HlEthVolRegimeGatedMomentumContinuationLS1H

Hyperliquid ETH 1H Volatility-Regime-Gated Momentum Continuation Long+Short

A long+short directional momentum-continuation strategy on ETHUSD perpetual at Hyperliquid using 1H bars. The strategy is OHLCV-only (no liquidation/funding/cross-venue dependencies — avoiding the 13+ infrastructure-blocked abandons this session). Core mechanism: gated z-score momentum. (1) A volatility-regime gate computes 30-day realized vol of 1H log returns and only allows entries when current realized vol sits in the middle 60% of its 30-day distribution — this kills both dead-quiet whipsaw ranges and extreme-vol blowoffs that cause IS/OOS divergence. (2) Within the allowed regime, compute the 4-bar (4H) log return; enter LONG when it exceeds +1.2σ of the trailing 48-bar return distribution; enter SHORT when it falls below -1.2σ. (3) Position sizing is volatility-targeted: notional = (target_vol / realized_vol) × equity × 0.30, capped at 0.50 equity, leverage ≤ 3x. Hyperliquid is chosen because (a) its venue is at 4.4% vs 20% quota target, (b) HL order flow has a different participant mix than Binance — heavier retail/momentum-chaser presence, fewer high-frequency market-makers — which empirically extends the half-life of intraday momentum past what Binance allows, (c) HL fees are 0.09% RT, friendlier than Binance perp (0.10% RT). The 1H timeframe produces ~150-300 entries/year across 3+ years of HL ETH history — well above the walk-forward floor (5+ trades/OOS window) that killed 13 daily strategies this session.

Iteration 2 fixes the Layer-3 'No trades / 0 evaluated bars' failure with a single targeted change. Root cause: the base class caps self._bars at 500, but min_bars_required() returned regime_lookback+vol_window+5 = 749. The warmup gate len(self._bars) < min_bars_required() was therefore permanently true (500 < 749), so _process_primary never ran and calculate_signal/should_enter were never reached -> exactly the diagnostic's '0 evaluated bars'. The fix lowers min_bars_required() to mom_period+mom_lookback+5 = 57 (< 500), so processing starts as soon as the first momentum z-score is computable. The long ~30-day regime history still accumulates in the strategy's private _closes buffer (cap 804, unaffected by the base 500-cap), and _compute()'s existing guard (dist.size < max(30, regime_lookback//4)) holds entries off until enough realized-vol samples exist -- so the regime gate's integrity is fully preserved and no signal logic changed. All entry/exit/sizing code and imports are byte-identical, so Layers 1-2 stay green; only the warmup-threshold collision with the base buffer cap was corrected.

Do not optimize — two independent, decisive failures. (1) DATA INSUFFICIENT: the hypothesis is premised on '3+ years of HL ETH history... well above the walk-forward floor,' but the backtest spans only 147 data_days (~7 months, 2025-10-11 to 2026-05-20, 5,317 1H bars). A 3-window walk-forward over 7 months yields tiny, unreliable OOS windows — the very floor the hypothesis claimed to clear does not exist for HL ETH 1H. (Symptom stated, not root-caused: whether HL history is genuinely this short or under-fetched, only ~7 months is available to validate against.) (2) NO EDGE AFTER COSTS + CAPACITY WALL: over the available window PF is 1.045, per-trade expectancy $13.85, Sharpe -0.22 with sharpe_ci_low -2.81 (CI deeply straddles 0) and omega 0.96; critically impact_cost_pct is 66.4% and commission is 12.2% of gross — trading costs consume nearly all gross PnL — while capacity_usd is only $226,937. A 1H long/short strategy with 288x turnover is far too heavy for HL ETH liquidity and is untradeable beyond toy size regardless of how parameters are tuned. This is not a sparsity or parameter problem fixable by optimization: the validation data is absent and the cost/capacity structure destroys the edge. If the momentum-continuation idea is to be pursued, it needs a venue/instrument with both sufficient history and the liquidity to absorb the turnover (a premise change), not optimization of the HL ETH 1H version — abandon this arm.

ETHUSD Hyperliquid 1H long+short volatility-regime-gated momentum continuation, OHLCV-only. A 30-day realized-vol percentile gate allows entries only when current 1H realized vol is in the middle 60% of its trailing distribution; within that regime it enters LONG/SHORT when the 4-bar log return z-scores beyond +/-entry_z against the trailing 48-bar return distribution. Volatility-targeted sizing (target_vol/realized_vol x base_frac, capped at 0.50 equity). Exits on ATR stop, momentum reversal past exit_z, or a 72-bar time stop.

Verification failed (Layer 3 — sandbox backtest): No trades produced Bar type used: ETHUSD.HYPERLIQUID-1-HOUR-LAST-EXTERNAL, Bars processed: 5317 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.

Healthy in-window trade count (195 trades over 7 months) and OHLCV-only design avoids the infrastructure-blocked failure mode

Mechanism executes as intended (long+short, vol-regime gate, vol-targeted sizing); low net beta (-0.01)

Data window is ~7 months (147 data_days, 2025-10-11 to 2026-05-20), not the 3+ years the hypothesis requires — a 3-window walk-forward would have tiny, unreliable OOS windows

No edge after costs: PF 1.045, expectancy $13.85/trade, Sharpe -0.22, sharpe_ci_low -2.81 (CI deeply straddles 0), omega 0.96

Hard capacity wall: impact_cost_pct 66.4% + commission 12.2% of gross consume nearly all PnL; capacity_usd only $226,937 — 288x turnover is far too heavy for HL ETH 1H liquidity

Untradeable at scale regardless of optimization; the hypothesis's own premise (long HL history + friendly costs) is contradicted by the data

HlEthVolRegimeGatedMomentumContinuationLS1H aimed to exploit a longer intraday-momentum half-life on Hyperliquid's retail-heavy ETH order flow with a vol-regime-gated, vol-targeted long/short design that cleanly avoided the session's infrastructure-blocked failure mode. The mechanism executed as intended — 195 long and short trades with near-zero net beta — but the backtest spanned only ~7 months rather than the claimed 3+ years, and over that window it produced no edge after costs (PF 1.045, Sharpe -0.22 with a CI low of -2.81). Worse, market impact (66.4%) and commission (12.2% of gross) consumed nearly all PnL against a ~$227k capacity wall, contradicting the hypothesis's own friendly-cost premise. With both the validation data absent and the cost/capacity structure destroying the edge, the analyst abandoned the arm on its second iteration without running optimization.

A venue chosen for friendly fees and supposed long history must actually deliver both sufficient data and enough liquidity to absorb the turnover — a high-turnover 1H long/short strategy on a thin venue is untradeable beyond toy size and unvalidatable when only months of history exist, neither of which optimization can fix.

The analyst abandoned it at the backtest-review gate before optimization on two independent, untunable failures: insufficient validation data (only ~7 months available versus the 3+ years the hypothesis premised, leaving any walk-forward OOS windows tiny and unreliable) and no edge after costs plus a hard capacity wall (66.4% impact plus 12.2% commission consuming nearly all gross PnL on 288x turnover too heavy for HL ETH 1H liquidity).

A long+short, OHLCV-only directional momentum-continuation strategy on Hyperliquid ETHUSD perpetual 1H bars — gating entries to a middle-60% realized-volatility regime, then entering long/short on z-scored 4-bar momentum past ±1.2σ with volatility-targeted sizing, ATR stops, momentum-reversal exits, and a time stop.

Over the available ~7-month window (147 data_days) the strategy took 195 trades (93 long, 102 short) for just +2.82% total return with profit factor 1.045, per-trade expectancy $13.85, a 43% win rate, max drawdown 17.4%, and Sharpe -0.22 (CI low -2.81). Trading costs were crippling — impact_cost_pct 66.4% and commission 12.2% of gross — against a capacity of only $226,937.