SolHyperliquidDualTfTrendConfluenceVolRegimeLS

Hyperliquid SOL Dual-Timeframe Trend Confluence with Volatility-Regime Gate, Long-Short (Single-Instrument SOLUSD.HYPERLIQUID, 4H Primary + 1D Trend Gate — Take a Directional Position ONLY When the 4H and 1D Trends Agree AND Realized Vol Is in a Trending Regime; Flat in Chop; ATR Trailing Exit, 3-Parameter)

A LONG-SHORT, SINGLE-INSTRUMENT, PURE-OHLCV directional trend-follower on SOLUSD.HYPERLIQUID perpetual, deliberately built to (a) extend the factory's ONE demonstrated survivor — 'BTC Dual-Timeframe Momentum Confluence' (Sharpe 1.99, now in paper) — onto a DIFFERENT venue and instrument, and (b) fill the single most under-represented venue bucket (HYPERLIQUID at 5.9% vs a 20% target) while also pushing the direction mix toward long_short (currently only 13%). It avoids every currently-dead idea family: no COIN-M (L51), no options (L50), no cross-sectional momentum basket (L52), no non-price gate on a price impulse (L46), no liquidation/quarterly feed that risks data_unavailable (L47), and no long-only structural spot basket (L48). Signal is computed on 4H bars (~2.3 years of usable Hyperliquid history — safely above the sub-hourly HL history limit) with a 1D trend confirmation (full HL history). Only ONE side is ever held, and only when BOTH timeframes agree in direction AND the market is in a trending (not choppy) volatility regime; otherwise the book sits flat. Three parameters only, to resist the overfit failure mode (301 prior overfit deaths).

Iteration-2 fix for the Layer-3 'should_enter() returned a side 0 times over 2017 bars'. The entry logic was NOT the problem — my pre-submit dry-run against the real catalog showed the full compound gate (both timeframes agreeing AND vol_rank >= 0.35) satisfied on 781 of 2,031 evaluable sandbox bars. The failure was config plumbing: InstrumentResolver.resolve_extras() iterates `for i, extra_instr in enumerate(extra_instruments)` and reads extra_bar_types[i] inside that loop, so with `extra_instruments: []` the loop body never executes and my 1-DAY bar type was silently dropped from the sandbox config. No daily bars were ever delivered, `_daily_ready` stayed False forever, and `should_enter` bailed on the `not self._daily_ready` guard on every single bar — exactly the reported diagnostic. PRIMARY FIX (config): `extra_instruments` now lists SOLUSD.HYPERLIQUID alongside the 1-DAY `extra_bar_types` entry, so the resolver emits the pair and the daily bars reach `on_extra_bar`. The instrument is the same as the primary, which the runner dedupes via `needed_ids`, and since 1-DAY differs from the 4-HOUR primary the base class's same-timeframe alignment barrier still (correctly) does not gate it — no deadlock. SECONDARY FIX (code, 12 lines): if the daily buffer is short or absent, daily_mom is resampled from the primary 4H closes over the identical horizon (bars_per_day * trend_lookback_d = 60 4H bars = 10 days). This is the SAME quantity on the SAME instrument and price series — not a substitute edge mechanism, and dormant whenever real daily bars are present — so a missing or still-filling daily subscription can never again turn the strategy into a silent no-op in the sandbox, in paper, or after a live restart. Everything that already passed Layers 1-2 is untouched: identical imports, class structure, continuous fast_mom signal, confluence entry, daily-flip + ATR-trail exits, cooldown, and risk-capped sizing. The only other change is min_bars_required now also covering the resample span (still 167 bars, unchanged in practice).

Class-prior abandon at BACKTEST_REVIEW (L41): single-symbol momentum/trend-confluence on Hyperliquid has 0/97 HL and 0/213 OHLCV survival, and direct siblings (ETH HL dual-TF Sharpe 1.79/PF 2.13, SOL HL Donchian) each went optimize→abandon after wasting the full 3-phase run. This SOL HL dual-TF confluence is a near-clone with WEAKER metrics — Sharpe 0.77 with a CI [-0.26, 1.75] that straddles zero, i.e. not distinguishable from no edge before the optimizer's best-of-N selection even inflates it. The +49% return is carried by ~4 outsized days; the per-day edge is thin and regime-dependent and will collapse in walk-forward OOS/holdout exactly as the siblings did. No robust parameter region is evident to justify 2 hours of optimization. Not a fee or trade-count failure (avg_trade_return_pct 1.04%, 120 trades) — the mechanism is simply in a family that does not generalize here.

Long/short pure-OHLCV trend-follower on SOLUSD.HYPERLIQUID perpetual, 4H bars with a 1-DAY trend confirmation of the same instrument. Every 4H bar it computes fast_mom (4H momentum over fast_lookback bars), daily_mom (daily momentum over trend_lookback_d days), and vol_rank (percentile rank of current realized vol within its own trailing distribution). It goes LONG only when both horizons are positive, SHORT only when both are negative, and only while vol_rank >= vol_regime_pct; on disagreement or compressed volatility it sits flat. Positions exit when the daily trend flips against them or an ATR trailing stop is hit, with a short cooldown after each exit to prevent fee churn. Sizing risks 1% of equity against the initial 1.5x ATR stop, hard-capped at 50% of equity notional.

Verification failed (Layer 3 — sandbox backtest): No trades produced Bar type used: SOLUSD.HYPERLIQUID-4-HOUR-LAST-EXTERNAL, Bars processed: 2183 Diagnostics: should_enter() returned a side 0 times over 2017 evaluated bars -> your ENTRY CONDITION never triggered. Loosen the entry logic / thresholds. Ensure your strategy produces trades with the given data and parameters.

Acceptable to proceed, but treat optimizer parameter estimates as low-confidence; if the full-history run returns materially fewer than ~100 trades, loosen vol_regime_pct rather than fitting a thin sample.

Trade-count is near the measurability floor. Sandbox produced 44 trades over 364 days; extrapolating to the full ~2.3yr usable Hyperliquid 4H history implies only ~100-110 trades — right at the ~100-trade threshold below which edge cannot be cleanly separated from noise. The walk-forward IS/OOS windows and 15-day holdout will each see very few events.

Log a one-time warning when the resample fallback engages so a genuinely missing daily subscription is visible in paper/live.

The daily confirmation leg silently falls back to a 4H-resampled proxy when real 1-DAY bars are short/absent. The config plumbing is now fixed so real daily bars should arrive, but the fallback masks any future delivery gap — the strategy keeps trading on the proxy without surfacing a dead daily subscription. Same quantity/instrument, so not a correctness defect, but it removes an observability signal.

Optional: rank against the trailing window excluding the current point (hist[:-1]), or leave as-is.

_update_vol_rank appends the current vol before ranking (mean(hist <= vol) includes the point itself), so vol_rank can never fall below 1/vol_rank_lookback. With vol_regime_pct=0.35 the effect on the gate is immaterial; the percentile is very slightly upward-biased. Cosmetic only.

Mechanism faithfully implements the hypothesis: balanced long/short (64L/56S), 120 trades over 922 days, trades match the dual-timeframe confluence + vol-regime design

avg_trade_return_pct 1.04% is comfortably above the ~0.15% Hyperliquid fee floor — not a fee-fragility death

Reasonable risk profile on the surface: PF 1.57, max_drawdown 15.9%, only 3 tuned parameters (resists selection inflation)

Single-symbol momentum/trend-confluence on Hyperliquid is a graveyard class: 0/97 HL and 0/213 OHLCV survival (L41). Direct siblings ETH HL dual-TF and SOL HL Donchian both went optimize→abandon after burning the full 3-phase run

Metrics are WEAKER than the dead siblings (Sharpe 0.77 vs their 1.79) — with the Sharpe CI [-0.26, 1.75] straddling zero, the point estimate is not statistically distinguishable from no edge before the optimizer even inflates it via best-of-N

Returns are concentrated in ~4 outsized days (2026-02-08 +12.66%, 2026-06-12 +9.56%, 2025-03-02 +7.47%); the underlying per-day edge is thin and regime-dependent — walk-forward OOS will not reproduce it

information_ratio -0.06 and alpha barely positive on a non-meaningful benchmark; nothing here suggests a robust parameter region for optimization to find

SolHyperliquidDualTfTrendConfluenceVolRegimeLS ported the factory's one demonstrated survivor — BTC dual-timeframe momentum — onto Hyperliquid SOL as a long-short, vol-regime-gated trend-follower, filling the under-represented HYPERLIQUID venue and long-short buckets. Its first backtest looked superficially healthy (+49.5% return, PF 1.57, 120 balanced trades, 1.04% avg trade, 15.9% max drawdown), but Sharpe was just 0.77 with a confidence interval spanning zero and the gains leaned on about four outsized days. At backtest review the analyst issued an abandon verdict on class-prior evidence: single-symbol momentum on Hyperliquid has never survived (0/97), stronger-metric siblings had already burned the full optimization only to abandon, and no robust parameter region was visible. It was abandoned after 2 iterations before optimization, analysis, or risk review — not on fees or trade count, but because the mechanism does not generalize on this venue.

A backtest that clears fee and trade-count floors can still be correctly abandoned pre-optimization when its Sharpe CI straddles zero and it belongs to a strategy family with a proven zero-survival track record — profitable-looking headline returns carried by a handful of days will not reproduce in walk-forward OOS.

The analyst abandoned it at backtest review (pre-optimization) on class-prior grounds: single-symbol momentum/trend-confluence on Hyperliquid is a graveyard family (0/97 HL, 0/213 OHLCV survival), direct siblings with stronger Sharpe (1.79) had already gone optimize→abandon, and with a thin regime-dependent edge whose Sharpe CI straddles zero, no robust parameter region justified spending the full 3-phase optimization.

A long-short, single-instrument pure-OHLCV trend-follower on SOLUSD.HYPERLIQUID (4H signal + 1D trend gate, 3 tuned parameters) that took a directional position only when the 4H and 1D trends agreed and realized volatility sat in a trending regime, exiting on a daily-trend flip or ATR trailing stop — extending the factory's one paper survivor (BTC dual-timeframe momentum) onto a new venue.

The initial backtest (SOLUSD.HYPERLIQUID 4H, 922 data days, 2024-01 to 2026-08) returned +49.5% with a balanced 120 trades (64 long/56 short), profit factor 1.57, max drawdown 15.9%, win rate 0.44, and avg_trade_return_pct 1.04% — comfortably above the fee floor. But Sharpe was only 0.77 with a confidence interval of [-0.26, 1.75] straddling zero, and the return was concentrated in roughly 4 outsized days.