BtcEthHourlyFundingDifferentialSpreadLS

BTC-ETH Hourly Funding Rate Differential Spread on Hyperliquid (Relative-Value Long-Short Pair, 1H Bars)

Market-neutral cross-asset RELATIVE-VALUE strategy trading the differential between BTC and ETH hourly funding rates on Hyperliquid perpetuals. This hypothesis EXPLICITLY addresses the post-mortem from the abandoned `BTC Coin-M Inverse Perp Extreme Funding Rate Mean-Reversion` strategy, which recommended: 'Future funding-based strategies should consider: (a) funding-rate regime-change detection, (b) cross-asset funding spreads (BTC vs ETH funding differential as a relative-value signal), or (c) funding-rate momentum.' This proposal implements option (b) — a fundamentally different mechanism from the failed 'fade-the-spike' single-asset reversion. Core thesis: BTC and ETH perpetual funding rates are both real-time gauges of leverage demand for the same asset class, so they share most common factors and their DIFFERENCE is a stationary mean-reverting series even when individual funding rates trend. When the BTC-ETH funding spread becomes extreme (z > 2σ over rolling 168-hour window), one asset's longs are over-positioned relative to the other → trade the relative-value convergence by longing the under-positioned leg and shorting the over-positioned leg. Hyperliquid is chosen because (a) its HOURLY funding cycle (vs Binance 8h) gives 8x more signal samples and lets us trade intraday relative-value drift, (b) HL is heavily under-represented (4.3% vs 20% target), (c) HL has both BTC and ETH perps with deep history. Fills FOUR under-represented quotas simultaneously: HYPERLIQUID venue, pairs scope, long-short direction, and uses funding-rate data (different data class from the bar-only pattern strategies that have flooded the failure log). NOT a classical chart pattern, NOT single-asset funding extreme (dead), NOT a directional bet on either asset.

The failure was structural fee drag (22.2% of gross, -6.66 Sharpe) from churning the entry_z=1.5/exit_z=0.5 band on ~6.7h round-trips whose ~1-sigma captured move could not clear the ~0.09% HL round-trip x2 legs. Per the analyst's directive #1, the smallest structural fix is a frequency cut, implemented with three confined edits and no change to the (already-passing) z-score signal, leg routing, or ETH hedge reconciliation: (1) entry_z 1.5->2.5 so we only enter at genuine extremes, cutting trade count sharply; (2) exit_z 0.5->0.25 so each convergence trade captures ~2.25-sigma instead of ~1-sigma, widening per-trade move relative to fixed fees; (3) a new min_hold_bars=12 gate that blocks the convergence-exit path (the churn source) until a trade is old enough for captured convergence to exceed cost, while hard stops (stop_z, lengthened max_hold_bars=72) stay always-active so risk control is preserved. Imports are unchanged (no banned additions) and the signal-computation code is byte-identical, so Layers 1-2 remain green. This is the minimal change that can flip the fee math; if a low-frequency book still bleeds every month on the ~146-day sample, that confirms the BTC-ETH funding differential has no tradable net-of-fee RV edge here and the strategy should then be abandoned rather than further optimized.

Negative-expectancy market-neutral funding-differential pair on a single-regime, data-limited sample — not worth 2 hours of optimization. Funding is now credited by the engine (3,793 events), so this is a genuine economic result, and it loses: total_return -9.17%, profit_factor 0.773, avg_trade_return_pct -0.017% (negative), expectancy -$37/trade over 236 trades, Sharpe -1.17 with CI [-3.32, 0.86] and PSR 0.22, red in nearly every month. The payoff is inverted (avg_loss $322 > avg_win $262 at a 48.7% win rate) and commission is still 16.95% of gross — the BTC-ETH HL funding spread's convergence does not clear the two-leg cost even after the iteration-2 frequency cut. It is also unvalidatable: Hyperliquid BTCUSD/ETHUSD hourly history spans only ~11 months (2025-10 → 2026-09, only_one_regime true), so walk-forward/holdout would sit in one regime. This is the L57/L44 cross-venue funding relative-value family crossed with the L45 market-neutral pairs class (0/84, near-zero survival); benchmark_meaningful is correctly false. No parameter tuning flips a negative-expectancy, inverted-RR funding pair, and the premise (profitable mean-reversion of the BTC-ETH funding spread) is falsified on the available data. Failure pattern: no_edge/fee_edge market-neutral funding-differential pairs, single-regime data-limited (L44/L45/L57).

Market-neutral BTC-ETH relative-value pair on Hyperliquid perps (1H bars) trading the hourly funding-rate differential (BTC_funding - ETH_funding). Computes the spread's z-score vs a rolling 168h window; at a 2.5-sigma extreme it fades the divergence (short the funding-rich leg, long the funding-cheap leg), dollar-neutral. Iteration 2 attacks the fee-drag that sank iteration 1: entry raised to 2.5-sigma, exit tightened toward the mean (0.25-sigma) so each trade captures ~2.25-sigma of convergence, and a hard 12-bar minimum-hold gate forbids the fast in-out churn while regime-break (4-sigma) and 72-bar time stops remain always active.

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.

Analyst to decide abandon (developer's recommendation) vs a structural rethink (widen the stop so the fade has room, or confirm the spread's half-life supports the hold) rather than an entry_z sweep that cannot manufacture mean-reversion the data does not show.

The convergence thesis is contradicted by trade behaviour: entries at 2.5-sigma are frequently stopped at the 4-sigma regime-break level (the only exit that can fire before the 12-bar min-hold, hence avg holding 10h33m < 12h), meaning the funding spread continues diverging after entry rather than reverting. A relative-value fade only works if the series mean-reverts within the hold; here it does not clear fees, so avg_loss ($342) > avg_win ($305), PF 0.745. Not a code defect — entry/exit/stop logic is correctly wired — but the edge is absent in this regime.

Clean dollar-neutral construction; funding is now credited by the engine so the result is economically real, and the iteration-2 min-hold gate did cut churn

Low drawdown (9.9%), no liquidation

Negative expectancy: total_return -9.17%, PF 0.773, avg_trade_return_pct -0.017%, expectancy -$37/trade, Sharpe -1.17 (CI [-3.32, 0.86]), PSR 0.22 — losing in nearly every month

Inverted payoff: avg_loss $322 > avg_win $262 at 48.7% win rate; commission still 16.95% of gross

Data-limited to a single 11-month regime (HL BTC/ETH hourly 2025-10 → 2026-09, only_one_regime true) — unvalidatable across regimes

L57/L44 cross-venue funding relative-value + L45 market-neutral pairs (0/84); benchmark_meaningful correctly false

data_unavailable: backtest data for BTCUSD.HYPERLIQUID could not be collected after 5 retries: - bars (bar_type=BTCUSD.HYPERLIQUID-1-HOUR-LAST-EXTERNAL): got 0, needed 200, or 30+ days span The data layer (not the strategy) is the blocker — re-evaluate this hypothesis once collection is fixed / the symbol is backfilled.