AvaxHlDailyReturnSkewnessRegimeLongShort

AVAX Hyperliquid Perpetual Daily-Return Skewness-Regime Long-Short (Higher-Moment Signal + Trend Confirmation)

Single-instrument long-short strategy on AVAXUSD.HYPERLIQUID perpetual using daily bars. Signal combines 30-day rolling skewness of daily log returns (third moment, regime indicator) with a 50-day SMA trend filter (first moment, direction confirmation). The hypothesis: persistent return-distribution skewness reflects ongoing positioning/leverage regimes that have momentum — sustained negative skew (large losses dominating) marks ongoing deleveraging cycles that continue; sustained positive skew (large gains dominating) marks momentum bursts that continue. Higher-moment regimes persist across multi-week horizons (Boyer/Mitton/Vorkink 2010 documented this for equities; same dynamics apply in crypto via leverage-cycle mechanics). Targets the HYPERLIQUID venue (4.1% vs >=20% quota target — biggest absolute venue gap after COIN-M became unverifiable) AND long-short direction (12.8% vs target — long-only dominates at 87.2%) simultaneously. Mechanistically NOVEL for the portfolio: the only prior skewness strategy was cross-sectional (BinanceTop3CrossSectionalSkewnessLS); this is single-asset time-series skewness, a different mechanism. Avoids all session failure modes: (1) not classical TA pattern (no candle shapes, no breakouts); (2) not first-moment momentum (already done as BTC/ETH TSM); (3) not ratio mean-reversion; (4) not vol-expansion entry (vol-clustering edge is different from skew-regime); (5) zero supplementary data dependency (pure OHLCV from daily bars); (6) STATIC single instrument (avoids the COIN-M verification-resolution failure mode — Hyperliquid AVAX perp is a known catalog instrument); (7) only 4 parameters total (skew lookback, skew threshold, SMA lookback, max hold) — chosen ex-ante from literature rounds, minimizes overfit risk.

The backtest reached Layer 3; signal/sizing logic already passed earlier layers, so this is the smallest additive change that fixes the reported data-coverage defect. Issue #1 was that Hyperliquid's AVAX perp fed flat/stale bars for ~2.5 years (2021-01 -> 2023-02), every daily return exactly 0.0, producing 26 degenerate zero-PnL trades and an all-zero early walk-forward zone. A strategy cannot change the optimizer's configured date range, but it CAN refuse to trade dead data: the new _data_live() gate requires the recent skew window to have realized daily-log-return std >= min_recent_vol (1%) AND >= min_nonzero_frac (70%) non-zero returns before any entry, which is true for genuine AVAX HL data (daily vol ~4-7%, ~100% non-zero) and false for the pre-2023 flat/forward-filled zone. A price-based liveness floor is more robust than hardcoding a 2023-04 start: it adapts to wherever real data actually begins and survives any future collector backfill. This directly addresses the symptom (trades booked on stale bars) and supports issue #2's outlier/expectancy review by ensuring every booked trade sits on live data, so the analyst's positive-per-trade-expectancy and not-outlier-dependent checks are computed on the real 2023+ sample. All skew, SMA, entry-direction, exit (regime flip / trend flip / time stop), and sizing logic is byte-identical to the previous iteration, so Layers 1-2 stay green and Layer 3 still produces trades over the real-data period. Venue stays HYPERLIQUID per the hypothesis (the point is to fill the HL quota gap), and leverage stays 1.0 referenced only as a non-multiplier so no unused-leverage gate applies. Note for the analyst per the decision rule: if the clean 2023+ window is too thin (~43 trades/3yr) or the edge collapses excluding the 2-3 largest days, abandon; otherwise sweep only the 4 core params.

Failed deflated Sharpe: DSR=0.0 (vs 0.95 bar), expected-max=7.215 over 225 trials — the optimized Sharpe of 1.83 is FAR below the best-of-N luck bar, statistically indistinguishable from selection noise (is_significant=false), with PBO 0.7884 and sharpe_ci_low -0.1662 (CI straddles 0). Root cause is fat-tail outlier dependence: optimized return_kurtosis 104.4, return_skew 9.16, tail_ratio 36.55, with the entire result carried by ~2 days (2024-12-06 +51.5%, largest_win $68.8K vs avg_win $14.7K) on just 30 trades — DSR's skew/kurtosis adjustment correctly crushes it to 0 and the edge is concentrated in 2024 (+51.4%) then decays. The PASSING holdout (2.15) and clean sensitivity (0 cliffs) cannot rescue it: a holdout on a 2-tail-day edge is fragile to a single tail flip, and a smooth in-sample surface measures consistency, not plausibility. Not iterate: sensitivity already passes with 0 cliffs (optimizer on its robust plateau), so there is no under-explored region to tune toward, and one walk-forward OOS window already produces ZERO trades — a second best-of-225 sweep re-selects another sparse, outlier-carried config below the 7.215 luck bar. Not revise_hypothesis: single-asset time-series skewness has no proven sibling stranded on a dead target, and the premise does not survive deflation on AVAX. FAILURE PATTERN: single-asset higher-moment (skewness-regime) long-short on a high-beta alt perp optimizes to a flashy Sharpe (1.83) with a passing holdout and clean sensitivity, yet fails deflated Sharpe (0.0, far below a 7.215 luck bar) with PBO 0.79 and a CI straddling 0, because the result is carried by a handful of fat-tail pump days (kurtosis 104, tail_ratio 36.5) on only 30 trades — the same outlier-luck signature as the abandoned SOL Keltner (kurtosis 124), NEAR Supertrend (kurtosis 175), and DOGE drawdown-accumulation siblings. The PSR=0.9975 / DSR=0.0 split is the diagnostic tell; a passing holdout and a zero-cliff grid do not validate an edge resting on 2 days.

Single-asset time-series skewness-regime long-short on AVAXUSD.HYPERLIQUID perpetual, daily bars. The continuous signal is the 30-day rolling skewness of daily log returns; a 50-day SMA confirms direction. It goes long when skew >= +0.5 and close > SMA (positive-skew momentum burst in an uptrend), short when skew <= -0.5 and close < SMA (negative-skew deleveraging in a downtrend), and exits on skew-sign flip, price crossing the SMA, or a 20-bar time stop. Iteration 2 adds a data-liveness gate: entries are blocked unless the recent return window shows genuine activity (realized daily-vol >= 1% and >= 70% non-zero daily returns), so the strategy does not trade on the flat/stale Hyperliquid AVAX bars that precede the perp's real ~2023 trading history. Pure OHLCV, no supplementary data, leverage 1.0, four core parameters plus the liveness floor.

Distinguishable edge pre-optimization: Sharpe 1.78 with sharpe_ci_low +0.198, PF 2.22, tail_ratio 2.30, asymmetric payoff (avg_win 2.1x avg_loss)

Iteration-2 data-liveness gate fixed the stale-HL degenerate-trade problem; real 2023-2026 window has 71 trades (~24/yr), long+short (34/37)

Novel single-asset time-series skewness mechanism, fills HL-venue and long-short quota gaps, deployable capacity ($50M, impact 4.4%), only 4 parameters

Positive across recent years (2023 +15%, 2024 +54%, 2025 +50%)

Short HL sample (~3 years of real data) — high cross-window Sharpe variance inflates the deflated-Sharpe expected-max luck bar; this is exactly what failed the HL SOL TSM sibling (DSR 0.0)

Outlier-influenced edge: kurtosis 41, win_rate 32% — much of the return rides a few large moves (2024-11-26 +28.7%); deflated Sharpe will discount it

2026 -4.1% (partial year) — holdout must confirm recent-regime persistence

Sensitivity PASSED with 0 cliffs (smooth ~1.3-1.8 Sharpe surface across the grid)

Walk-forward is_overfitted=FALSE; holdout PASSED (holdout_sharpe 2.15, ratio 1.089)

Hits two portfolio gaps simultaneously (HYPERLIQUID venue + long-short direction); zero supplementary-data dependency

Failed deflated Sharpe: DSR=0.0 (vs 0.95), is_significant=FALSE; optimized Sharpe 1.83 is FAR below the expected-max luck bar 7.215 over 225 trials — indistinguishable from best-of-N selection noise

PBO 0.7884 (79% overfit probability); sharpe_ci_low -0.1662 (CI straddles 0 from -0.17 to +3.26)

Outlier/fat-tail dependence: return_kurtosis 104.4, return_skew 9.16, tail_ratio 36.55; the entire optimized return is carried by ~2 days (2024-12-06 +51.5%, largest_win $68.8K vs avg_win $14.7K). DSR's skew/kurtosis adjustment correctly discounts this to 0

Sparse and unstable: only 30 optimized trades, one walk-forward OOS window with ZERO trades (liveness gate starves it), one IS window also 0; optimized max_DD 24.8% with a 443-day drawdown duration

Edge is concentrated in 2024 (+51.4%) and decays (2025 +5.6%, 2026 +7.6%); information_ratio -0.42 (no benchmark-relative edge)

Do NOT optimize yet — two structural problems must be resolved first. (1) DATA COVERAGE: every daily return from 2021-01 through 2023-02 is exactly 0.0 (26 dated zero-PnL trades; annual_returns 2021=0.0, 2022=0.0). Hyperliquid's AVAX perp did not meaningfully trade before ~2023, so the catalog is feeding flat/stale bars for the first ~2.5 years of the configured 2020-09-23 -> 2026 range. A 3-window walk-forward would put its first window(s) entirely in this dead zone -> degenerate all-[0,0,0] OOS / 0.0 holdout-reference failure regardless of the real edge, wasting the 2-hour run. ACTION: confirm the on-demand collector is returning REAL HL data for the full range (it almost certainly is not), or restrict the evaluation to the genuine HL data period (~2023-04 onward) so the walk-forward sees real bars end to end. (2) OUTLIER/EXPECTANCY: return_kurtosis 41.9 and a -$331/trade expectancy (30% win rate) alongside a +136% compounded return suggest the positive equity may be a sizing artifact riding 2-3 outlier days. After fixing the data window, confirm per-trade expectancy is positive on the real 2023+ sample and the edge survives excluding the 2-3 largest days. DECISION RULE: clean real-data window with positive per-trade expectancy and not-outlier-dependent -> proceed to optimization (sweep only the 4 core params); too thin (~43 trades/3yr) or edge collapses without outliers -> abandon. Detailed feedback written to workspace/discussions/117570e0-570c-4324-a741-976e627efeaa/iteration_1_feedback.md

AvaxHlDailyReturnSkewnessRegimeLongShort introduced a novel single-asset time-series skewness-regime long/short on Hyperliquid AVAX, simultaneously filling the under-represented Hyperliquid-venue and long-short quotas, and its iteration-2 liveness gate fixed the stale-bar problem to yield a clean 158% base backtest that earned an 'optimize' verdict. But optimization exposed an outlier-carried edge: a deflated Sharpe of 0.0 far below a 7.215 luck bar, PBO 0.79, a CI straddling zero, and a result driven by ~2 tail days (kurtosis 104, tail ratio 36.5) over only 30 trades, with a zero-trade OOS window from the liveness gate. The passing holdout and smooth sensitivity surface measured consistency, not plausibility, and could not rescue a 2-day edge on a short HL sample. On its second iteration the analyst abandoned it, citing the same outlier-luck signature as the abandoned SOL Keltner, NEAR Supertrend, and DOGE siblings.

A passing holdout and a zero-cliff sensitivity grid cannot validate an edge resting on a handful of fat-tail pump days — on a short ~3-year sample with extreme cross-window Sharpe variance, deflated Sharpe's skew/kurtosis penalty correctly crushes the headline to zero, the PSR-high/DSR-zero split being the diagnostic tell of outlier luck.

Despite a passing holdout (Sharpe 2.15), clean sensitivity (0 cliffs), and a non-overfit walk-forward, the analyst abandoned it post-optimization because the deflated Sharpe was 0.0 against an inflated 7.215 luck bar (is_significant=FALSE), PBO was 0.79, the CI straddled zero (low -0.166), and the entire optimized return rode ~2 fat-tail days (kurtosis 104, tail ratio 36.5) on just 30 trades — with one walk-forward OOS window producing zero trades.

A single-instrument, OHLCV-only long/short higher-moment strategy on Hyperliquid AVAXUSD perpetual daily bars — combining 30-day rolling skewness of daily log returns (a positioning/leverage-regime indicator) with a 50-day SMA trend filter: going long on positive skew above the SMA and short on negative skew below it, exiting on a skew flip, trend flip, or time stop, with a data-liveness gate to block stale early-Hyperliquid bars and 4 parameters.

After the iteration-2 liveness gate fixed degenerate stale-bar trades, the real 2023-2026 window gave a respectable base backtest: 158.3% return (CAGR 18.8%), Sharpe 1.78 (CI low +0.198), profit factor 2.22, 71 trades (37 long, 34 short), and an 18.1% drawdown. Optimization kept the headline strong (Sharpe 1.83, PF 2.79) but thinned the book to 30 trades with kurtosis 104 and tail ratio 36.5.