HyperliquidDiversifiedTsmomVolTargetedBasketLS

Hyperliquid Diversified Time-Series (Absolute) Momentum — Vol-Targeted Managed-Futures Book Across ~15 Perps: Each Name Independently LONG or SHORT by the Sign of Its Own Trailing Trend, Inverse-Vol Risk-Weighted to Equal Risk Contribution, Book Scaled to a Target Volatility, Weekly Rebalance (HYPERLIQUID, Long-Short, 2-Parameter)

A LONG-SHORT, MULTI-INSTRUMENT ABSOLUTE (time-series) MOMENTUM book on Hyperliquid perps — a diversified managed-futures / CTA-style trend strategy, NOT a single-name trend and NOT a cross-sectional ranking. For each of ~15 liquid HL perps (≥2y daily history + adequate ADV), the sign of its OWN trailing `lookback`-day return sets its position: uptrend → long that name, downtrend → short that name, small dead-band → flat (avoids churn near zero). Each name is INVERSE-VOL weighted (size ∝ 1/trailing realized vol) so every name contributes equal risk, and the whole book is scaled to a target annualized volatility (`vol_target`). Rebalance weekly: flip a name only when its trend sign changes, and resize only when the vol-target drifts outside a no-trade band. This is STRUCTURALLY DISTINCT from everything in my pipeline: it is NOT the cross-sectional momentum basket (that is RELATIVE rank, dollar-neutral, market-neutral; THIS is absolute per-name sign, net exposure floats with the aggregate trend and can be net-long, net-short, or hedged); NOT funding carry / cash-and-carry / VRP options (no funding, no options, no basis); NOT the single-name OHLCV trend class that dies 0/213 (L9/L41/L62 are about ONE symbol or symbol-swapping a fixed rule — here the LOAD-BEARING edge is the diversification of ~15 independent trend bets plus vol-targeting, which is precisely why diversified TSMOM survives where single-name trend does not); and NOT the L60 'always-in multi-factor composite' (a composite is always-in and averages noise toward BETA — TSMOM is directional-ADAPTIVE, it goes flat or NET-SHORT the entire book in downtrends, which is the opposite of always-in beta). The genuine, documented value it adds to THIS portfolio: it is a LONG-SHORT trend book that delivers CRISIS ALPHA — it flips net-short in sustained bear regimes (e.g. 2022) and profits while the portfolio's 85.8%-long-only strategies draw down. It fills the two largest bucket gaps at once (HYPERLIQUID 6.8% vs ≥20%; long-short 14.2% vs 85.8% long-only) plus multi-instrument scope. L61 feed check: PURE OHLCV daily bars only (no supplementary feed, no dated future) on HL majors/liquid alts that have full multi-year DAILY history — avoids the sub-hourly HL data-span trap (daily = full history back to listing) and the data_unavailable deaths entirely. TIMEOUT-SAFETY: signals are trailing returns + rolling vol, computed incrementally (rolling deques), O(1) per bar — no per-bar rescan/sort. FEE ARITHMETIC (checklist #1): a ~90-day trend sign flips only ~2–4x/yr per name; each captured trend leg on a trending crypto perp is typically several % of notional, versus a ~0.09% HL round trip — >20x the fee on realized legs — and vol-target resizes are gated by a no-trade band so tiny rebalances don't bleed fees. ~15 names × ~3 flips/yr × multi-year history ⇒ ~130+ completed trades, well above the measurability floor for walk-forward + holdout. This is emphatically NOT a sub-fee fast-cadence mechanism.

Implements the hypothesis exactly: absolute per-name TSMOM sign (not cross-sectional rank), inverse-vol equal-risk weighting, book-level volatility targeting, weekly rebalance, ~15 HL perps, pure daily OHLCV with no supplementary feed. Universe selection was verified against the catalog: all 15 chosen HYPERLIQUID perps have continuous 1-DAY history from 2020-2021 through 2026-08 and 8-dp size precision, so there is no data_unavailable or round-to-zero risk (L61/L4). Two core tunables (lookback, vol_target) are exposed as the hypothesis specifies; the rest of the construction is frozen, and every parameter my code clamps is declared in _param_bounds. Timeout safety: per-bar work is O(1) (a single list index for the trailing return); the only O(n_assets * vol_lookback) work runs once per calendar week at the rebalance, so there is no per-bar rescan or sort. Blow-up safety: the vol scalar is capped at 1.0 and leverage is 1.0, so gross notional can never exceed equity - the strategy cannot reach the -100% sizing failures seen in leveraged neutral books (L59/L15). Fee arithmetic: a 90-day trend sign flips ~2-4x/yr per name, so ~15 names over 5+ years of HL daily data gives well over 100 completed legs, each a multi-percent move against a ~0.09% HL round trip; the no-trade band prevents small vol-target drifts from churning. The rebalance schedule is derived purely from the bar's own timestamp (days-since-epoch // rebalance_days), never from a bar counter, so it is restart- and replay-identical. One caveat as required: the load-bearing assumption is that per-name daily trend persistence survives on crypto perps at 90 days - I expect the diversification and vol-targeting to matter more than the exact lookback, which is why lookback and vol_target are the two search dimensions.

No measurable edge before optimization in a 0/84-survival class (long-short major-perp basket, L45). Sharpe 0.279 with CI straddling zero (low -0.398) is below the 0.5 OOS floor; information_ratio -0.542 (worse than the equal-weight benchmark) and PF 1.14 in the class fail band. The 37.4% headline over ~6 years is regime/outlier-driven (kurtosis 7.05, a handful of +13-17% weeks plus 13.1% end-unrealized MTM), and the claimed crisis alpha fails: 2022 +9.6% is more than erased by 2023 -20.3%. Not tunable — the base Sharpe is the ceiling for an always-adaptive book, so ~225-trial deflation + holdout would only curve-fit the outlier weeks and lower it. Failure pattern: long_short_major_perp_basket_no_edge (Sharpe CI straddles zero, PF in fail band, negative IR, outlier-driven returns).

Diversified time-series (absolute) momentum managed-futures book across 15 liquid HYPERLIQUID perps (BTC, ETH, SOL, BNB, XRP, DOGE, AVAX, LINK, LTC, ADA, ATOM, AAVE, DOT, NEAR, UNI) on 1-DAY bars, pure OHLCV. Each name is independently LONG or SHORT by the sign of its OWN trailing 90-day return, with a 2% dead-band that keeps it flat near zero. Active names are inverse-vol weighted (equal risk contribution, per-name weight capped at 20% and renormalized) and the whole book is scaled by a bounded vol-target scalar clip(vol_target/portfolio_vol, 0, 1) so gross exposure never exceeds equity. Net exposure floats with the aggregate trend: net-long in broad uptrends, net-SHORT in sustained bear regimes (crisis alpha), hedged when trends conflict. Rebalance on a calendar-anchored weekly grid (days-since-epoch // 7 from the bar's own timestamp); a name only flips when its trend sign changes and same-side resizes are gated by a 30%-of-target no-trade band, so fee drag stays far below the multi-percent trend legs being captured.

Acceptable for daily BTC bars; optionally also check the calendar grid from on_extra_bar so a missing BTC bar cannot silently stall rebalancing.

The entire book is rebalanced only when a PRIMARY (BTCUSD) daily bar arrives, since _rebalance() runs inside calculate_signal() which the base dispatches on primary bars only. A missing/gapped BTC bar stalls the whole book's weekly rebalance even if the other 14 names have fresh bars. BTC gaps are unlikely, but book cadence is coupled to one instrument's stream.

Ensure the full backtest spans the multi-year HL daily history so walk-forward + holdout get the ~130+ trades and both regimes.

Sandbox covered only data_days=364 (one year); long_win_rate 10.5% vs short_win_rate 71% is regime-skewed over a single year. The crisis-alpha and >100-trade claims depend on the full multi-year window. This is an analyst edge/robustness question, not a code defect.

Well-engineered, timeout-safe implementation with genuine per-name absolute-momentum sign logic, inverse-vol weighting and vol-targeting

Measurable sample: 502 trades over 2181 days (253 long / 249 short), full multi-year daily HL history — not a low-trade-count or data-span failure

Fills real portfolio bucket gaps (Hyperliquid, long-short) in principle

Sharpe 0.279 with sharpe_ci_low -0.398 — CI straddles zero, below the 0.5 OOS floor before optimization runs

information_ratio -0.542: worse than the equal-weight basket benchmark; no active edge demonstrated

profit_factor 1.14 sits in the class fail band (<1.2); fee-fragile with 5.18% impact_cost_pct

Returns carried by a few outlier weeks (kurtosis 7.05) plus 13.1% end-unrealized MTM, not a persistent edge; 2023 -20.3% erases the 2022 'crisis alpha'

Class prior: long-short major-perp basket (L45, 0/84 survival) with mediocre PF + 33.7% drawdown — the exact abandon signature

HyperliquidDiversifiedTsmomVolTargetedBasketLS aimed to bring a classic diversified managed-futures / TSMOM book to Hyperliquid — 15 perps each independently long or short by its own trailing trend, inverse-vol weighted and scaled to a target volatility — arguing that diversification across ~15 trend bets is the load-bearing edge that lets it survive where single-name trend dies, and that it would deliver crisis alpha by flipping net-short in bears. The implementation was clean and produced a healthy, measurable sample (502 trades over 2181 days), but there was no edge: Sharpe 0.279 with a CI straddling zero, information ratio -0.54 (worse than the equal-weight basket), PF 1.14, and a 33.7% drawdown, with the crisis-alpha claim falsified as 2023's -20.3% erased 2022's gain. The analyst abandoned it on the first iteration at the backtest_review gate, applying the L45 long-short major-perp basket prior (0/84 survival) since the base Sharpe was the ceiling and optimization would only curve-fit outlier weeks. It never advanced to optimization, analysis, or risk review.

Diversification and vol-targeting do not by themselves manufacture a trend edge on a crypto major-perp basket — a Sharpe near zero with a CI straddling zero, a negative information ratio, and outlier-driven returns means there is no active edge to optimize, and the claimed crisis alpha must actually persist across bear regimes (2022's gain was erased by 2023) rather than appear in a single window.

It was abandoned at the pre-optimization backtest_review gate on iteration 1: despite a well-engineered, timeout-safe implementation and a measurable 502-trade sample, the analyst found no edge before optimization — Sharpe below the 0.5 OOS floor with a CI straddling zero, negative information ratio, and PF in the class fail band — in the long-short major-perp basket class (L45) with 0/84 prior survival, where the base Sharpe is the ceiling for an always-adaptive book and further optimization would only curve-fit the outlier weeks.

A long-short, multi-instrument diversified time-series (absolute) momentum book across 15 liquid Hyperliquid perps on daily bars, where each name is independently long or short by the sign of its own trailing 90-day return (with a dead-band), inverse-vol weighted to equal risk contribution and scaled to a target annualized volatility, rebalanced weekly — a CTA-style managed-futures book aimed at delivering crisis alpha by flipping net-short in sustained bear regimes, with two core parameters (lookback and vol_target).

The backtest was a measurable but edgeless result: total_return 37.4% over ~6 years with Sharpe 0.279 (CI low -0.398, straddling zero), profit_factor 1.14, win_rate 0.40, information_ratio -0.542 (worse than the equal-weight benchmark), and max_drawdown 33.7% across 502 trades (253 long / 249 short). Returns were carried by a few outlier weeks (kurtosis 7.05) plus 13.1% end-unrealized MTM, and the claimed crisis alpha failed — 2022's +9.6% was more than erased by 2023's -20.3%.