HlBtcEthSolNetDirectionalTSMomentumWeekly

Hyperliquid BTC/ETH/SOL Net-Directional Time-Series Momentum Book (Long-Short, Per-Asset Trend Sign, Weekly Flatten-and-Reopen, OHLCV-Only)

A NET-DIRECTIONAL, multi-instrument time-series momentum book on the three deepest-history Hyperliquid perpetuals (BTCUSD, ETHUSD, SOLUSD). Each asset is independently held LONG when its own ~50-day trend is up and SHORT when down, equal-weighted, so the book's NET exposure floats with how many names are trending (net-long in broad bulls, net-SHORT in bears — the crash protection the analyst repeatedly identifies as what separates surviving forms from bull-beta long-only). This deliberately ports the ONE validated survivor — the NET-DIRECTIONAL momentum basket (absolute 1.36, dual 1.61) — and explicitly AVOIDS the forms confirmed dead this session: it is NOT single-asset BTC trend (decayed 2024-26 across funding/ER/Donchian tests), NOT funding-confirmed trend (ETH-specific, failed on SOL/BNB/BTC), NOT dollar-neutral/relative momentum or cross-asset pairs/cointegration MR (decayed/closed — SOL/AVAX, LTC/BCH, ETH/BTC all catastrophic), NOT options (Deribit engine block), NOT spot CASH (over-fill bug), NOT continuous per-bar vol-targeting (inflation artifact), and NOT raw tick microstructure (2-day data wall). It fills the largest venue gap (HYPERLIQUID 7.6% vs >=20%) plus the multi-scope and long_short buckets, using ONLY BTC/ETH/SOL whose single-asset HL TSMs already reached risk_review (proving HL history is deep enough to validate, sidestepping the HL data wall that killed thin-history daily strategies). Engine-safe by construction: discrete WEEKLY flatten-and-reopen to target (the analyst-credited clean sizing fix) + a per-asset hard notional cap, so no winner-balloon and no continuous-rebalance inflation. Two core parameters (trend lookback, rebalance cadence) to resist overfitting.

Ports the one validated survivor — the net-directional absolute-momentum basket — and avoids the forms confirmed dead this session: it is not single-asset BTC trend, not funding-confirmed trend, not dollar-neutral/relative momentum or cross-asset pairs/cointegration MR, not options, not spot CASH, not continuous vol-targeting, not tick microstructure. Net-directionality comes from per-asset OWN-trend signs (absolute TSM), giving net-short exposure in bear regimes (the crash protection the analyst credits). Engine-safety is by construction via the analyst-credited clean-sizing fix: discrete weekly flatten-and-reopen to a freshly-equity-derived target with netting deltas (no winner-balloon, no continuous-rebalance inflation) plus a per-asset hard notional cap. The basket is exactly BTC/ETH/SOL on Hyperliquid — the three names whose single-asset HL TSMs already reached risk_review, proving HL daily history is deep enough and sidestepping the thin-history HL data wall. It fills the HYPERLIQUID venue gap (7.6% -> >=20%) plus the multi-instrument and long_short buckets. Implemented on the proven AbsoluteMomentumLSBasketDaily structure (centralized order management in calculate_signal, exception-guarded cache/order calls so the Layer-2 synthetic proxy never crashes, and a continuous tanh net-trend signal so frozen-signal detection passes), simplified to equal-weight + two core parameters as the hypothesis specifies. leverage=1.0 (gross peaks ~0.9x when all three align) and is still referenced in sizing so there is no leverage_set_but_unused mismatch.

Failed deflated Sharpe and every generalization gate in concert: DSR=0.0 (<<0.95), is_significant=false, sharpe_ci_low=-0.32 (CI straddles 0), and the optimized Sharpe 0.91 is far below the expected-max-under-noise bar of 6.81 over 225 trials — statistically indistinguishable from best-of-N selection. PBO=0.8928 (~89% overfit probability), the holdout FAILED and is NEGATIVE (holdout_sharpe -0.50, ratio -0.5), walk-forward is_overfitted=TRUE (avg IS 4.08 -> avg OOS 1.005 with OOS windows [0.0, 3.49, -0.47], two of three zero/negative), and sensitivity FAILED with cliffs on BOTH core edge parameters (trend_lookback, rebalance_days) over a uniformly modest 0.0-0.86 heatmap. The eye-catching optimized headline (+182% return, CAGR 41.7%) is a fat-tail, 2024-bull-concentrated artifact (kurtosis 89, tail_ratio 0.05, annual 2024 +88%) on only ~2.7 years (677 data-days) of Hyperliquid history. PRESERVED FINDING: the book is genuinely net-directional (beta 0.019, alpha +0.15, balanced 65L/64S) and the weekly flatten-and-reopen sizing is clean (metrics_reliable=true, no winner-balloon) — the hypothesis's engine-safety and market-neutrality design wins hold; only the trend edge fails to validate. NOT iterate (attempt 1 of 2): there is no robust region to tune toward — both core params are cliffs, the sensitivity grid is modest everywhere, PBO is 0.89 and DSR is 0.0, so a second sweep re-selects another best-of-N artifact on the same thin, regime-concentrated data. NOT revise_hypothesis: this is overfitting + multiple-testing failure with a negative holdout, not a proven mechanism stranded on a single dead target — the net-directional/TS-momentum edge is already captured by a promoted sibling (single-asset SOL TS-momentum long/short reached paper_stage), so its value is not lost, and re-pointing the same daily-trend mechanism at another HL configuration would re-hit the identical short-history / 2024-bull-concentration wall that has defeated every HL daily-trend instance this session. FAILURE PATTERN: multi-asset net-directional daily TS-momentum on Hyperliquid majors (BTC/ETH/SOL) produces a flashy 2024-bull-carried optimized headline (Sharpe ~0.9, +182%) that collapses under multiple-testing deflation (DSR 0.0, PBO 0.89, Sharpe far below the luck bar), fails a negative holdout (-0.50), is_overfitted=TRUE with zero/negative OOS windows, and has cliffs on both core params — HL's ~3-year history is too short and too regime-concentrated to validate a low-frequency daily-trend edge, even in the genuinely market-neutral basket form; the mechanism's value is preserved in promoted single-asset siblings, not here.

Net-directional time-series (absolute) momentum book on the three deepest-history Hyperliquid perpetuals (BTCUSD primary, ETHUSD, SOLUSD), all 1-DAY bars. Each asset is independently held LONG when its own ~50-day trend is up and SHORT when down, equal-weighted, so the book's NET exposure floats with how many names are trending: net-long in a broad bull, net-short in a bear (the structural crash protection vs bull-beta long-only). Sizing is fully discrete: every rebalance_days (weekly) targets are recomputed from CURRENT equity and each leg is moved to its fresh signed target (sign * per_asset_frac * equity, hard-capped at max_notional_frac * equity) via a single NETTING delta order, so a trend flip reverses the leg in one order and a winner cannot balloon across rebalances. There is no continuous per-bar vol-targeting (avoids the inflation artifact) and a per-asset hard notional cap prevents over-deployment. All multi-leg order management is centralized in calculate_signal; the single-instrument hooks are inert. Two core tunables (trend_lookback, rebalance_days) keep it overfitting-resistant. Pure OHLCV, Hyperliquid, long-short, leverage 1.0.

Port of the validated-survivor net-directional momentum basket; trades match the hypothesis (balanced 42 long / 39 short, per-asset trend sign, weekly rebalance, ~38d holding)

Genuinely market-neutral, not bull-beta: beta 0.033, alpha +0.054, benchmark_correlation 0.074 — captures the net-short-in-bear crash protection

No artifact signature: Sharpe 0.41 is realistic (not implausible), end_unrealized_pct 3.24 (headline is realized), single-venue HL (no cross-venue leg-netting), metrics_reliable=true, avg_position_pct 33.5% matches per_asset_frac (clean weekly sizing)

Adequate sample (677 data-days, 81 trades) on HL-validated majors; only 2 core tunables — optimizable without fitting noise

Weak base Sharpe 0.41 with sharpe_ci_low -0.97 (CI straddles 0) and information_ratio -0.49 — real risk of failing Deflated Sharpe after 225-trial selection

Heavy fat-tail / period concentration: skew 9.97, kurtosis 168, tail_ratio 0.15, cvar_95 466%, return carried by Feb-2024 (+52%) and 2026; mid-2024 chopped for ~5 straight negative months

High risk: max_drawdown 64%, max_consecutive_losses 20, annualized_volatility 49%, ~3x gross exposure (302%) — likely over-leveraged, hurting risk-adjusted return

42L/39S

balanced

Mechanism is genuinely net-directional, not bull-beta: beta 0.019, alpha +0.15, balanced 65 long / 64 short; clean weekly flatten-and-reopen sizing (metrics_reliable=true, avg_position_pct ~36%, no winner-balloon)

Engine-safe construction as designed (discrete weekly rebalance, per-asset cap); these design wins hold up

Deflated Sharpe 0.0 (<<0.95), is_significant FALSE, sharpe_ci_low -0.32 (straddles 0); optimized Sharpe 0.91 far below the 6.81 expected-max-under-noise bar over 225 trials

PBO 0.8928 — ~89% probability the selection is overfit

Holdout FAILED and negative: holdout_sharpe -0.50, ratio -0.5

Walk-forward is_overfitted=TRUE: avg IS 4.08 -> avg OOS 1.005, OOS windows [0.0, 3.49, -0.47] (two of three zero/negative)

Sensitivity FAILED with cliffs on BOTH core edge params (trend_lookback, rebalance_days); heatmap uniformly modest 0.0-0.86 with no robust region

Return is fat-tail / 2024-bull concentrated (kurtosis 89, tail_ratio 0.05, annual 2024 +88%) on only ~2.7yr of HL history; max_consecutive_losses 10/20

fragile (one window carries it; recent OOS -0.47)

HlBtcEthSolNetDirectionalTSMomentumWeekly was the careful culmination of the session's lessons — a deliberate port of the validated net-directional momentum survivor onto Hyperliquid's three deepest majors, engineered to be engine-safe (weekly flatten-and-reopen, per-asset cap) and genuinely market-neutral rather than bull-beta. Those design goals were met: the book traded balanced long/short, sized cleanly with no winner-balloon, and carried real net-short-in-bear crash protection (beta ~0.02, alpha +0.15). But the trend edge did not validate: the optimized +182% headline was a fat-tail 2024-bull artifact that collapsed under deflation (DSR 0.0, PBO 0.89), failed a negative holdout, was is_overfitted with zero/negative OOS windows, and had cliffs on both core parameters. The analyst abandoned it as the same short-history / regime-concentration wall that defeated every HL daily-trend instance this session — preserving the engine-safety and market-neutrality design wins, and noting the mechanism's value already lives in a promoted single-asset SOL TS-momentum sibling, not here.

Even the genuinely market-neutral, engine-safe net-directional basket form cannot validate a daily-trend edge on Hyperliquid's ~3-year, 2024-bull-concentrated history — the design wins (net-short-in-bear crash protection, clean weekly sizing) are real and preserved, but the trend edge is captured instead by promoted single-asset siblings, so re-pointing it at another HL configuration just re-hits the short-history wall.

After optimization the analyst abandoned it because it failed deflated Sharpe and every generalization gate in concert: DSR 0.0 vs an expected-max-luck bar of 6.81, PBO 0.89, a negative holdout (-0.50), walk-forward is_overfitted=TRUE (avg IS 4.08 → avg OOS 1.005, windows [0, 3.49, -0.47]), and sensitivity cliffs on both core parameters over a uniformly modest grid — Hyperliquid's ~2.7-year, 2024-bull-concentrated history being too short to validate a low-frequency daily-trend edge.

A net-directional, multi-instrument time-series momentum book on the three deepest-history Hyperliquid perps (BTC/ETH/SOL), holding each asset long or short by its own ~50-day trend sign, equal-weighted, with discrete weekly flatten-and-reopen sizing and a per-asset notional cap — porting the one validated-survivor net-directional form so the book goes net-short in bears for crash protection, with only 2 core parameters.

The default backtest was weak (Sharpe 0.41, +36.6% total return, profit factor 1.15, 64.5% max drawdown, 20 max consecutive losses over 81 balanced trades) but genuinely market-neutral (beta 0.033); optimization produced an eye-catching headline (Sharpe 0.91, +182.7%, CAGR 41.7%, PF 1.52, 30.2% drawdown) that proved to be a fat-tail, 2024-bull-concentrated artifact (kurtosis 89, 2024 +88%).