LinkHyperliquidVolTargetedTimeSeriesMomentumLS

LINK Hyperliquid Volatility-Targeted Time-Series Momentum, Long-Short (Single-Instrument LINKUSD.HYPERLIQUID Perp, 4H Bars, Classic Sign-of-Trailing-Return TSMOM with MA Confirmation, DYNAMIC Volatility-Targeted Position Sizing — Bigger in Calm Trends / Smaller in Chaos, 3-Parameter)

A LONG-SHORT, single-instrument, pure-OHLCV strategy on LINKUSD.HYPERLIQUID that implements the single most-replicated anomaly in the academic literature — TIME-SERIES (absolute) MOMENTUM (Moskowitz-Ooi-Pedersen) — with the standard robustness overlay that makes it survive out-of-sample: DYNAMIC VOLATILITY TARGETING of position size. This is deliberately differentiated from the five momentum strategies already in this factory's pipeline (SOL Donchian breakout, ETH RSI pullback, AVAX efficiency-ratio gate, BTC dual-TF confluence, SOL BTC-regime gate) in TWO ways they all share and this one changes: (a) the signal is the plainest possible robust momentum — the SIGN of the trailing K-bar return, confirmed by price vs a moving average — rather than a bespoke breakout/pullback/gate construction that invites overfitting (overfit is this factory's #1 killer at 249x); and (b) position size is not a fixed 1.5% ATR-risk but a VOLATILITY-TARGET: notional is scaled inversely to recent realized volatility so the strategy takes larger exposure in calm, clean trends and automatically de-risks in chaotic high-vol regimes — the mechanism that historically stabilizes momentum's OOS Sharpe and tames drawdowns. Fills the most under-represented data-safe buckets: HYPERLIQUID venue (6.0% vs 20% quota) and long_short direction (13.7% vs target), on a fresh asset (LINK) not yet used on HL. Uses HYPERLIQUID 4H bars (~2.3 years usable history, above the sub-hourly wall), single instrument (no cross-venue schema-merge landmine, no spot-lookup timing bug), pure OHLCV, no supplementary feed (best-surviving tag family at 4%). Avoids every recent death mode: no options-expiry cadence, no funding fee-floor, no 90-day liquidation-feed wall, no COIN-M booking defect, and — by using the canonical robust signal rather than a fragile fitted one — directly targets the overfit failure that killed the LINK spot dip-buyer (PBO 0.89) and most bespoke signals.

This is a one-line-class-of-change fix to a self-inflicted deadlock, not a re-tune. The Layer-3 diagnostic said should_enter() returned a side 0 times over ZERO evaluated bars across 2,199 processed bars — zero *evaluated* bars means the base class never called the signal path at all, so no threshold could have been at fault. I traced it to a hard interaction in base_template: line 288 trims self._bars to _max_bars = 500, and line 309 gates the whole signal path on len(self._bars) >= min_bars_required(). Iteration 2's canonical-horizon change made min_bars_required() return max(360, 540, 30) + 2 = 542, which the 500-cap buffer can never reach — so calculate_signal() was never invoked, the indicator deques never filled, and the run was structurally incapable of trading. Iteration 1 returned 52 from that method, which is precisely why it traded 117 times; the zero-trade regression was introduced by the horizon fix, not by the entry logic. The fix is to decouple warmup from the tunable windows: min_bars_required() now returns a small locked constant (32), because this strategy's indicators never read self._bars — they own their own deques (_closes, _ma_win, _rets), which are unaffected by the 500-bar cap — and the genuine warmup gate already lives in should_enter(), which refuses to trade until the momentum, MA and realized-vol buffers are each individually full. That also makes the strategy optimizer-safe: the walk-forward optimizer can now raise ma_window past 500 without silently re-deadlocking the run, which would otherwise have produced phantom zero-trade windows during optimization. Nothing else changed. The signal, entry/exit logic, hysteresis, disaster backstop, cooldown, vol-targeted sizing, imports and all risk constants are byte-identical to iteration 2, so the Layer-1 and Layer-2 passes are preserved and the previously measured behaviour stands: with the 540-bar MA warm after ~540 of the sandbox's 2,199 bars, my vectorized sim of this exact configuration produced 15 trades in the 365-day window (+7.3% return, 12.2% max drawdown, PF 2.12) and 50 trades over the full 2.5 years (+19.9%, 18.3% drawdown, PF 1.68), so Layer 3's >=1-trade requirement is comfortably met. The thin-sample caveat from iteration 2 still stands and should drive the analyst's decision: ~50 trades over 2.5 years is a wide confidence interval, and the horizon was selected after observing a grid, so this must be judged on holdout and OOS results rather than in-sample ratios.

Statistically empty, outlier-carried single-perp momentum on a zero-survivor venue/mechanism class — not worth 2 hours of optimization. Only 46 trades over 2.6 years (below the ~100-trade measurability floor, L10), so the 3-window walk-forward and 20% holdout would each contain single-digit trades and any optimized Sharpe would be best-of-N noise. Sharpe 0.466 with bootstrap CI [-0.797, 1.649] straddles zero by a wide margin and PSR 0.76 (<0.95) — not distinguishable from no-skill even before the ~225-trial deflation. The 33.4% return is outlier-driven: 2025-01-31 alone (+22.9%) is ~two-thirds of it, with a +12.7% final bar and kurtosis 9.1 — strip a handful of days and the edge is gone. This is the Hyperliquid single-perp OHLCV momentum family (0/128, L82) on LINK, whose momentum-confluence already died 'no latent edge', and the developer honestly concedes the K=360/MA=540 horizon was selected as the only monotone-positive grid cell (selection bias baked in) and that it must not be promoted on in-sample ratios. Not fee-dead and DD contained, but no parameter tuning conjures significance from 46 outlier-driven trades in a zero-survivor class. Failure pattern: overfit/no_edge low-trade-count HL single-perp momentum, unmeasurable outlier-driven sample (L10/L82).

Long-short time-series momentum on the LINKUSD Hyperliquid perpetual, 4H bars, at the canonical (months-long) TSMOM horizon. Direction is the sign of the trailing 360-bar (60-day) return, confirmed for entry by price versus a 540-bar (90-day) moving average: long when both are positive, short when both are negative, flat on disagreement. Exits use hysteresis — only the trailing-return sign needs to flip — plus a locked 25% adverse-move disaster backstop, with a 12-hour calendar-anchored cooldown after each exit. Position size is volatility-targeted: notional equals equity times (15% annual vol target / current annualized realized vol), clipped to between 5% and 35% of equity. Leverage 1.0. Exactly three tunable parameters (momentum_window, ma_window, vol_target_annual); warmup, vol window, annualization factor, notional cap/floor, disaster stop and cooldown are locked constants.

CLEAN RESTART 2026-09-04 — this run's verdict history and learning records were removed and it was restarted from verification. Its previous abandonment came from the pipeline, not from the market: the Layer-2 harness mis-bound @staticmethod helpers (fixed), QA issued terminal performance verdicts on an unoptimized smoke test (removed — QA now judges correctness only), and sandbox timeouts came from backtest-slot starvation (fixed). The hypothesis and the strategy code are unchanged. Verify the code on its merits; performance is decided later by the full backtest and the optimizer.

Low frequency at the canonical horizon: 46 trades / ~2.6y, Sharpe CI straddles zero. Inherent to months-long TSMOM, not a bug; per-trade edge (4.2%) clears fees, but in-sample metrics can't stand alone and the holdout will have few trades. Weight OOS/holdout heavily; require holdout >0 Sharpe with >=10 trades on its own. If the optimizer shortens the horizon for more trades, re-check net per-trade return stays positive (developer's grid shows short horizons flip negative).

avg_trade_return_pct 4.14% is well above the ~0.15% Hyperliquid fee floor — not fee-fragile

max_drawdown 17.6% and PF 1.77 are contained; disaster stop and vol-targeting kept the iteration-1 blowup from recurring

Clean canonical construction (sign-of-trailing-return TSMOM), 3 parameters, no supplementary-feed or cross-venue landmines

46 trades over 2.6 years — below the ~100-trade measurability floor (L10); walk-forward folds and the 20% holdout would each hold single-digit trades, so any optimized Sharpe is best-of-N noise

Sharpe 0.466 with bootstrap CI [-0.797, 1.649] straddling zero widely and PSR 0.76 (<0.95) — not statistically distinguishable from no-skill before deflation

Outlier-driven: 2025-01-31 (+22.9%) is ~2/3 of the 33.4% total, plus a +12.7% final bar; kurtosis 9.1 — the edge is a few days, not a distribution

HL single-perp OHLCV momentum is a zero-survivor class (L82, hyperliquid 0/128) and LINK momentum already died 'no latent edge'; horizon was chosen after grid observation (developer admits the selection bias)

Verification failed (Layer 4 — QA review): - CATASTROPHIC DRAWDOWN / RISK BLOWUP past the >50% hard-abandon line. max_drawdown 63.4% (CI up to 99.86% — near-liquidation) with total_return -61.2% over 365 days. The hypothesis's central claim — that vol targeting TAMES momentum drawdowns — is refuted by its own backtest: with no stop-loss and inverse-vol sizing averaging 83.9% of equity notional, losing streaks in the 72% of trades that lose compound into a near-account-wipe. The vol estimate is backward-looking (de-risks only AFTER a spike) and the 0.40 target sizes far too aggressively for LINK. - MISLEADING RISK-ADJUSTED METRICS. Sharpe +1.48 / Sortino 2.53 / PSR 0.97 are internally CONTRADICTED by total_return -61.2%, PF 0.49, win_rate 0.28. The positive arithmetic Sharpe is an artifact of extreme positive skew (+1.59; one $14,690 win ~10x a normal position) while the GEOMETRIC outcome compounded to -61% via volatility drag from ~84% notional sizing. A promotion on these ratios would trust a strategy that lost 61% and nearly liquidated. Trust total_return / max_drawdown / PF instead. - NEGATIVE PER-TRADE EDGE. PF 0.49 (loses ~2:1), avg_trade_return_pct -0.636% (below fee floor), win_rate 0.28 over 117 trades — the sign-of-trailing-return TSMOM signal has negative expectancy on LINK/HL. Vol targeting only SCALES returns; it cannot flip a negative-edge signal positive. Lowering vol_target shrinks the loss/drawdown but still leaves a loser, so the drawdown fix and the edge problem are different levers. - The code is CORRECT and faithfully implements the hypothesis: causal trailing K-bar return signal, MA-confirmed long/short entry (signal>0 & close>MA -> BUY; mirror -> SELL; flat on disagreement), signal-flip exit (no stop, by design), inverse-vol sizing clipped [0.05,1.0], calendar-anchored cooldown, O(1) running-sum indicators, correct venue routing (HL futures, shorts). NO code bug — noted so the developer does not chase a phantom defect; the failure is risk/edge, not logic.

Verification failed (Layer 3 — sandbox backtest): No trades produced Bar type used: LINKUSD.HYPERLIQUID-4-HOUR-LAST-EXTERNAL, Bars processed: 2199 Diagnostics: should_enter() returned a side 0 times over 0 evaluated bars -> your ENTRY CONDITION never triggered. Loosen the entry logic / thresholds. Ensure your strategy produces trades with the given data and parameters.

Verification failed (Layer 4 — QA review): - STRUCTURALLY UNMEASURABLE SAMPLE (L16). The canonical 60-day/90-day TSMOM horizon produces only 14 trades in the sandbox and ~50 over the full 2.5-year history — below the ~100-trade floor needed to distinguish edge from noise. Sandbox Sharpe 0.29 with CI [-1.25, +1.92] straddles zero. Unlike the short-horizon HL momentum siblings (57-77 sandbox / ~100-150 full trades) where the full backtest CAN settle the edge, this long-horizon system cannot reach a validatable sample at any point, and walk-forward would have ~17 trades/window — too few for a reliable IS/OOS ratio. The full backtest cannot resolve this; per L16 reject at Layer 4 rather than burning the 3-phase optimization. - THE POSITIVE RESULT IS REGIME-SPECIFIC, NOT A ROBUST EDGE. The +5.96% is driven almost entirely by shorts in a LINK down-year: short_win_rate 0.80 (8/10) vs long_win_rate 0.25 (1/4). Realized reward:risk is UNFAVORABLE (avg_win $1664 < avg_loss $1805), so PF 1.66 rests entirely on a 64% win rate over 14 trades — unlikely to persist and would reverse in a LINK up-regime where the short-dominated book loses. - OVERFIT-BY-GRID-SELECTION. The developer honestly discloses the shipped horizon (K=360/MA=540) is the ONLY strongly-positive point in a grid where every shorter horizon is negative/marginal, chosen after observing that grid. The TSMOM-months prior mitigates but does not eliminate the selection bias; overfit is the factory's #1 killer and the sibling LINK spot dip-buyer died at PBO 0.89. A ~50-trade sample chosen at the one working grid point is a classic overfit setup walk-forward cannot reliably clear. - CREDIT WHERE DUE — the iteration-3 fixes are correct and genuine. The zero-trade deadlock is properly fixed (min_bars_required now a locked constant 32, decoupled from tunable windows; indicators own deques independent of the base 500-bar cap; warmup enforced in should_enter). The iter-1 risk blowup is genuinely fixed: avg_position_pct 26.9% (was 83.9%), max_drawdown 10.0% (was 63.4%) via vol_target 0.15, cap 0.35, 25% disaster stop. Code is correct and faithful with NO L17 defect signature. The block is NOT code or risk.