Strategies

ETH/BTC Ratio Mean-Reversion Pairs Spread on Hyperliquid Perps (Market-Neutral, OHLCV-Only, Z-Scored, 4H)

Outcome: Abandoned

EthBtcRatioMeanReversionPairsHL4H

Outcome Summary

EthBtcRatioMeanReversionPairsHL4H was hailed at backtest review as the first genuinely novel, clean mechanism in a long failed streak — a market-neutral ETH/BTC ratio stat-arb on the canonical crypto pair, dodging every active engine defect and built on just 3 parameters — so despite a losing base case it earned a trip to optimization. Optimization then proved the edge negative everywhere: every sensitivity cell lost, walk-forward collapsed from +2.62 in-sample to -1.55 out-of-sample, the holdout failed, and DSR was 0.0, with the best-of-225 config still down 21%. The root cause was a falsified premise: the ETH/BTC ratio trends with rotation flows rather than reverting at a tradeable 4H scale, so the fade sold convergences that never came while fees compounded the bleed. The analyst abandoned it as a dead mechanism class — cross-asset pairs mean-reversion failed across multiple pairs this session — concluding the market-neutral construction is genuine but there is simply no reversion edge to harvest, and re-pointing the fade at another correlated pair would only repeat the decay.

Strategy report

Hypothesis

A market-neutral, two-leg statistical-arbitrage pairs strategy on Hyperliquid perpetuals: trade the ETHUSD/BTCUSD price ratio when it dislocates from its rolling mean, holding a beta-balanced long-short pair so net crypto-beta is ~zero. ETH and BTC are the two most correlated crypto majors, so their price RATIO is far more stationary than either leg alone — relative dislocations (driven by rotation flows, ETH-specific catalysts, or transient liquidity imbalances) tend to revert because both assets share the same macro beta. This deliberately AVOIDS every active failure mode in the recent log: it is NOT a single-name directional/reversion bet (ETH single-name on HL has failed momentum, breakout, AND VWAP-reversion — this trades the residual ratio, not ETH's direction); NOT funding-as-signal (the entire funding-directional family is a graveyard — this uses OHLCV only); NOT supplementary-data-gated (no OI/liquidations/options-IV, which only cover a ~1-month recent sliver and render strategies untestable); NOT options (dodges the ~0-capital options sizing bug); and NOT CASH-spot (perps on a MARGIN venue, dodging the multi-instrument cash-overfill engine defect). It fills three under-represented quota buckets simultaneously: HYPERLIQUID (7.6%->target 20%), pairs scope, and long-short/market-neutral (14%->more headroom). Only 3 core parameters (z-window, entry-z, exit-z) to resist the overfitting that has recurred this session.

Implements the hypothesis exactly: a market-neutral, two-leg statistical-arbitrage pair on the ETH/BTC ratio using Hyperliquid perps and OHLCV only. It trades the residual ratio (not either coin's direction), uses no funding/OI/liquidations/options/supplementary data (so it's fully testable over the long history), and runs on a MARGIN futures venue (HYPERLIQUID) with both long and short legs -- avoiding the spot CASH multi-instrument overfill defect and the options-sizing bug. Only 3 core parameters (zscore_window, entry_z, exit_z) drive the edge, with stop_z and a time stop as risk controls, resisting overfitting. Timestamp-keyed close dicts guarantee leg alignment (no index-trim bug); the z-score excludes the current bar to avoid look-ahead. Dollar-neutral equal-notional legs approximate beta-balance since beta(ETH/BTC) ~= 1 for the two majors, so net crypto-beta ~= 0. Leverage (2.0) is wired into per-leg sizing (equity * per_leg_pct * leverage / price), making the margin venue's leverage actually consumed. Round-trip cost on Hyperliquid is 4 x 0.045% ~= 0.18%; an entry_z=2 -> exit_z=0.5 reversion captures a multi-sigma ratio move well above that, keeping avg trade return viable. Fills three under-represented quota buckets: HYPERLIQUID, pairs scope, and long-short/market-neutral.

No edge anywhere in the parameter space — not fixable by iteration or by reframing. Both backtests lose money (initial Sharpe -0.63 / -30.8%; optimized best-of-225 Sharpe -0.70 / -21.1%, expectancy -$106/trade), the sensitivity grid is NEGATIVE in every cell (best -0.46, so no profitable region to tune toward), walk-forward is_overfitted=TRUE (avg IS +2.62 -> avg OOS -1.55, windows [-4.14, -2.39, +1.88]), the holdout failed (sharpe -0.03, ratio 0), and robustness is conclusive (deflated_sharpe 0.0, is_significant false, PSR 0.20, sharpe_ci_low -2.06, optimized Sharpe far below the 5.10 expected-max-under-noise bar). The mechanism executes correctly and is genuinely market-neutral (beta 0.087, balanced 86L/86S), so this is a true no-edge result, not a bug: the ETH/BTC ratio trends with rotation flows rather than mean-reverting at a tradeable, fee-clearing 4H scale, so the z-score fade systematically sells a convergence that doesn't arrive while paying ~3.5% of gross in commissions + impact. NOT iterate (attempt 1 of 2): the edge is negative across the entire grid and the optimizer's own best config still loses 21% with DSR 0.0 — a second sweep re-searches the same uniformly-losing surface. NOT revise_hypothesis: cross-asset pairs/cointegration mean-reversion has NO promoted sibling in this factory and has failed across multiple pairs this session (SOL/AVAX, LTC/BCH, ETH/BTC all catastrophic), so this is a dead mechanism, not a proven one stranded on one dead target — re-pointing the same fade at another correlated pair would re-hit the identical decay. FAILURE PATTERN: crypto-major ratio/pairs mean-reversion (z-scored ETH/BTC on HL 4H) is a graveyard — the ratio is directionally trending not stationary, so the fade loses in-sample and out-of-sample (is_overfitted=TRUE, all-negative sensitivity grid, DSR 0.0, failed holdout) while fees compound the bleed; market-neutral construction is genuine but there is simply no reversion edge to harvest.

Implementation

Market-neutral ETH/BTC ratio mean-reversion pairs trade on Hyperliquid perpetuals (4H bars, OHLCV-only). Each 4H close, ETHUSD and BTCUSD closes are timestamp-aligned, the log ratio log(ETH/BTC) is computed and z-scored over a 90-bar (~15-day) rolling window with the current observation excluded (look-ahead-safe). When |z| >= entry_z (2.0) the strategy opens a dollar-neutral two-leg position (z high: short ETH + long BTC; z low: long ETH + short BTC) in equal USD notional, sized capital- and leverage-relatively (per_leg = equity * 0.50 * 2.0x). Positions close when |z| <= exit_z (0.5, reverted), |z| >= stop_z (4.0, relationship broke), or after max_holding_bars (60, ~10-day time stop).

Backtest Review

Sharpe: -0.99
Total return: -36.57%
Max drawdown: 37.40%
Trades: 190
Win rate: 51.6%
Profit factor: 0.89

First genuinely novel, clean mechanism in a long failed streak: market-neutral two-leg ETH/BTC ratio stat-arb, a legitimate well-established edge class (the canonical crypto pair). Genuinely neutral (beta 0.087, correlation 0.118, benchmark_meaningful=false) and distinct from every failed single-name directional book.

Dodges ALL active engine defects: perps on a MARGIN venue (not CASH multi-instrument cash-overfill), OHLCV-only (not funding-directional or supplementary-gated), not options. Sizing is plausibly clean (avg_position_pct 109%, ~100% gross for 50%/leg, metrics_reliable=true).

Fully validatable cadence: deep 4H data (5004 bars, ~1.3yr from 2024-01-22), 172 trades (86 balanced pairs) — the walk-forward windows and 15-day holdout will be well-populated, so optimization is statistically meaningful and the post-opt gates can rule rigorously.

Only 3 core parameters (z-window, entry-z, exit-z) plus risk controls — built to resist the overfitting that recurred this session.

Near breakeven (PF 0.91, win_rate 50.6%) with losses regime-concentrated in 2025 (2024 ≈flat, 2026 +3.9%), not a uniform bleed.

Base case is a net loser: total_return -30.8%, Sharpe -0.63, avg_trade_return -$234, max_drawdown 56%.

Heavy left tail / short-gamma geometry: skew -3.93, kurtosis 44, largest_loss -$34.5k (≈-34% on one pair), avg_loss $5035 > avg_win $4457. stop_z=4.0 (wider than entry_z=2.0) lets adverse moves bleed before cutting — the structural risk of fade-the-dislocation pairs.

Premise tension: the ETH/BTC log-ratio is not reliably stationary — 2025's -41.5% came from a sustained ratio trend that z-reversion got run over. Optimization must show the edge survives a trending-ratio regime, not just curve-fit around 2025.

Analysis

Mechanism executes cleanly and is genuinely market-neutral: 172 trades balanced 86 long / 86 short, beta 0.087, benchmark_correlation 0.118, metrics_reliable=true — no sizing/leg-netting artifact

Sound engineering: look-ahead-safe z-score (current obs excluded), timestamp-keyed leg alignment, dodges the CASH-overfill and options-sizing defects as designed

Sensitivity is internally stable (no chaotic cliffs) — the negativity is consistent, not noise

Loses money in BOTH backtests: initial Sharpe -0.63 / total_return -30.8% / PF 0.91; optimized (best of 225) Sharpe -0.70 / total_return -21.1% / PF 0.94 — negative expectancy per trade in both

Sensitivity grid is negative in EVERY cell (best -0.46) — no profitable parameter region exists to tune toward

Walk-forward is_overfitted=TRUE: avg IS +2.62 -> avg OOS -1.55, OOS windows [-4.14, -2.39, +1.88]; holdout failed (holdout_sharpe -0.03, ratio 0)

Robustness conclusive: deflated_sharpe 0.0, is_significant false, PSR 0.20, sharpe_ci_low -2.06, optimized Sharpe far below the 5.10 expected-max-under-noise bar

Fee-fragile: commission 3.48% of gross + $8.6k impact fighting a ratio that trends rather than reverts

Outcome Summary

Crypto-major ratio/pairs mean-reversion is a graveyard this session (SOL/AVAX, LTC/BCH, ETH/BTC all failed) — the ratio is directionally trending, not stationary, so the z-fade loses in- and out-of-sample regardless of parameters; combined with the failed ratio-momentum sibling, both directions on ETH/BTC are dead, and a genuinely market-neutral construction cannot manufacture a reversion edge that isn't there.

After optimization the analyst abandoned it because there is no edge anywhere in the parameter space: the sensitivity grid is negative in every cell, walk-forward is_overfitted=TRUE (avg IS +2.62 → avg OOS -1.55), the holdout failed (-0.03), and robustness was conclusive (DSR 0.0, is_significant false, PSR 0.20) — the ETH/BTC ratio trends with rotation flows rather than mean-reverting at a fee-clearing 4H scale, so the z-fade sells a convergence that never arrives while paying ~3.5% of gross in costs.

A market-neutral, OHLCV-only ETH/BTC ratio pairs stat-arb on Hyperliquid 4H perps that z-scores the log(ETH/BTC) ratio and fades dislocations — shorting the rich leg and longing the cheap leg as a beta-balanced pair — with only 3 core parameters (z-window, entry-z, exit-z) to harvest the canonical crypto-pair convergence while dodging the CASH-overfill, options, and funding-directional failure modes.

It lost money in both backtests: initial total return -30.8% (Sharpe -0.63, PF 0.91, avg_trade_return -$234, 56% max drawdown, heavy left tail with skew -3.93 and avg_loss $5,035 > avg_win $4,457) over 172 balanced pairs, and the best-of-225 optimized config still lost -21.1% (Sharpe -0.70, PF 0.94); the construction was genuinely market-neutral (beta 0.087) and metrics_reliable=true.

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