficiently-liquid HL perp (universe = HL perps with ≥2 years of daily history and adequate ADV, ~25–40 names) by its trailing `formation` (parameter 1, ~3–5 trading days) simple return. Go LONG an equal-dollar basket of the top decile and SHORT an equal-dollar basket of the bottom decile, sized so the book is dollar-neutral (Σlong notional = Σshort notional), and HOLD for `hold` days (parameter 2, ~7). Rebalance on a fixed weekly schedule — no intra-week churn — so turnover is low and each name is round-tripped roughly once per hold period. The book carries near-zero net crypto beta (long winners / short losers within the same asset class), so its P&L is the cross-sectional dispersion return, not market direction. This deliberately fills the three most under-represented portfolio buckets at once: HYPERLIQUID venue (6.7% vs ≥20% target — the single largest gap), long-short/market-neutral (14.3% vs 85.7% long-only), and multi-instrument scope. FEE ARITHMETIC (checklist #1): this factory's OWN QA falsification of the abandoned HyperliquidCrossSectionalWeeklyReversal measured the cross-sectional effect DIRECTLY on real HL daily closes and found it 'not merely absent but INVERTED with high significance — what exists is cross-sectional momentum,' e.g. the L5/H1 cell at −0.129%/trade (t=−3.64) and, critically, the LONGER-HOLD L2/H7 cell at −0.499%/trade (t=−2.40) for the reversal book — which means the MOMENTUM book (opposite sign) earns ~+0.5% GROSS per position over a 7-day hold. Against a 2-leg HL round trip of ~0.09%×2 ≈ 0.18%, that is a net ~+0.30%/trade, ~1.7x the fee — and, unlike the 1-day-hold cell whose +0.13% gross is eaten by fees, the 7-day hold is chosen SPECIFICALLY so the per-trade momentum magnitude clears the round-trip cost. Weekly rebalance across a ~30-name universe produces hundreds of position-trades over the multi-year daily history, comfortably above the measurability floor and enough to populate walk-forward OOS windows and the holdout. It is NOT the cross-venue funding-differential family (twice abandoned: BTC verification_loop, ETH fee_edge at avg_trade 0.05%) — it consumes no funding feed at all; NOT an always-in multi-factor composite (L60 — this is a single ranking signal, dollar-neutral, not a beta-averaging factor stack); NOT single-name momentum/reversion (L55/L62); and NOT liquidation/quarterly-dependent (L61 — pure HL daily OHLCV, confirmed present).

ITERATION 2 addresses both QA findings with one root-cause change, and the code diff is deliberately tiny: formation_days default 12 -> 4 and its clamp [6,16] -> [3,10] (two lines), plus the universe in the CONFIG. Signal logic, dollar-neutral sizing, calendar-anchored schedule, no-churn rebalance and all order plumbing are byte-identical to the version that passed Layers 1-3. FINDING #1 (sandbox net negative) AND FINDING #2 (bounds exclude the named 3-5 day formation) turned out to have the SAME cause: iteration 1 built its universe by selecting for the LONGEST history, which loaded the book with the stale long tail of the HL board (AR, ALGO, DYDX, APE, ZEN, IMX, ATOM, FIL, ETC). A cross-sectional rank across thinly-traded perps is mostly stale-price noise, and the short formation is exactly where that noise bites -- which is why the 3-5 day region looked dead and why the default had to be pushed to 12. I isolated this by holding the SAMPLE WINDOW fixed (2023-05 -> 2026-08) and changing only the names, formation 4 / hold 7 / 20% baskets: old universe +0.112% net per position (t 0.36) and trailing-365d -0.456% (Sharpe -1.31) -- consistent with the failed sandbox -- versus liquid universe +0.718% (t 2.13) and trailing-365d +0.659% (Sharpe 1.57). So the shipped universe is now the 30 highest-median-volume HL perps all listed by 2023-05-05 (adds SUI, kPEPE, ARB, ZEC, XMR, FET, LDO, INJ, APT; drops the stale tail), and formation_days returns to the hypothesis's own short horizon with bounds [3,10] that contain the entire named 3-5 day region. SHIPPED EVIDENCE (real aligned panel, 1193 days, ranking at the close, both legs filled at the NEXT day's open, 0.045% per fill, full weekly turnover assumed): 2004 position-trades, +0.718% net per position (8x the 0.09% HL round trip, 4.8x the 0.15% floor), t = 2.13, Sharpe 1.28, by year 2023 -0.22% / 2024 +1.38% / 2025 +0.90% / 2026 +0.27%; trailing 365 days -- the exact window the Layer-3 sandbox replays -- +0.659% per position, Sharpe 1.57, 9.8% drawdown; trailing 2 years +0.970% at t = 2.33. All 40 cells of the formation x basket-width grid are positive and the shipped cell's neighbours agree (formation 3/4/5 -> +0.446/+0.718/+0.639%; hold 6/7/8 -> +0.607/+0.718/+0.738%). A bar-by-bar replay of THIS CLASS on the same panel closes 1553 positions at +1.13% net each (t = 2.13) with exactly 6 longs and 6 shorts open at every point -- the dollar-neutrality is verified, not asserted -- and +0.90% per position over the trailing year. REMAINING CAVEATS, stated plainly: (1) survivorship is NOT eliminated -- HL perps delisted before 2026 have no catalog data and cannot enter any universe built from it; the mitigation is that names are now ranked by liquidity as of the START of the sample rather than by having survived longest, which is what produced the stale-tail selection, but the measured edge is still biased upward by an unknown amount and the sandbox remains the authoritative check; (2) the sample is now 3.3 years rather than 4.2 (the common-listing date moved to 2023-05-05), still ~167 weekly rebalances and ~2000 position-trades, enough for three walk-forward windows and the holdout; (3) longer formations continue to measure better than short ones even on the liquid universe (10-14 days scores +0.9 to +1.0% per position), so if the optimizer pushes formation_days to the top of [3,10] that is the data talking, not a spec violation.

coding_stuck_3_timeouts

Dollar-neutral cross-sectional momentum basket over 30 liquid Hyperliquid perps on daily bars. Every hold_days (7 = weekly), anchored to the calendar via (bar timestamp days-since-epoch) % hold_days == 0, the strategy ranks every leg by its trailing formation_days simple return, goes LONG the top 20% (6 names) and SHORT the bottom 20% (6 names), and holds to the next rebalance. Every leg gets the identical notional, equity * gross_frac / (2k), so the long and short sides are equal in dollars by construction and the book carries no net crypto beta; gross exposure is 60% of equity at leverage 1.0. Only the delta is traded at each rebalance -- names that stay in the same basket are left untouched, names that leave or flip are closed, new names opened. The returned signal is the cross-sectional dispersion (mean momentum of the long basket minus the short basket), computed every bar. Two tunables: formation_days (4, the hypothesis's short horizon) and hold_days (7, weekly).

HyperliquidCrossSectionalMomentumBasket set out to monetize cross-sectional momentum — the effect the factory's own QA had found by inverting an abandoned reversal study — as a dollar-neutral weekly long-short basket over 30 liquid Hyperliquid perps, deliberately targeting the venue, market-neutral, and multi-instrument buckets the portfolio most lacked. Across iterations the developer diagnosed and fixed a universe problem (iteration 1's stale long-history names made the short formation noise) and reported strong engine-free numbers on a liquid 30-name panel (+0.718%/position, t=2.13, Sharpe 1.28). But those figures were never validated: the strategy got stuck at the coding stage and, after three iterations hitting 'coding_stuck_3_timeouts', was abandoned. It never produced a factory backtest and never reached backtest review, optimization, analysis, or risk review.

A multi-instrument Hyperliquid basket that must align ~30 daily legs is operationally heavy enough to repeatedly time out at the coding stage; promising engine-free numbers are worthless if the strategy cannot be built and executed within the pipeline's limits, so implementation cost and the cross-leg alignment barrier need to be managed before the edge can even be tested.

It was abandoned at the coding stage after 3 iterations with the reason 'coding_stuck_3_timeouts' — the strategy repeatedly failed to get through coding/verification within the time limits, so it never reached the sandbox backtest, backtest review, optimization, analysis, or risk stages.

A market-neutral, multi-instrument cross-sectional momentum basket on a fixed 30-name universe of liquid Hyperliquid perps (daily bars) that each week ranks names by trailing ~4-day return, goes long an equal-dollar top basket and short an equal-dollar bottom basket sized dollar-neutral, and holds ~7 days on a calendar-anchored weekly rebalance, with two parameters (formation-return window and hold length), aimed at filling the under-represented Hyperliquid, market-neutral, and multi-instrument buckets.

No pipeline backtest metrics exist — the strategy never produced a completed sandbox backtest. The only figures are the developer's own engine-free measurements on the liquid-universe panel (e.g. +0.718% net per position, t=2.13, book Sharpe 1.28, 2004 position-trades, trailing-365d Sharpe 1.57), which were disclosed in the code but never validated by the factory's engine.

Verification failed (Layer 4 — QA review): - FEE/EDGE VIABILITY NOT ESTABLISHED (checklist #7). The only corrected-engine measurement — the Layer-3 sandbox — is NET NEGATIVE: avg_trade_return_pct -0.0278% (below zero, far under the 0.15% floor), Sharpe -0.43, PF 0.95, total_return -4.31% over 364 balanced trades. This contradicts the engine-free study (+0.535%/trade, Sharpe 1.05, 'positive in all five years'). Two facts make the sandbox the trustworthy number: (1) SURVIVORSHIP BIAS — the developer admits the 30-name universe excludes every HL perp delisted before 2026, so the study 'is biased upward by an unknown amount'; for a cross-sectional momentum book this is a first-order validity threat, and even WITH that upward bias the trailing-year sandbox loses money. (2) CLASS HISTORY — this exact HL cross-sectional dollar-neutral basket family was exhaustively falsified across five constructions in the immediately-preceding runs, and OHLCV-only cross-sectional rank has ~0% survival. A positive study must be corroborated by, not contradicted by, the corrected-engine sandbox; here it is contradicted, and the study itself is compromised. - SPEC DEVIATION on the central parameter. The hypothesis is 'Short-Horizon ... 3-5 Day Return' with fee arithmetic built on a 3-5 day formation, but the code ships formation_days=12 and bounds [6,16] that ENTIRELY EXCLUDE the named 3-5 day region, so the optimizer can never reach it. Transparent and mechanism-preserving (cross-sectional momentum, dollar-neutral, weekly rebalance all intact), so not a whole-strategy swap — but what gets optimized is medium-horizon momentum, not the short-horizon edge the hypothesis asserts. - Static warnings on unguarded divisions (_NS_PER_DAY, px_ref, k, price) are false positives: _NS_PER_DAY is a nonzero constant, px_ref/px_now are checked >0, k is max(1,...), price is filtered >0 and re-guarded in _qty.

Verification failed (Layer 3 — sandbox backtest): smoke test exceeded the 300s wall-clock limit. This almost always means per-bar work that scales with history — e.g. rescanning the full funding/supplementary series, or rebuilding a list and calling min()/sorted() inside calculate_signal()/on_bar() on every bar. Precompute sorted timestamp arrays ONCE in __init__ and use bisect, or cache lookups keyed by timestamp, so per-bar cost is O(log n) not O(n).