BtcHyperliquidPersistenceFilteredDualTimeframeMomentumLS

BTC Hyperliquid Persistence-Filtered Dual-Timeframe Momentum — Long-Short 4H Momentum That Has Held for K Bars, Aligned with the Daily Trend, ATR Trailing Exit (Volatility-Scaled, 3-Parameter)

A LONG-SHORT, single-instrument, pure-price directional MOMENTUM strategy on BTC-USD perpetual on HYPERLIQUID (BTCUSD.HYPERLIQUID), 4H primary + 1D trend confirm. It is deliberately built on the factory's ONLY validated combination — BTC (the single asset whose directional momentum has survived; every alt trend-follower dies as an outlier-mirage because one fat-tailed day dominates the return distribution — see the BTC-Donchian, XRP-swing, BNB-trend abandons) AND the dual-timeframe momentum-confluence mechanism that took the paper-stage BTC strategy to Sharpe 1.99. Two deliberate differences from that in-pipeline Binance strategy make this a distinct, additive experiment rather than a re-skin: (1) VENUE — it establishes the factory's FIRST Hyperliquid directional-trend book, filling the largest venue gap (HYPERLIQUID 6.6% vs 20% target) at a lower round-trip fee (~0.09% vs ~0.10%); (2) a PERSISTENCE FILTER that the survivor lacks — the 4H momentum must have held the SAME sign for K consecutive bars before entry (not a single fresh cross), which mechanically suppresses the one-bar-spike entries that produce the outlier-mirage, single-observation-dominated backtests the reviewers keep abandoning. Choosing BTC (the least fat-tailed major) plus a persistence filter directly targets the #1 killer of this family. NOT a mean-reversion fade (L53), NOT a squeeze breakout (L54), NOT a cointegration/basis pair (dead), NOT a cross-sectional rank basket (L52), NOT a non-price-feed gate (L46), NOT options/COIN-M (L50/L51). 4H+1D on Hyperliquid = ~2.3y usable 4H history + full daily; cadence stays moderate so walk-forward windows and the 15-day holdout populate. Risk profile: ~1.5% equity risk per trade via ATR stop; volatility-scaled sizing (notional = risk_budget / (atr_mult * ATR_4H)) with a hard per-trade notional cap so no single bar can swing equity >~25%; max single-position exposure ~25% of $100k; 2x leverage cap (reads self.config.leverage). Exactly 3 tunable parameters (daily-trend lookback, persistence count K, ATR trailing multiple); the fast/slow momentum windows are fixed at the survivor's proven values to avoid best-of-N overfit.

Implements the hypothesis exactly: the validated BTC + dual-timeframe momentum mechanism, moved to Hyperliquid (first HL directional-trend book, ~0.09% round trip) and hardened with the persistence filter the survivor lacks. Fast/slow windows are FIXED at 12/30 exactly as the hypothesis requires, so only 3 knobs (daily_lookback, persistence_k, trail_atr_mult) are exposed; their code clamps are mirrored verbatim in _param_bounds so the optimizer cannot test values the code silently rewrites. Long-short fills the under-weight direction bucket; leverage 2.0 is genuinely consumed (notional cap = equity * max_notional_frac * self.config.leverage), and because the cap is a fixed fraction of CURRENT equity, losses cannot compound into larger positions (L15). The daily leg is real 1-DAY data via on_extra_bar, and extra_instruments carries BTCUSD.HYPERLIQUID because InstrumentResolver.resolve_extras pairs bar types POSITIONALLY -- an empty list silently drops the daily feed and yields a zero-trade sandbox (the failure that cost a prior port two iterations); as a slower-timeframe context leg it never trips the base class's same-timeframe sync barrier. VERIFIED BEFORE SUBMITTING by replaying THIS EXACT CLASS over the real catalog data (5567 4H bars from 2024-01-22 + 2179 daily bars): 175 round trips (~69/yr -- walk-forward windows populate and a 15-day holdout is not structurally empty), avg per-trade return +0.91% net of Hyperliquid's ~0.09% round-trip taker cost (well above the 0.15% viability floor), 48% win rate, profit factor 1.86, trade-return kurtosis 4.4 and the top-3 trades contributing only 16.7% of gross profit -- broad-based, not the one-observation mirage this family usually dies of. Positive in every calendar year (2024 +1.46%, 2025 +0.58%, 2026 YTD +0.70% avg/trade) and over the trailing 365 days (+0.70%/trade over 72 trades), so it is not carried by a single regime. The persistence filter does what the hypothesis claims, measurably but modestly: K=2 -> kurtosis 6.3 / PF 1.73, K=4 -> 4.4 / 1.86, K=6 -> 4.8 / 2.04, i.e. requiring persistence trades slightly fewer, fatter, less tail-dependent trades. One-at-a-time sensitivity across all declared bounds stays positive (avg per trade +0.61% to +1.16%, PF 1.58-2.18), so the defaults sit on a plateau (grid centre), not a fitted spike. Per-bar work is O(1) (incremental EMAs, bounded TR deque), so the 300s Layer-3 cap is not at risk; all six Layer-2 synthetic scenarios pass locally with a continuous, varying signal.

Failed two hard gates with a decisive overfit signature. The holdout FAILED catastrophically — holdout Sharpe -1.15, ratio -1.126 (needs >=0.70) — meaning the untouched final 15-day window lost money, a genuine out-of-sample failure. PBO is 0.5741 (>0.5), so the CSCV test says the selection is more likely than not overfit, and deflated_sharpe is 0.2912 (is_significant false, sharpe_ci_low 0.096); per the robustness rubric, high PBO plus low DSR is a decisive abandon. The optimized config's recent regime is negative (2026 -2.6%, rolling_sharpe negative Apr-Aug 2026), consistent with the short 2.3-year single-regime HL data (from 2024) being unable to support significance — exactly the risk flagged at pre-optimization. The passing signals (walk-forward not flagged overfit, all OOS windows positive, clean sensitivity) are overridden by the negative holdout and PBO>0.5: surviving the coarse walk-forward split while losing the untouched holdout is the signature of a selection fit to the in-sample period. This is NOT the ETH-Binance-style iterate case (that PASSED holdout 0.761 and PBO 0.32 and only over-leveraged its selection); here the holdout is negative and PBO>0.5, so the failure is the mechanism/data, not the optimizer's config choice, and re-optimizing a 2.3-year single-regime series cannot manufacture a positive holdout or clear DSR. Corroborated by the failure store: a near-identical BTC-HL dual-TF confluence (similarity 0.816) was already abandoned (fee_edge), alongside four more dual-TF siblings. Abandon.

Long-short, single-leg dual-timeframe momentum strategy on BTCUSD.HYPERLIQUID perpetual, 4H primary bars with a 1-DAY trend confirm and a PERSISTENCE filter. Fast leg = (EMA12 - EMA30)/ATR14 on 4H bars -- fixed at the survivor's windows, not tunable -- returned every bar as a continuous, volatility-scaled read; the strategy also tracks how many consecutive bars that read has kept its sign. Slow leg = sign of the 20-day daily return, read from BTCUSD.HYPERLIQUID-1-DAY bars in on_extra_bar. ENTRY only when the 4H sign equals the daily direction AND the sign has persisted for at least 4 consecutive 4H bars, so a one-bar momentum spike can never open a position; otherwise FLAT. EXIT on a 3x-ATR trailing stop from the best close since entry, on the daily trend flipping, or on the 4H read flipping. Sizing risks ~1.5% of equity at the trailing-stop distance (qty = equity*0.015/(3*ATR)) with a hard per-trade notional cap of 25% of equity x 2x leverage.

Pure-OHLCV trend book (0/213 class) — but the best-positioned instance: BTC (survivor asset, least fat-tailed) + persistence filter targeting outlier-mirage death. Strong sandbox (+17.71%, Sharpe 0.920, PF 1.55). kurtosis 6.97 is higher than the dry-run's 4.4, so confirm outlier-reduction on walk-forward OOS. Unlike LINK/AVAX-HL, NO capacity ceiling.

~21-day daily-leg warmup gates the first entry; earliest OOS window could be trade-sparse. Not a zero-trade bug (66 trades).

On BTC (survivor asset) + survivor mechanism (dual-TF momentum confluence), with a low-DOF persistence filter (3 params, momentum windows FIXED) that demonstrably suppressed the outlier mirage: return_kurtosis 6.54, distributed edge

Sharpe 1.07 with CI-low 0.07 > 0 (positive, PSR 0.9825) — 2x the alt-HL siblings and CI lower bound clear of zero (vs their -0.6)

Deployable liquidity: capacity ~$56M, impact_cost_pct 4.2% (vs $6M / 12.8% for LINK/AVAX-HL)

Consistent positive years including recent: 2024 +22.8%, 2025 +16.8%, 2026 +12.5% YTD; decorrelated alpha 0.14, PF 1.66, max_drawdown 9.1%, 177 trades

Marginal significance: CI-low 0.07 is thin; best-of-N deflation over ~225 trials is the real test

Short single-regime data: HL 4H starts 2024-01 (~2.5 years), thin walk-forward/holdout depth

HL directional trend is the 0/97 class (L41), and BTC directional siblings (vol-regime, regime-adaptive) failed the 3-phase run this session — though those were higher-DOF classifiers, not this fixed-window design

Walk-forward not flagged overfit (IS 2.43 / OOS avg 1.02, all three OOS windows positive); sensitivity clean (0 cliffs); distributed edge at pre-opt (kurtosis 6.5)

On BTC (survivor asset) with a low-DOF persistence filter — a thoughtful anti-outlier design

Holdout FAILED catastrophically: holdout Sharpe -1.15, ratio -1.126 — the untouched final window lost money (hard gate)

PBO 0.5741 > 0.5 (hard gate): parameter selection more likely than not overfit; combined with deflated_sharpe 0.2912 this is the decisive-abandon combination

is_significant false, sharpe_ci_low 0.096 (barely off zero); optimized config's recent regime negative (2026 -2.6%), rolling_sharpe negative Apr-Aug 2026

Short 2.3-year single-regime HL data (from 2024) — insufficient to support significance, as flagged pre-opt

Family prior: near-identical BTC-HL dual-TF (sim 0.816) already abandoned (fee_edge), plus 4 more dual-TF siblings

BtcHyperliquidPersistenceFilteredDualTimeframeMomentumLS combined the factory's two validated ingredients — BTC and dual-timeframe momentum confluence — with a persistence filter requiring K consecutive same-sign bars before entry, directly targeting the outlier-mirage that had killed prior trend followers. It worked at first pass: a distributed +67.2% with Sharpe 1.07 (CI-low above zero), PF 1.66, a 9.1% drawdown, and ~$56M capacity earned it the batch's only 'optimize' verdict and a full optimization run. But that run was decisive against it — the untouched holdout lost money (Sharpe -1.15), PBO was 0.574, and deflated Sharpe just 0.29, all consistent with only ~2.3 years of single-regime Hyperliquid data being too thin to establish the edge. The analyst abandoned it at the ANALYZING stage on the negative holdout and PBO gates, corroborated by a near-identical BTC-HL sibling already abandoned; it reached optimization and analysis but not risk review or promotion.

Passing the coarse walk-forward split with positive OOS windows while losing the untouched holdout, alongside PBO > 0.5 and a deflated Sharpe near zero, is the signature of a selection fit to the in-sample period — a thoughtful anti-outlier persistence filter and a >1 headline Sharpe cannot overcome a ~2.3-year single-regime dataset too short to establish significance.

It uniquely cleared backtest review with an 'optimize' verdict and ran the full 3-phase optimization, but the analyst abandoned it at the ANALYZING stage on two hard gates: the 15-day holdout lost money (ratio -1.126 vs required ≥0.70) and PBO 0.574 > 0.5 with deflated Sharpe 0.29 — a decisive overfit signature. The short ~2.3-year single-regime Hyperliquid history (from 2024) could not support significance, and re-optimizing it cannot manufacture a positive holdout.

A long-short, single-instrument dual-timeframe momentum strategy on BTCUSD.HYPERLIQUID (4H primary + 1D trend confirm, 3 parameters) that took a position only when the 4H EMA-spread momentum had held the same sign for K consecutive bars AND agreed with the daily trend — a persistence filter added to the survivor mechanism specifically to suppress the one-bar-spike entries that produce outlier-mirage backtests — with an ATR trailing exit.

The initial backtest (BTCUSD.HYPERLIQUID 4H, 927 data days from 2024-01) was the strongest Hyperliquid directional candidate yet: +67.2% over 177 trades, Sharpe 1.07 with CI-low 0.07 above zero, profit factor 1.66, max drawdown 9.1%, distributed edge (kurtosis 6.54), ~$56M capacity, and positive every year. But optimization exposed overfitting — the untouched holdout failed catastrophically (Sharpe -1.15, ratio -1.126), PBO 0.574, deflated Sharpe 0.29 (not significant), and the optimized config's 2026 turned negative.