HyperliquidCrossSectionalLowBetaDefensiveNeutralLS

Hyperliquid Cross-Sectional Low-Beta Defensive Anomaly, Beta-Neutral Basket (Long Low-Realized-Beta Perps, Short High-Realized-Beta Perps, Inverse-Vol Weighted, Weekly Rebalance)

A LONG-SHORT, BETA-NEUTRAL, MULTI-INSTRUMENT relative-value strategy on Hyperliquid perpetuals that harvests the low-volatility / low-beta anomaly — the persistent tendency for LOW-beta assets to deliver better risk-adjusted returns than HIGH-beta assets, which in crypto is amplified because high-beta alts structurally bleed via leverage-decay, liquidation churn, and negative convexity while low-beta majors compound more cleanly. This is a THIRD, ORTHOGONAL cross-sectional factor distinct from the two pending HL baskets (funding-carry uses funding; short-horizon-reversal uses recent return) — here the selection signal is REALIZED BETA/VOLATILITY, not price momentum, not funding, not recent return, so it satisfies L52's demand for a non-momentum-rank selection edge and does not spray any dead template. Every week, estimate each universe perp's 30-day realized beta to BTC (and realized vol); go LONG the 3-4 LOWEST-beta names and SHORT the 3-4 HIGHEST-beta names. Legs are inverse-vol weighted, then the SHORT sleeve is scaled so the book is BETA-neutral to BTC (not just dollar-neutral), removing directional crypto exposure so the residual return is the defensive factor premium, not market beta or a lucky outlier day (the specific failure mode that killed prior 'composite' baskets, kurtosis 18.5). It also fills the most under-represented HORIZON bucket (macro_1w_plus at 0.5%) plus HYPERLIQUID venue (6.7% vs >=20%), long_short direction (13.6% vs long_only 86.4%), and multi-instrument scope. Daily bars over HL's full history; betas are slow-moving so turnover — and therefore fee drag — is very low. Minimal fixed parameters (universe, beta lookback=30d, names/sleeve, rebalance=weekly) to resist overfitting.

Selection variable is REALIZED BETA/VOL, not momentum, funding, or recent return, so the factor is orthogonal to the two pending Hyperliquid baskets, as the hypothesis requires. Beta is estimated over a deque capped at beta_lookback_days, so per-bar work is O(12 x 30) and never grows with history (no Layer-3 timeout risk). All 12 legs share ONE aligned return-tuple buffer that is appended only when every leg reports a valid price on the same bar, so the covariance can never mix a stale leg with a fresh one; the base class's cross-leg barrier additionally defers the primary signal until all same-timeframe legs have a bar at the primary timestamp. The rebalance is calendar-anchored ((ts_event // 1 day) % rebalance_days == 0), a pure function of the bar's date, so it is identical across restarts, backtests and paper replay. Beta-neutralization is the hypothesis's core construction and it necessarily makes the long (low-beta) notional larger than the short notional -- a plain dollar-neutral version of this trade would carry residual SHORT market beta and its PnL would be market direction, not the defensive premium. Per prior neutral-book blowups, leverage stays 1.0 and gross is capped at 60% of equity, with a 30% resize band so the book de-levers as equity falls instead of compounding into effective leverage. Warmup is ~51 daily bars (21 min_bars + 30 beta window), a small fraction of the 365-day sandbox window, and the book re-forms every 7 days thereafter, so the strategy trades on the order of 45 rebalances in that window. Venue is Hyperliquid futures because the basket is long+short (spot CASH cannot short); leverage is 1.0 by choice, not by venue. One reservation, stated as required and implemented as asked: the market proxy is BTC itself, so BTC's beta is identically 1.0 and its rank is deterministic; an equal-weight universe index would give every name including BTC a non-degenerate beta and is the natural refinement if the analyst wants it.

Well-constructed but sub-DSR beta-neutral basket in the confirmed-dead betting-against-beta / long-short major-perp class (L45, 0/84). Despite genuine beta-neutrality (realized beta 0.043), no liquidation, an above-floor avg_trade_return_pct 0.565%, and moderate 23% drawdown, the base Sharpe is only 0.285 with sharpe_ci_low -0.366 — the CI straddles zero, so the result is indistinguishable from no-skill even before the optimizer's best-of-225-trial selection deflates it; there is no path from 0.285 to clearing the deflated-Sharpe 0.95 hard gate (AAVE failed DSR at a much higher post-opt ~1.0, and every BTC probe this session with a CI-low near zero DSR-failed). profit_factor 1.076 is fee-fragile and impact_cost_pct is 11% of gross with only $8.2M capacity. The 57.7% headline is front-loaded in 2021 (+65.9%, early-crypto low-beta premium) and has decayed to negative in the recent regime (2023 -5.6%, 2026 -11.7%, rolling Sharpe -3 to -3.9) — precisely the window the walk-forward OOS and 15-day holdout occupy, so optimization would fit the strong early years and fail out-of-sample. IR -0.596. Abandon at BACKTEST_REVIEW rather than spend the 2-hour optimization budget to reach the AAVE/DSR outcome. Failure pattern: betting_against_beta_neutral_basket_sub_dsr.

Betting-against-beta basket on 12 liquid Hyperliquid perpetuals (BTC, ETH, SOL, BNB, XRP, DOGE, AVAX, LINK, LTC, ADA, ATOM, AAVE) on daily bars. Every 7 calendar days it estimates each name's 30-day realized beta to BTC (cov(r_i, r_btc)/var(r_btc)) and realized daily vol from a single timestamp-aligned return buffer, then goes LONG the 3 lowest-beta names and SHORT the 3 highest-beta names. Legs are inverse-volatility weighted inside each sleeve, and the two sleeves are sized so N_long*beta_long == N_short*beta_short, making the book beta-neutral to BTC rather than merely dollar-neutral. Gross exposure is 60% of equity at leverage 1.0; legs are re-sized when held notional drifts more than 30% from target. calculate_signal returns the primary leg's continuous standing in the beta cross-section, (universe mean beta - primary beta) / beta stdev.

Use an equal-weight universe index (or ex-BTC composite) as the market proxy so every name including BTC gets a non-degenerate estimated beta — the developer's own suggested refinement.

The market proxy is BTC itself (self._market_idx = 0 = primary), so BTC's realized beta is identically cov(r_btc,r_btc)/var(r_btc) = 1.0 on every bar. BTC's cross-sectional rank is therefore deterministic, not estimated: in a crypto universe where most alts run beta > 1 to BTC, BTC will almost always fall into the LOW-beta (LONG) sleeve, and its continuous score is near-deterministic. Disclosed in the rationale and not a correctness bug (BTC genuinely IS the low-beta major, so the placement is economically sensible), but one of the 12 names carries no cross-sectional information and the primary leg is effectively hard-wired long.

On restart, reconstruct _side_state from cache.positions_open() before deciding to exit.

should_exit() returns True whenever _side_state[primary] == 0 (state lost on restart), so on the first bar after any live/paper restart correctly-held positions are exited and re-entered — an avoidable fee round-trip per restart. Harmless in backtest.

At backtest review, check whether the fat-tail losses cluster on beta-regime breaks (correlation-spike crashes); consider beta-shrinkage or a vol-cap on the high-beta short sleeve before spending optimization budget.

Informational for the BACKTEST_REVIEW analyst (not a QA block): the 362-day sandbox is net-negative (total_return -5.1%, Sharpe -0.52, PF 0.88) despite a healthy per-trade edge. avg_trade_return_pct is +1.169% (well above the ~0.15% fee floor) and turnover is low (11.5, 94 trades), so this is NOT a fee-drag/no-edge failure — the loss is concentrated in a few fat-tail days (largest_loss -8720 vs largest_win 3201, kurtosis 7.49, negative skew). This is a performance judgment for the analyst, not a correctness defect.

Genuinely beta-neutral (realized beta 0.043) — the beta-scaling construction works, unlike a plain dollar-neutral version; no liquidation, clean risk build

avg_trade_return_pct 0.565% clears the fee floor, moderate max_drawdown 23.2%, 744 trades (well-populated), positive alpha 0.042

Fills under-represented buckets (Hyperliquid, long_short, multi-instrument, macro horizon) with a non-momentum selection signal

Sharpe 0.285 with sharpe_ci_low -0.366 — CI straddles zero, so the edge is not distinguishable from no-skill; after best-of-225-trial selection it cannot clear the deflated-Sharpe 0.95 gate

profit_factor 1.076 is in the fee-fragile band; impact_cost_pct 11% of gross with capacity only $8.2M

Edge front-loaded and decayed: 2021 +65.9% carries the entire 57.7% headline, then 2022 -13.3%, 2023 -5.6%, 2026 -11.7% with rolling Sharpe -3 to -3.9 — exactly the window WF-OOS and the holdout evaluate

information_ratio -0.596; betting-against-beta / beta-neutral major-perp basket is the 0/84 class (L45) that dies overfit/no_edge

HyperliquidCrossSectionalLowBetaDefensiveNeutralLS was a well-built third orthogonal cross-sectional factor — long the lowest-realized-beta Hyperliquid perps, short the highest, scaled to true beta-neutrality (not just dollar-neutral) to isolate the defensive premium and avoid the fat-tail and concentration failures that had killed prior baskets. The construction worked: realized beta 0.043, no liquidation, +57.7% over 744 trades, moderate 23% drawdown, and controlled kurtosis. But the edge was not there — base Sharpe 0.285 with a CI straddling zero (-0.366), fee-fragile PF 1.076 on tiny $8.2M capacity, and a headline entirely front-loaded in 2021 that decayed to negative across 2022-2026. The analyst abandoned it at backtest review as a sub-DSR basket in the confirmed-dead betting-against-beta major-perp class (L45, 0/84) whose decayed recent regime would fail walk-forward and holdout; it never reached optimization, analysis, or risk review.

Fixing the risk construction is necessary but not sufficient — this book was genuinely beta-neutral with no liquidation (unlike its funding-carry sibling), yet the betting-against-beta premium on crypto majors is a confirmed-dead 0/84 class whose entire edge was an early-2021 low-beta premium that decayed to negative; a Sharpe of 0.285 with a CI on zero and a decayed recent regime cannot clear DSR no matter how sound the neutrality.

The analyst abandoned it at backtest review: despite clean, genuine beta-neutrality and no blowup, a base Sharpe of 0.285 with a CI straddling zero is indistinguishable from no-skill and has no path to clearing the deflated-Sharpe 0.95 gate after best-of-225-trial selection, the edge is fee-fragile and decayed to negative precisely in the recent window the walk-forward OOS and holdout evaluate (so optimization would fit the strong 2021 and fail out-of-sample), and it is the confirmed-dead betting-against-beta / long-short major-perp class (L45, 0/84).

A long-short, beta-neutral, multi-instrument relative-value strategy on 12 Hyperliquid perpetuals (daily bars) harvesting the low-beta/low-volatility (betting-against-beta) anomaly — each week estimating every name's 30-day realized beta to BTC, going long an inverse-vol-weighted sleeve of the 3 lowest-beta names and short the 3 highest-beta names, then scaling the sleeves so the book's beta to BTC (not just its dollar exposure) is zero, using realized beta as a non-momentum selection signal to isolate the defensive factor premium and fill under-represented Hyperliquid/long-short/multi-instrument/macro-horizon buckets.

The backtest (2179 daily bars, 2020-2026, no liquidation) returned +57.7% over 744 trades and was genuinely beta-neutral (realized beta 0.043) with moderate drawdown (23.2%), positive alpha 0.042, avg_trade_return_pct 0.565% (above the fee floor), and lower kurtosis (4.08) than prior fat-tailed baskets. But the base Sharpe was only 0.285 with sharpe_ci_low -0.366 (CI straddles zero), profit factor a fee-fragile 1.076, impact cost 11% of gross on just $8.2M capacity, information ratio -0.60, and the edge was front-loaded (2021 +65.9% carries the whole headline) then decayed to negative (2022 -13.3%, 2023 -5.6%, 2026 -11.7%, rolling Sharpe -3 to -3.9).