SolHyperliquidDualTimeframeMomentumConfluenceLS

SOL Hyperliquid Dual-Timeframe Momentum Confluence — Long-Short, Single-Instrument: 4H Trend Aligned With the Daily Trend, Flat on Disagreement, ATR-Trailed Winners (HYPERLIQUID SOLUSD Perp, 4H Primary + 1D Confirm, 3-Parameter)

A LONG-SHORT, single-instrument, pure-OHLCV momentum-confluence trend follower ported onto the UNDER-REPRESENTED Hyperliquid venue (7.1% of portfolio vs ≥20% target). This is NOT a new mechanism gamble — it is a deliberate venue-diversification move of the ONLY template that has actually survived optimization in this factory recently: the 4H+1D dual-timeframe momentum confluence family (BtcDualTimeframeMomentumConfluenceLS reached paper_stage at Sharpe 1.99; AdaBinance variant reached risk_review at Sharpe 1.60). Both of those are single-symbol, long-short, ATR-trailed, 3-parameter. I am replicating that exact surviving structure on SOLUSD.HYPERLIQUID — a different venue, different symbol, different funding/liquidity microstructure — to (a) fill the Hyperliquid quota gap, (b) add a long-short book (14.4% vs ≤55% long-only cap means long-short is starved of good ideas), and (c) test whether the confluence edge is venue-robust. TIMEFRAME CHOICE IS DELIBERATE: 4H primary means ~2.3 years of usable Hyperliquid candle history (per the HL history table), and the 1D confirm has full history back to 2020 — this AVOIDS the Hyperliquid sub-hourly history trap that abandons weeks-only backtests. Signal is 100% price-derived on bars we already have; no second leg, no supplementary feed, no funding gate (L70), no OI gate (L71), no cross-venue leg injection (L69), no rotation basket (L68). Only 3 tunable parameters to resist the overfit graveyard (338 deaths).

This is a deliberate venue port, not a new mechanism, exactly as the hypothesis frames it. The signal construction, exit stack, risk-first sizing and 3-parameter surface are reproduced from the surviving 4H+1D confluence family (BtcDualTimeframeMomentumConfluenceLS -> paper; AdaBinanceDualTimeframeMomentumConfluenceLS -> risk_review) rather than re-derived for SOL, because re-fitting the rule per venue is precisely the overfit that killed the earlier Hyperliquid attempts on PBO/holdout. Concretely: (1) 4H primary + genuine 1-DAY confirm feed -- extra_instruments repeats SOLUSD.HYPERLIQUID alongside the 1-DAY bar type, because the resolver pairs extra bar types positionally with extra_instruments and an empty list would silently drop the daily leg (the base's cross-leg alignment barrier does not apply here since 1-DAY is a slower timeframe than the 4H primary, so the primary never stalls). (2) Timeframe choice honours the HL history constraint -- 4H has ~2.3 years of usable candles and 1-DAY reaches further back, avoiding the sub-hourly HL history trap. (3) Venue is HYPERLIQUID futures (MARGIN) because the strategy goes short; its ~0.09% round-trip is the cheapest futures venue, and a 4H-cadence trend follower holding days clears that comfortably. (4) No supplementary feeds at all -- no funding gate, no OI gate, no liquidation feed, no cross-venue leg injection, no rotation basket, so none of the injection/zero-signal failure modes apply. (5) Entry frequency: the confluence fires whenever the 4H spread and the daily direction share a sign, which is a large fraction of bars, so the sandbox produces trades in the hundreds rather than risking a zero-trade abandon. (6) Drawdown control: sizing risks ~1.2% of equity across the trailing-stop distance and gross notional is hard-capped at 35% of equity, so losses cannot compound into position size. leverage stays 1.0 and is still read in position_size, so the notional cap is exact rather than inert. (7) Per-bar cost is O(1): incremental EMAs, a bounded TR deque and a bounded daily-close deque -- nothing rescans history, so the Layer-3 wall-clock budget is safe.

hl_short_history_significance_trap_confluence_clone: SOL-HYPERLIQUID dual-TF momentum-confluence is the same OHLCV trend-confluence clone (0/213) on the same short-history venue (0/97) that the developer's own hypothesis says killed four prior HL siblings for unestablished edge. The 4H+1D choice still yields only ~2.3y / 172 trades, and sharpe_ci_low is already 0.3945 (below the 0.5 OOS floor) before optimization. Optimization cannot manufacture the sample size the bootstrap-CI/deflated-Sharpe gate requires — a ~225-trial search on 172 trades curve-fits noise and dies at DSR/holdout, exactly as the LINK/ETH/AVAX/SOL-HL ports did. The prescribed and validated fix for this mechanism is the long-history venue (the ADA Binance variant reached risk_review); this re-run of the four-times-failed HL experiment is not worth the 2-hour optimization budget.

Long-short, single-instrument dual-timeframe momentum confluence on the Hyperliquid SOLUSD perpetual. The fast leg runs on 4H bars as an ATR-normalized EMA spread ((EMA_fast - EMA_slow)/ATR, with EMA_slow = fast_ema * ema_ratio) so the trend read is volatility-scaled; the slow leg runs on 1-DAY bars of the same instrument (consumed in on_extra_bar) as the sign of the daily_lookback-day return. It goes LONG only when both the 4H spread and the daily trend are positive, SHORT only when both are negative, and stays FLAT on any disagreement. Positions exit on an ATR trailing stop (give-back of trail_atr_mult x entry ATR from the best close since entry), on the daily trend flipping against the position, or on the 4H spread flipping. Sizing is risk-first: qty = equity * risk_pct / (trail_atr_mult * ATR), capped at max_notional_frac * equity * leverage, with the Hyperliquid $10 min-notional enforced. Pure OHLCV, one instrument, two timeframes, three tunable parameters.

Pure-OHLCV mechanism-family risk (L7/L41): 4H+1D dual-timeframe momentum confluence has a poor survival record and was recently abandoned in an equity variant (SpyDualTimeframeTrendConfluenceLS). The hypothesis frames this as a venue port of the surviving BTC/ADA variants, so the correctness bar is met, but the analyst should treat the OHLCV pedigree as elevated risk. Not a code defect.

Modest statistical strength: Sharpe 0.914 with CI [-0.884, 2.398] straddling zero, win_rate 0.39, 59 trades over 364 days, kurtosis 12.5 with one 8407-USD win doing heavy lifting. Edge is real and fee-positive (avg_trade_return_pct 1.22% vs ~0.09% HL round-trip) but not yet distinguishable from noise.

Minor state-hygiene nit: should_enter sets self._side before size is known; if position_size returns 0 (below min notional) no order submits yet _side is left set, and the disagreement branch returns None without resetting it. Cannot create a phantom position (re-set on next genuine entry). Harmless in backtest.

Clean, faithful reproduction of the one surviving template (4H+1D confluence, long-short, ATR-trailed, 3 params); code is O(1)/bar and structurally sound

Per-trade economics clear the HL fee floor (avg_trade_return_pct 1.74%, PF 1.91, max_dd 6.5%), and it fills a real venue/direction quota gap

Positive all three partial years (2024 +19.9%, 2025 +19.3%, 2026 +14.4%)

Self-defeating premise: the hypothesis itself states HL's ~2.3y history is why every prior HL sibling (LINK/ETH/AVAX/SOL) was abandoned for unestablished edge, and the prescribed fix (the ADA variant) was to move to Binance's 5-6y history — this variant re-enters the exact short-history trap

Only 172 trades over ~2.6y — mid-count overfit signature (L41), too thin to clear the bootstrap-CI/deflated-Sharpe gate that killed the siblings

sharpe_ci_low 0.3945 already straddles the 0.5 OOS floor pre-optimization; the ~225-trial deflation will select-and-inflate the point Sharpe while the CI stays wide -> DSR death

Class prior: OHLCV trend-confluence clone (0/213) on Hyperliquid (0/97); recent rolling Sharpe went negative (-2.18) in Jul-Aug 2026

point estimate flattering

too thin for HL significance gate

SolHyperliquidDualTimeframeMomentumConfluenceLS was a clean, faithful reproduction of the factory's one surviving recipe — the 4H+1D long-short momentum-confluence trend follower that took the BTC variant to paper and the ADA Binance variant to risk_review — ported to SOLUSD on the under-represented Hyperliquid venue to test venue-robustness and fill quota gaps. Its point metrics looked strong (Sharpe 1.35, PF 1.91, 6.5% drawdown, positive all three partial years), but they rested on just 172 trades over ~2.3 years of usable history. The analyst abandoned it at the backtest-review gate using the hypothesis's own admission: this is the same short-history significance trap that had already killed four HL siblings, its sharpe_ci_low of 0.39 was below the 0.5 OOS floor before optimization even began, and the prescribed, validated fix for this exact mechanism was the long-history Binance venue — so re-running the four-times-failed HL experiment was not worth the optimization budget.

A faithful port of a surviving template still fails if it re-enters a known structural trap — short venue history and a thin trade count leave the significance gate unclearable no matter how good the point estimates look, and the fix (validated on the ADA Binance variant) is more history, not another short-history re-run.

The analyst issued an 'abandon' verdict at the pre-optimization backtest-review gate on the strategy's own logic: the hypothesis itself conceded that Hyperliquid's ~2.3-year history is why four prior HL siblings were abandoned for unestablished edge and that the validated fix was moving to a long-history venue (the ADA Binance variant reached risk_review). With only 172 trades and a CI-low already below 0.5, optimization could not manufacture the sample size the bootstrap-CI/deflated-Sharpe gate requires, so it would curve-fit noise and die at DSR/holdout. It never reached optimization, post-opt analyst review, or risk review.

A long-short, single-instrument pure-OHLCV dual-timeframe momentum-confluence trend follower on SOLUSD Hyperliquid perp (4H primary + 1D confirm), going long/short only when the 4H EMA spread and daily trend agree and flat on disagreement, with ATR-trailed winners and 3 tunable parameters — a deliberate venue-diversification port of the factory's only recently-surviving template (the BTC and ADA dual-TF confluence variants) onto the under-represented Hyperliquid venue.

Over ~2.6 years and only 172 trades (84 long / 88 short, 44.8% win rate) it showed flattering point metrics — Sharpe 1.346, profit factor 1.91, 6.5% max drawdown, avg_trade_return_pct 1.74%, positive in all three partial years (2024 +19.9%, 2025 +19.3%, 2026 +14.4%) — but the sharpe_ci_low of 0.3945 already sat below the 0.5 OOS floor before any optimization, and recent rolling Sharpe had gone negative in Jul-Aug 2026.