ETH-BTC Ratio Z-Score Mean Reversion Pairs Long-Short on 4H

A dollar-neutral pairs trading strategy on the ETH/BTC ratio using 4-hour bars on Binance USD-M perpetuals. ETHUSDT.BINANCE is the primary instrument; BTCUSDT.BINANCE is the secondary (hedge) leg. Every 4H bar close, the strategy computes the log-price ratio R = log(ETH_close / BTC_close), then a 30-bar rolling mean (mu) and standard deviation (sigma) of R, yielding a z-score Z = (R - mu) / sigma. When Z falls below -2.0 the ratio is depressed relative to its recent equilibrium and is statistically likely to mean-revert upward, so the strategy goes LONG ETH + SHORT BTC (in equal USD notional, ~25% of equity each leg, dollar-neutral pair). When Z rises above +2.0 the ratio is elevated and likely to revert downward, so the strategy goes SHORT ETH + LONG BTC. Exits occur when Z crosses back through 0 (full mean reversion), or after a 20-bar time stop, or on a -8% per-leg drawdown emergency stop. This hypothesis is structurally orthogonal to every strategy in the pipeline (none are pairs trades, none use a relative-value ratio as the signal) and orthogonal to every failure pattern in the log: it is NOT single-instrument (so it cannot suffer the rare-event regime concentration that killed XRP/ADA/DOGE/LINK/ETH-trend), it is NOT a single-bar pattern (the z-score aggregates 30 bars of co-movement), it is NOT a directional breakout (it is dollar-neutral relative-value), it is NOT mean-reversion-with-technical-gates (no Bollinger or RSI overlay; pure z-score), and it does NOT depend on supplementary data (avoiding the verification framework's no-trades failure on OI/L-S-ratio strategies). Two free parameters total: lookback (30) and z-threshold (2.0) — the lowest parameter count of any hypothesis this session, minimizing overfitting surface. Expected trade frequency: |z| > 2.0 events at 4H frequency occur ~80-150 times per year on the ETH/BTC ratio based on historical co-movement statistics, well above the walk-forward viability floor and the explicit 25-trade/year minimum cited in multiple abandonment notes. The dollar-neutral pair structure also means the strategy is market-direction-agnostic — it earns from co-movement disequilibrium rather than from beta exposure to crypto, which means 2022 bear-market PnL is structurally as productive as 2021/2024 bull-market PnL, breaking the regime-concentration curse that killed every single-alt strategy in the failure log.

Pure relative-value pairs trade with the lowest parameter count (lookback, z_entry) and no supplementary-data dependency, structurally orthogonal to the single-instrument/single-bar/breakout strategies in the pipeline and market-direction-agnostic (earns from ETH/BTC co-movement disequilibrium, not beta). BTC is configured as a SAME-4H-timeframe extra leg, so the base class's cross-leg alignment barrier defers primary processing until the contemporaneous BTC bar arrives -- the z-score and fills use same-timestamp prices, avoiding the phantom-basis bug from a stale T-1 hedge leg. The log-ratio is stored in a single synced buffer (one append per processed primary bar) so the rolling mean/std stay index-aligned, per the multi-instrument alignment rule. Because the base template only submits the primary leg, the BTC hedge is submitted as a side effect in should_enter and closed in should_exit so both legs move together; equal 25%-equity notional per leg makes the position dollar-neutral with 50% gross exposure, so no leverage is needed and leverage stays 1.0 (the leverage_set_but_unused gate does not trigger). calculate_signal returns the continuous z-score in its natural units and the thresholds are compared in matching z-units (entry |z|>=2, exit through 0), never clipped. Venue is BINANCE USD-M futures because the strategy shorts one leg and needs a margin account; the ~0.10% round-trip fee on two legs is comfortably covered by 2-sigma ratio reversions.

The ETH/BTC ratio z-score mean-reversion premise is empirically refuted: the strategy is a consistent, catastrophic loser with Sharpe -1.97 (CI entirely negative), profit_factor 0.813, total_return -63.3%, expectancy -$66/trade, and a 65% max drawdown — losing money in EVERY one of the 7 years (2020 -23.7% through 2026 -2.4%) with rolling Sharpe pinned between -2 and -8. Judged on absolute metrics (benchmark_meaningful is correctly FALSE for a market-neutral pair). The code is implemented correctly (468 paired dollar-neutral entries, 0 dropped, cross-leg alignment barrier) and the sample is ample, so this is not sparsity or a bug — it is a sign-wrong edge. The symptom (consistent losses on both legs across all regimes from a fade-the-deviation rule) is consistent with the ETH/BTC log-ratio TRENDING rather than mean-reverting at the 30-bar/4H scale, so shorting the elevated ratio and longing the depressed ratio systematically fights a persistent trend (likely cause, stated as symptom not a verified data claim). Not optimize: no lookback/z-threshold tuning flips a systematically-losing fade with Sharpe -1.97 and every year negative into a winner. Not iterate: the code correctly implements the stated premise; the only 'fix' is to invert the signal into a momentum/trend-following trade, which is a different hypothesis, not a parameter change. Not worth 2 hours. FAILURE PATTERN: z-score mean-reversion on the ETH/BTC log-ratio at 4H produces regime-agnostic LOSSES (Sharpe -1.97, PF 0.81, -63% total, 65% DD, every year negative) because the ratio is non-stationary/trending at this scale rather than mean-reverting — fading 2-sigma deviations of a trending crypto-major spread is a systematically-losing trade that the dollar-neutral structure does not rescue; relative-value reversion requires a genuinely cointegrated/stationary spread, which ETH/BTC is not at a 30-bar 4H lookback.