HyperliquidAbsoluteMomentumLSBasketDaily

Time-Series (Absolute) Momentum Long-Short Basket on Hyperliquid Perps (Pure OHLCV, Managed-Futures Style, Weekly Rebalance, Trend-Flip Exit, Equal-Weight Net-Directional)

A LONG-SHORT, MULTI-INSTRUMENT TIME-SERIES (absolute) MOMENTUM basket across the most liquid Hyperliquid perpetuals (BTC, ETH, SOL, and 4-6 other HL majors with confirmed multi-year history). Each asset is independently held LONG when its OWN trailing trend is up and SHORT when its own trend is down — pure managed-futures / trend-following, NOT cross-sectional ranking and NOT funding-based. This is a deliberate VENUE PORT of the factory's single most robust validated mechanism (the promoted AbsoluteMomentumLSBasketDaily, Sharpe 1.48 on Binance USD-M) onto Hyperliquid, where the portfolio is critically under-represented (5.2% vs 20% target) and where the factory's live trading is migrating. Hyperliquid's ~0.09% round-trip taker cost is LOWER than Binance USD-M (~0.10%), so a mechanism already net-profitable after Binance fees has a structurally wider margin here. Equal-weight per active leg, sized relative to $100k equity. Pure daily OHLCV — no funding rate, no options, no sub-1H grain, no cross-venue legs — so it sidesteps every recently-abandoned failure family (funding-carry fee-domination, options-unwired, sub-1H impulse fee drag).

Faithful venue port of the factory's single most robust validated mechanism (AbsoluteMomentumLSBasketDaily, Sharpe 1.48 on Binance USD-M) onto Hyperliquid, filling the under-represented HL bucket (5.2% vs 20%) where live trading is migrating. HL's ~0.09% round-trip taker cost is below Binance USD-M's ~0.10%, so a mechanism already net-profitable after Binance fees has a structurally wider margin here. Implementation is identical to the validated parent except for the equal-weight sizing the hypothesis specifies (the parent's inverse-vol simplified away — one fewer tunable to overfit) and the HL basket/instrument IDs. The basket is trimmed to BTC/ETH/SOL + 4 majors (AVAX, LINK, DOGE, LTC) all listed since HL's 2023 launch with confirmed multi-year daily history, minimizing data_unavailable risk while matching the '4-6 other majors' spec. All order management is centralized in calculate_signal (a continuously-varying close-to-close proxy is returned so Layer-2 frozen-signal detection passes while the inert should_enter/should_exit/position_size hooks keep single-instrument logic out of a basket strategy). The two robust-region clamps ([48,56] trend, [7,8] weekly rebalance) are applied in __init__ so neither the Optuna IS search nor the ±20% sensitivity perturbation can resolve back to the overfit short-trend / 6-day-cliff corners that produced IS 3.96 → OOS -2.44 on the parent. leverage=1.0 with gross capped at 1x equity keeps the port net-directional but un-leveraged, matching the parent's risk profile; sizing references self.leverage so it stays consistent if a future iteration raises it. Pure daily OHLCV sidesteps every recently-abandoned failure family (funding-carry fee-domination, options-unwired, sub-1H impulse fee drag, cross-venue fee floor).

Failed deflated Sharpe: DSR=0.0, expected-max=6.60 over 225 trials — the optimized Sharpe of 0.44 is far below the best-of-N luck bar and statistically indistinguishable from selection noise (is_significant=false, PBO=0.9028>>0.5, sharpe_ci_low -0.76 straddles 0). The walk-forward is genuinely overfit: is_overfitted=TRUE with avg IS 2.07 collapsing to avg OOS -3.37 and windows [-9.40, -2.09, +1.38] — two strongly negative, so the HL port's edge does not generalize out-of-sample. The holdout's passed=true is a degenerate artifact: ratio=0 because the WF-OOS reference (-3.37) is negative, so holdout_sharpe 0.40 against a negative denominator is meaningless and does not rescue it. Sensitivity passing (1 cliff) is necessary but nowhere near sufficient against DSR 0, PBO 0.90, and deeply negative OOS. The optimized result is also outlier/regime-driven (tail_ratio 0.014, kurtosis 63, skew 6.3; annual 2024 +13.3 / 2025 -9.7 / 2026 +24.2), confirming the pre-optimization concern that the headline was carried by a handful of 2024/2026 bull-spike days rather than a generalizable edge. Not iterate: the iterate exception requires a not-overfit walk-forward AND a genuine positive holdout, and both fail here; the core edge params were already clamped to the parent's 'validated-robust' plateau and still produce -3.37 avg OOS, so there is no robust region to tune toward. Not revise_hypothesis: this is overfit (is_overfitted=TRUE, OOS collapse), not a structurally-dead target for a proven mechanism — the parent AbsoluteMomentumLSBasketDaily is already promoted on Binance USD-M and captures this edge, so re-pointing the same mechanism at another venue would just re-overfit. FAILURE PATTERN: venue-porting a Binance-promoted absolute-momentum LS basket onto Hyperliquid majors fails because HL's shorter, regime-concentrated history (deeply negative 2023, gains bunched in 2024/2026 bull spikes) yields a fat-tail in-sample Sharpe that does not survive deflation (DSR=0), is_overfitted=TRUE with negative walk-forward OOS, and a degenerate holdout — a promoted Binance mechanism does not automatically transfer to a thinner-history venue.

Long-short time-series (absolute) momentum basket across 7 liquid Hyperliquid perpetuals (BTCUSD primary + ETHUSD, SOLUSD, AVAXUSD, LINKUSD, DOGEUSD, LTCUSD), all 1-DAY bars. Each asset is independently held LONG when its own trailing trend (close/close[-1-trend_lookback]-1) is up and SHORT when down — pure managed-futures / CTA mechanism, net-directional, NOT cross-sectional and NOT funding-based. Equal-weight per active leg: every leg with a non-zero trend sign gets the same base notional (equity*gross_exposure/N), signed by its trend. Rebalances weekly via single netting delta orders (a trend flip reverses the leg in one order). Risk controls: per-leg adverse stop-loss checked every bar (closes the leg and blocks same-direction re-entry until the trend flips) and an aggregate gross-exposure cap that scales all legs down so total gross never exceeds ~1x equity. trend_lookback is hard-clamped to [48,56] and rebalance_days to [7,8] in __init__ so the optimizer/sensitivity sweep cannot land on the parent's known overfit corners. Pure daily OHLCV — no funding, options, sub-1H grain, or cross-venue legs.

Venue port of a promoted, validated mechanism (AbsoluteMomentumLSBasketDaily, Sharpe 1.48 on Binance USD-M) onto a fee-favorable venue (HL ~0.09% RT < Binance ~0.10%)

Trades implement the stated hypothesis: 86 long / 81 short, net-directional own-trend momentum across 7 HL majors, weekly rebalance — pattern and frequency match managed-futures premise

167 trades over ~5.7 years is a sufficient sample for meaningful walk-forward/holdout analysis

Positive after-fee expectancy (PF 1.16, commission only 1.2% of gross), pure daily OHLCV so it sidesteps recently-abandoned fee-domination / unwired-data families

Result is heavily concentrated in 2026 (+44.7%) with 2023-2025 roughly flat-to-negative — strong regime/recency dependence

Fat-tail driven: return_skew 7.0, kurtosis 72, tail_ratio 0.012 — a few outlier days carry the equity curve; sharpe_ci_low -0.95 straddles 0

Weak risk-adjusted baseline: Sharpe 0.42, max_drawdown 54%, 760-day max DD duration

Both edge tunables are hard-clamped to narrow ranges, so optimization can only move sizing/stop params — useful but bounded

PF 1.16, commission 1.2% of gross

viable

Mechanism is a venue port of a genuinely promoted parent (AbsoluteMomentumLSBasketDaily, Sharpe 1.48 on Binance USD-M); code is functioning (250 trades, balanced 124L/126S), low fee drag (commission 1.1-1.3% of gross)

Sensitivity phase passed with only 1 cliff (stop_loss_pct)

Failed deflated Sharpe: DSR=0.0, is_significant=false, PBO=0.9028 — selection is almost certainly overfit; optimized Sharpe 0.44 far below expected-max luck bar 6.60

Walk-forward is_overfitted=TRUE: avg IS 2.07 collapses to avg OOS -3.37 with windows [-9.40, -2.09, +1.38] — the edge does not generalize out-of-sample on HL

Holdout 'passed' is degenerate: ratio=0 because the WF-OOS reference is negative; holdout_sharpe 0.40 against a negative denominator is meaningless

Result is fat-tail/outlier-driven and regime-concentrated: tail_ratio 0.014, kurtosis 63, skew 6.3, annual 2025 -9.7% — exactly the 2024/2026-spike dependence flagged pre-optimization

sharpe_ci_low -0.76 straddles 0

HyperliquidAbsoluteMomentumLSBasketDaily ported the factory's most robust validated mechanism — a managed-futures own-trend long-short basket — from Binance USD-M to seven Hyperliquid majors, betting that lower HL fees and quota gaps justified the venue swap, with edge params clamped to the parent's plateau. The baseline looked plausible enough (PF 1.26, +56.7%, low fee drag) to earn optimization, but that exposed an overfit mirage: optimized Sharpe 0.44 far below the 6.60 luck bar, DSR 0.0, PBO 0.90, and a walk-forward whose 2.07 in-sample average collapsed to -3.37 out-of-sample with two deeply negative windows, the result carried by a handful of 2024/2026 bull-spike days. The analyst ruled it overfit rather than a structurally dead target — the parent already captures the edge on Binance — with the holdout pass a degenerate artifact. It ended after one iteration as abandoned, reaching optimization and analysis but never risk review.

A promoted Binance mechanism does not automatically transfer to a thinner-history venue: Hyperliquid's shorter, regime-concentrated record (deeply negative 2023, gains bunched in 2024/2026 bull spikes) yielded a fat-tail in-sample Sharpe that collapsed under deflation and walk-forward — clamping the edge params to the parent's robust plateau did not help, and a holdout 'pass' is meaningless when its OOS reference is negative; the parent already captures this edge on Binance, so re-pointing it at another venue just re-overfits.

It passed the pre-optimization backtest-review gate (verdict: optimize) but was abandoned at the post-optimization analyst gate (verdict: abandon): the optimized Sharpe was indistinguishable from best-of-N selection noise (DSR 0.0, PBO 0.90, not significant), the walk-forward genuinely overfit with deeply negative OOS, and the passing holdout was a degenerate artifact of a negative denominator — so it never advanced to risk review.

A long-short, multi-instrument time-series (absolute) momentum basket across seven Hyperliquid majors (BTC, ETH, SOL, AVAX, LINK, DOGE, LTC) on daily bars — holding each asset long when its own trailing trend is up and short when down, equal-weight per active leg with a per-leg stop and gross-exposure cap, weekly rebalance — a deliberate venue port of the factory's promoted AbsoluteMomentumLSBasketDaily (Sharpe 1.48 on Binance USD-M) onto fee-favorable, under-represented Hyperliquid with both edge tunables hard-clamped to the parent's validated-robust plateau.

The baseline backtest (250 trades, balanced 124 long / 126 short, commission ~1.1% of gross) showed a weak, outlier-driven edge — Sharpe 1.15 (CI [-1.35, 2.90]), profit factor 1.26, total return +56.7%, but 50.3% max drawdown, skew 3.1, kurtosis 15, and heavy regime concentration; optimization confirmed the failure: optimized Sharpe 0.44 against a 6.60 expected-max luck bar, DSR 0.0, PBO 0.90, walk-forward is_overfitted=true with avg IS 2.07 collapsing to avg OOS -3.37 (windows [-9.40, -2.09, +1.38]), and a degenerate holdout (ratio 0 against a negative reference).