Skip to content
All strategies

Strategies

Hyperliquid SOL Volume-Confirmed Dual-Timeframe Momentum, Long-Short (SOLUSD.HYPERLIQUID Perp, 4H Primary + 1D Trend Confirm, Trade Directional Momentum ONLY When a Volume Surge Confirms Participation, Flat Otherwise, ATR-Trailing Exit, 3-Parameter)

Outcome: Abandoned

SolHyperliquidVolumeConfirmedDualTimeframeMomentumLS

Outcome Summary

This strategy deliberately ported the factory's only empirically-surviving mechanism — volume-confirmed dual-timeframe momentum — onto the under-represented Hyperliquid venue as a single-instrument long/short SOL trade, aiming for genuine venue novelty rather than a Binance re-skin. It cleared the pre-optimization gate (backtest-review verdict 'optimize') with a modest but viable profile: Sharpe 0.516, profit factor 1.235, per-trade edge above the venue cost floor, and a low 17.2% drawdown. Optimization barely moved the headline (Sharpe 0.559) and, crucially, the robustness statistics collapsed: PBO 0.7474 breached the hard 0.5 gate, the deflated Sharpe (0.234) sat below the best-of-N luck bar, and the candidate failed programme-level FDR while losing money in the stressed volatility regime the surge filter was meant to exploit. After one iteration the analyst abandoned it as an overfit, statistically-insignificant Hyperliquid dual-TF momentum port — consistent with the negative class prior of every prior HL dual-TF attempt.

Strategy report

Hypothesis

A LONG-SHORT, SINGLE-INSTRUMENT, pure-OHLCV momentum strategy on SOLUSD.HYPERLIQUID perpetual (4H primary bars, 1D trend confirm). This deliberately ports the factory's ONLY empirically-surviving mechanism family - volume-confirmed dual-timeframe momentum confluence, which reached paper stage as EthVolumeConfirmedMomentumLS (Sharpe 1.05), BtcDualTimeframeMomentumConfluenceLS (1.99) and AdaBinanceDualTimeframeMomentumConfluenceLS (1.60) - onto a genuinely under-represented VENUE (Hyperliquid, 7.2% of the corpus vs……Show moreShow less

A LONG-SHORT, SINGLE-INSTRUMENT, pure-OHLCV momentum strategy on SOLUSD.HYPERLIQUID perpetual (4H primary bars, 1D trend confirm). This deliberately ports the factory's ONLY empirically-surviving mechanism family - volume-confirmed dual-timeframe momentum confluence, which reached paper stage as EthVolumeConfirmedMomentumLS (Sharpe 1.05), BtcDualTimeframeMomentumConfluenceLS (1.99) and AdaBinanceDualTimeframeMomentumConfluenceLS (1.60) - onto a genuinely under-represented VENUE (Hyperliquid, 7.2% of the corpus vs the >=20% quota target; Binance USD-M is saturated at 70.9%) rather than re-skinning it on yet another Binance symbol. The novelty is real and mechanism-relevant, not cosmetic: Hyperliquid's participant base (on-chain-native, vault-driven, no US retail) produces a different momentum-persistence and mean-reversion signature than Binance USD-M, so a mechanism that overfit on Binance-ETH-HL confluence does NOT trivially reproduce here (this is single-instrument volume-confirmed, not the plain dual-TF confluence that overfit on ETH-HL). It is single-instrument (HL baskets crash at backtest_failed per L88 - explicitly avoided). It is NOT any dead class: not funding/OI/liquidation/tick-gated (L70/71/80/83/87), not a cross-sectional/TSMOM basket (L68/77), not intraday sub-hourly reversion (L89), not a premium/basis fade (L85), not COIN-M (L78), not dated-futures (L84), not an unbacktestable options/prediction feed (L76), not a US-session seasonality trade (L79).

Implements the hypothesis exactly: volume-confirmed dual-timeframe directional momentum on SOLUSD.HYPERLIQUID perp, 4H primary with a daily-scale trend confirm, flat when unconfirmed, ATR-trailing exit, three tunable parameters (mom_lookback, vol_surge_mult, trail_atr_mult) with declared _param_bounds matching the code's clamps. It ports the construction of the surviving siblings (EthVolumeConfirmedMomentumLS, BtcDualTimeframeMomentumConfluenceLS4H) onto the under-represented Hyperliquid venue, single-instrument……Show moreShow less

Implements the hypothesis exactly: volume-confirmed dual-timeframe directional momentum on SOLUSD.HYPERLIQUID perp, 4H primary with a daily-scale trend confirm, flat when unconfirmed, ATR-trailing exit, three tunable parameters (mom_lookback, vol_surge_mult, trail_atr_mult) with declared _param_bounds matching the code's clamps. It ports the construction of the surviving siblings (EthVolumeConfirmedMomentumLS, BtcDualTimeframeMomentumConfluenceLS4H) onto the under-represented Hyperliquid venue, single-instrument (no HL basket), pure OHLCV (no funding/OI/liquidation/tick feed, so no missing-supplementary path can change the edge). The daily-scale confirm is derived from the SAME 4H series with a 42-bar lookback rather than subscribing a separate 1-DAY bar type -- identical information, but it can never be one bar stale and it removes the extra-leg alignment barrier that has produced zero-evaluated-bar sandbox runs; this is the construction the surviving BTC 4H sibling uses. Venue is HYPERLIQUID futures (MARGIN) because the strategy takes shorts, which a CASH account cannot; leverage stays 1.0 and no sizing path reads it, so nothing is set-but-unused. Frequency was checked offline against the actual catalog series (5,728 4H bars from 2024-01): the FULL compound gate (surge >= 1.6x median AND sign agreement AND |signal| >= 1 ATR AND a fresh cross) fires 515 times, so the zero-trade sandbox failure mode is not in play, and the ~0.09% Hyperliquid round-trip cost is 20-60x smaller than a typical confirmed 4H SOL leg. Per-bar work is bounded (index-addressed windows, one 60-element sort, O(1) trailing extreme, memoised previous-bar eligibility) so the 300s smoke limit is comfortable, and there are no process-relative counters or schedules anywhere.

failed deflated Sharpe: DSR=0.234, expected-max=1.001 over 225 trials (selected Sharpe 0.559 is below the best-of-N luck bar) — indistinguishable from best-of-N noise. Compounded by a HARD, unwaivable gate failure (PBO 0.7474 > 0.5) and failure of programme-level FDR (keeps 3 of 237, candidate_survives=false), with is_significant=false and a Sharpe CI [-0.706, 1.717] straddling zero. This is NOT a fee-edge death (per-trade 0.303% clears the floor); it is an overfit / no-generalization + statistical-insignificance failure. OOS Sharpe 0.821 with a negative window and a -0.72 stressed-regime Sharpe confirm the volume-surge gate carries no robust edge on this venue. The same volume-confirmed dual-TF mechanism survived on Binance majors (ETH/BTC/ADA), but every Hyperliquid dual-TF momentum port has died (0/N: SOL/ETH HL confluence, SOL HL Donchian, SOL HL vol-gate, SOL HL TS-momentum) — re-pointing the same recipe at HL simply overfits again, so abandon rather than revise. Failure pattern: overfit / not-significant (Hyperliquid dual-TF momentum port).

Implementation

Long/short single-instrument momentum on the SOLUSD.HYPERLIQUID perpetual, 4H bars. Every bar it computes a continuous, ATR-normalised dual-timeframe confluence score -- 0.5*(fast 12-bar momentum + 42-bar (7-day, daily-scale) trend / 2.0) -- multiplied by a participation weight equal to the bar's volume over the median volume of the prior 60 bars, divided by the surge multiple and capped at 1. It goes long when that signal crosses up through +1.0 ATR and short when it crosses down through -1.0 ATR, but ONLY when the volume surge is genuinely met (participation >= 1) and the daily-scale trend has the same sign as the 4H momentum; otherwise it stands flat. Exits are structural: momentum reaching 0.25 ATR against the position (with hysteresis so an intra-leg pause does not truncate a winner and bill a second round trip), or a 3-ATR trailing stop from the best close since entry. Sizing is risk-first -- equity * 2% / (trail_atr_mult * ATR), capped at 1x equity notional -- so position size shrinks as volatility expands and no leverage is used.

Verification Results

No change required. The 4H-derived daily-scale confirm is a valid and intentional construction that avoids the extra-leg alignment/zero-bar sandbox failure mode; keep as-is.

The hypothesis title says '1D Trend Confirm', but the daily-scale trend leg is derived from a 42-bar (7-day) displacement on the SAME 4H series rather than subscribing a separate 1-DAY bar type. This is documented as deliberate (matches the surviving BTC 4H sibling) and is information-equivalent daily-scale trend, so the strategy is genuinely dual-timeframe and the core mechanism is faithfully implemented. Flagged as informational only — not a mismatch that changes the edge.

No fix needed — guards are present and correct.

Static analysis flags several division-by-variable anti-patterns (atr, stop_dist, price, step, base_vol). On review each is guarded: _atr returns None/positive-only, base_vol None/<=0 falls to participation=0, stop_dist/price/step all have finite-positive checks before use. No actual divide-by-zero path exists.

Regime attribution shows the edge is concentrated in calm/normal vol (Sharpe 1.25/1.01) and NEGATIVE in the stressed tercile (Sharpe -0.69, -15.6%, DD 29.9%) — the exact high-participation regime the volume-surge premise targets should ideally hold up, so the analyst should confirm the mechanism is not simply long crypto-beta in quiet periods. Sandbox Sharpe 0.51 with CI [-0.72, 1.69] straddles zero on 218 trades; not blocking, but the risk-adjusted edge is unproven pre-optimization.

Backtest Review

Sharpe
0.52
Total return
41.20%
Max drawdown
17.19%
Trades
218
Win rate
35.8%
Profit factor
1.24

Per-trade edge clears the venue cost floor: avg_trade_return_pct 0.238% > 0.15% Hyperliquid floor, commissions only 4.04% of gross and impact ~13% — this is not a fee-fragile / edge-below-cost case.

Reliable, adequate sample: 218 trades over 955 days (metrics_reliable=true), balanced 103 long / 115 short — the trades implement the stated long-short volume-confirmed dual-TF momentum mechanism, not a mismatch.

Profit factor 1.235, positive expectancy ($151/trade), max_drawdown a modest 17.2% (recovery_factor 1.92), and near-zero beta (-0.002) with positive alpha (0.149) — the return is not just crypto-beta.

This is the one empirically-surviving mechanism family (volume-confirmed dual-TF momentum, L82) as a single-instrument variant (deliberately NOT the HL confluence basket that died), on an under-represented venue — a legitimate candidate for the 3-phase run.

Risk-adjusted edge is unproven pre-optimization: Sharpe 0.516 with CI [-0.74, 1.69] straddling zero, PSR 0.768.

Regime attribution is a real caution: profitable in calm/normal terciles (Sharpe 1.25/1.02) but NEGATIVE in the stressed tercile (Sharpe -0.70, -15.6%, DD 29.9%) — optimization/walk-forward must confirm the edge is not just quiet-period momentum.

Losing calendar year in 2025 (-1.6%); performance is uneven across years.

Analysis

Per-trade edge clears the venue cost floor: avg_trade_return_pct 0.303% > 0.15% Hyperliquid floor; commissions only 3.54% of gross, so this is NOT a fee-edge/edge-below-cost death.

Reliable, adequate sample (190 trades, 955 days, metrics_reliable=true) and holdout technically passes the hard gate (holdout Sharpe 0.666, 25 trades, consistent_with_oos=true).

Walk-forward is_overfitted=false and sensitivity grid shows no destabilising cliffs in the tunables (mom_lookback/vol_surge/trail all smooth).

HARD gate failed and cannot be waived: PBO 0.7474 > 0.5 — the parameter selection is more likely than not overfitting. A promote is impossible.

Not statistically significant after multiple-testing correction: deflated_sharpe 0.234 << 0.95, is_significant=false, Sharpe CI [-0.706, 1.717] straddles zero, and it does NOT survive programme-level FDR (Benjamini-Hochberg q=0.10 keeps 3 of 237; candidate_p 0.766). The selected Sharpe (0.559) sits BELOW the expected-max-by-luck bar (1.001) over 225 trials — it is indistinguishable from best-of-N noise.

Weak generalisation: avg OOS Sharpe 0.821 with one window negative (-0.068), far below any promotion bar; optimization barely moved the headline (Sharpe 0.516 -> 0.559).

Loses money in the high-volatility third (stressed regime Sharpe -0.72, -14.3%) — the exact participation-driven regime the volume gate is meant to exploit — so the surge filter adds no robust edge.

Class prior is decisively negative: Hyperliquid dual-timeframe momentum ports keep dying (SOL HL confluence risk_reject, ETH HL confluence overfit, SOL HL Donchian overfit, SOL HL vol-gate fee_edge, SOL HL TS-momentum overfit). The venue novelty did not add robustness the deflation could not strip.

Benjamini-Hochberg at q=0.10 over 237 programme candidates keeps 3. A candidate that does not survive here is not distinguishable from the programme's own noise, however good its individual statistics look.

Outcome Summary

Re-pointing the volume-confirmed dual-timeframe momentum recipe (which survived on Binance ETH/BTC/ADA) at Hyperliquid keeps overfitting — every HL dual-TF momentum port has now died — so venue novelty alone does not add edge that deflation and PBO cannot strip.

The analyst abandoned it at ANALYZING: a hard, unwaivable PBO of 0.7474 (> 0.5), deflated Sharpe 0.234 below the best-of-N luck bar of 1.001 over 225 trials, failure of programme-level FDR (candidate_survives=false), is_significant=false and a Sharpe CI [-0.706, 1.717] straddling zero — an overfit / no-generalization failure, not a fee-edge death.

Ported the factory's one empirically-surviving mechanism family — volume-confirmed dual-timeframe momentum confluence — as a single-instrument long/short strategy on the SOLUSD.HYPERLIQUID perpetual (4H bars with a 4H-derived daily-scale trend confirm), trading directional momentum only when a volume surge confirmed participation and standing flat otherwise, with an ATR trailing exit.

The initial backtest ran 218 trades over 955 days with Sharpe 0.516, profit factor 1.235, 35.8% win rate, 17.2% max drawdown and avg per-trade return 0.238% (above the 0.15% Hyperliquid floor); after 3-phase optimization it produced Sharpe 0.559, profit factor 1.219, 190 trades and avg per-trade return 0.303%, and the holdout technically passed (holdout Sharpe 0.666, 25 trades).

Backtest and paper results are hypothetical. Trading involves risk of loss.