HyperliquidEthBtcRelativeStrengthTrendSpreadLS

Hyperliquid ETH/BTC Relative-Strength Dual-Timeframe Trend Spread, Dollar-Neutral Long-Short (ETHUSD.HYPERLIQUID vs BTCUSD.HYPERLIQUID — Trade the TREND of the ETH/BTC RATIO Itself: Long ETH/Short BTC When Relative Strength Trends Up and Both 4H + 1D Agree, Reverse When It Trends Down, Flat on Disagreement; Dollar-Neutral, ATR-Trailing Exit, 3-Parameter)

A DOLLAR-NEUTRAL, LONG-SHORT PAIRS strategy on Hyperliquid that trend-follows the ETH/BTC RATIO rather than either asset's absolute price. It deliberately reuses the ONLY mechanism the factory has validated out-of-sample — dual-timeframe trend confluence (the BTC 4H+1D momentum survivor at Sharpe ~1.99) — but applies it to the RELATIVE-STRENGTH spread between the two largest crypto assets, which converts an absolute-momentum edge into a market-neutral one. It fills four under-represented buckets at once: HYPERLIQUID venue (6.1% vs 20% target), pairs scope (13.0%), long_short direction (13.0% vs 87% long-only), and market-neutral exposure. It is distinct from every pipeline item (single-instrument BTC/SOL trend, cross-venue funding differential, long-only baskets) and avoids all dead families: NOT a cross-sectional momentum-RANK basket (L52 — this is one two-leg spread trend, no ranking/rotation), NOT COIN-M (L51), NOT options (L50), NOT a long-only structural basket that inherits full bear-market beta (L48 — dollar-neutrality bounds drawdown), NOT a non-price gate on a price impulse (L46 — the signal IS the ratio's own price trend), and NOT a liquidation/quarterly feed at risk of data_unavailable (L47 — both ETHUSD and BTCUSD perps have full Hyperliquid history; at 4H that is ~2.3 years, above the HL sub-hourly history limit, and 1D is full). Pure OHLCV, 3 parameters, to resist the 301-death overfit mode.

Iteration 3 makes exactly two targeted changes against the two named Layer-4 failures, leaving all Layer 1-3-passing scaffolding (imports, class structure, synchronised ratio buffer, continuous fast_mom signal, two-leg dollar-neutral submission, sizing) untouched. (1) NEGATIVE PER-TRADE EDGE: iteration 2 entered on a slow move above min_trend but exited only at the ZERO crossing, so each round trip surrendered the entire entry threshold back to the market before closing. The exit now fires at min_trend * exit_frac (exit_frac frozen at 0.5), halving the give-back per trade, and the trailing stop is tightened from 4.0 to 3.0 ratio-ATRs so a completed move cannot round-trip through the entry. The band between exit_level and min_trend preserves the hysteresis that stopped small oscillations from re-arming entries. (2) UNMEASURABLE TRADE COUNT (38 trades / 363 days): both legs move together from 1-DAY to 12-HOUR, doubling decision points while every horizon stays expressed in days (slow_bars = trend_lookback_d * 2), so the trend being traded is identical and only the entry/exit resolution is finer. With the earlier exit this should put the count in the 150-250 range over the ~2.3y span and 2-4 entries inside the 15-day holdout, clearing the zero-trades gate without reverting to the 4-HOUR cadence whose chop drove iteration 1 to -0.151%. Leverage stays 1.0 (gross 0.7x equity across two legs), so no unused-leverage flag. One caveat for the Analyst: the QA read that daily ETH/BTC relative-strength momentum may be structurally fee-bound on a four-way ~0.18% Hyperliquid round trip is plausible, and if this iteration's per-trade return is still negative the mechanism, not the tuning, is the problem.

No edge and capacity-dead market-neutral pairs (L45 0/84; L82 hyperliquid 0/128). profit_factor 1.005, total_return +2.05% over 3.4 years, Sharpe 0.186 with CI [-0.92,1.18] straddling zero. impact_cost_pct 88.7% with capacity_usd only $127k — market impact consumes nearly all gross PnL, so the edge is real only at toy scale. Rolling Sharpe collapses to -13 in 2026 (recent regime negative), so the last-20% holdout would fail. The hypothesis borrows L82's dual-TF confluence, but L82's survivable variant is a single-name directional Binance USD-M major (Sharpe >1.3/PF >2.0/DD <20%) — this is a HL pairs spread at Sharpe 0.19/PF 1.005, a different and zero-survivor object. Strictly worse rerun of the already-abandoned ETH/BTC ratio dual-TF spread. Base Sharpe 0.19 is the ceiling; deflation only lowers it. No parameter tuning flips a ~1.0 PF with 88% impact drag into a scalable edge. Failure pattern: no_edge/capacity market-neutral pairs (L45).

Dollar-neutral long/short pairs strategy on Hyperliquid that trend-follows the ETH/BTC ratio itself. Both legs trade 12-HOUR bars. Every bar it computes the ratio ETHUSD/BTCUSD into one synchronised buffer and derives two horizons on it: a fast momentum (fast_lookback bars) and a slow momentum (trend_lookback_d days x 2 bars/day). When both agree in sign AND the slow relative-strength move exceeds min_trend, it opens an equal-USD-notional spread (long ETH / short BTC when the ratio trends up, reversed when it trends down). It exits when the slow momentum decays back to min_trend * exit_frac (half the entry floor) or when a ratio-ATR trailing stop is hit, with a short cooldown to prevent immediate re-arming. Sizing is capital-relative: risk_pct of equity against the initial ratio stop, hard-capped at max_notional_pct per leg so gross exposure stays bounded and losses cannot compound into position size.

CLEAN RESTART 2026-09-04 — this run's verdict history and learning records were removed and it was restarted from verification. Its previous abandonment came from the pipeline, not from the market: the Layer-2 harness mis-bound @staticmethod helpers (fixed), QA issued terminal performance verdicts on an unoptimized smoke test (removed — QA now judges correctness only), and sandbox timeouts came from backtest-slot starvation (fixed). The hypothesis and the strategy code are unchanged. Verify the code on its merits; performance is decided later by the full backtest and the optimizer.

Dollar-neutral construction bounds drawdown to 15.8% and never liquidates (spread book, no naked beta)

262 trades — an adequate sample, so the flat result is a genuine no-edge read, not an unmeasured one

Data loaded cleanly (no data_unavailable), both legs contemporaneous

profit_factor 1.005 and total_return +2.05% over 3.4 years — no edge net of costs; Sharpe 0.186 CI [-0.92,1.18] straddles zero, PSR 0.62

impact_cost_pct 88.7% with capacity_usd only $127k — impact consumes nearly all gross PnL; edge exists only at toy scale (hard promotion blocker)

L45 market-neutral pairs family (0/84) and Hyperliquid dual-TF class (L82 hyperliquid 0/128) — both zero-survivor here

Decaying: rolling Sharpe falls to -13 across 2026; last-20% holdout sits in a negative regime and would fail the OOS/holdout gate

Strictly worse rerun of the already-abandoned ETH/BTC ratio dual-TF spread (that one Sharpe 0.44 / impact 26.7% / capacity $1.4M)

Verification failed (Layer 4 — QA review): - FEE/EDGE VIABILITY FAILURE (structural, not a code defect). Sandbox reports avg_trade_return_pct = -0.151% over 294 trades — negative and below the 0.10% low-fee-venue floor, against a ~0.18% four-way Hyperliquid taker round-trip. Corroborating: total_return -25.4%, max_drawdown 25.4%, profit_factor 0.73, Sharpe -0.77, win_rate 0.418. Signature is NOT a code-bug signature (no 0% win rate, no |Sharpe|>5, no PF=0), so there is no plumbing defect to fix. The developer's own scan confirms the edge is absent: negative in nearly every parameter cell (fast 6-24, trend_lookback_d 5-15, trail 2.5-6.0), negative in both sample halves, -0.46%/trade over the trailing 365d. The ETH/BTC ratio chops more than either outright leg while costing double the fees. Rejecting at Layer 4 rather than burning a 3-phase optimization that would near-certainly abandon on avg_trade_return_pct < floor.

Verification failed (Layer 4 — QA review): - PREMISE FALSIFIED — edge still negative net of fees after the iteration-2 redesign. Sandbox: avg_trade_return_pct = -0.035% (negative, below the venue floor), profit_factor 0.95 (<1), Sharpe -0.02, total_return -1.64% over 38 trades / 363 days. Moving both legs from 4H to 1-DAY and adding the min_trend=3% floor cut the per-trade fee drag (iteration 1 was -0.151%) but did NOT cross into positive territory. Daily ETH/BTC relative-strength trend does not generate a per-trade gross move that clears the ~0.18% four-way Hyperliquid round trip — the same mechanism failing for the same reason at a slower cadence. - STRUCTURALLY UNMEASURABLE trade count. 38 trades / 363 days implies ~87 over the full ~2.3y daily span — below the ~100-trade noise floor. With ~1 entry per 9.5 days and ~10-day holds, the Phase-3 15-day holdout will near-certainly produce 0-1 trades (HARD zero-trades gate) and walk-forward OOS windows will be sparse. Slowing to daily to cut fee drag directly created this measurability failure.

Verification failed (Layer 4 — QA review): - PREMISE FALSIFIED ACROSS THREE CADENCES — decisive fee/edge viability failure. avg_trade_return_pct is negative at every sampling rate tried: iteration 1 (4H) -0.151%, iteration 2 (1D) -0.035%, iteration 3 (12H) -0.246%. Iteration 3 is the WORST despite two targeted fixes (exit at min_trend*0.5 to halve give-back + a finer 12H cadence). Full sandbox: total_return -8.42%, Sharpe -0.64, PF 0.80, PSR 0.246, 68 trades. The ETH/BTC relative-strength trend does not produce a per-trade move that clears its double-leg fee load, measured now at 4H (chop), 1D (unmeasurable), and 12H (worse than both). The edge is absent, not mistuned.