EthBtcRatioConvergencePairsMarketNeutral4H

ETH/BTC Ratio Relative-Value Convergence, Market-Neutral Pairs (Long the Laggard / Short the Leader When the Log-Ratio Z-Score Deviates ≥2σ From Its Rolling Mean, USD-M Perps, 4h Bars, Adaptive-Mean + Hard Divergence Stop, 3-Parameter)

A MARKET-NEUTRAL statistical-arbitrage pairs trade on the two most cointegrated crypto majors: ETHUSDT.BINANCE and BTCUSDT.BINANCE (USD-M perps, 4h bars). Signal = the z-score of the log(ETH/BTC) price ratio versus its own rolling mean over a lookback window (default ~180 bars ≈ 30 days). When the ratio stretches ≥ entry_z (default 2.0) ABOVE its rolling mean (ETH richly outperforming BTC), SHORT ETH perp + LONG equal-notional BTC perp, betting on convergence; symmetric when it stretches ≥2σ BELOW (long ETH / short BTC). The position is delta-neutral in dollar terms (equal-notional legs) so it carries no crypto-beta — its P&L is the RELATIVE move of ETH vs BTC, not the market direction. Exit on convergence to |z| ≤ exit_z (default 0.5). This is a deliberately DIFFERENT family from the funding-carry basket that just collided (twice abandoned fee_edge): its edge is not a sub-basis-point funding cash flow that alt legs can't clear, but a multi-percent ratio reversion — ETH/BTC 2σ deviations are typically 2–4%, an order of magnitude above the ~0.20% two-leg USD-M round-trip (4 fills × ~0.05%), so it is structurally fee-viable where the N-leg carry baskets and cross-sectional reversal baskets (L12) are not. A ROLLING (adaptive) mean absorbs slow structural re-ratings of the ratio, and a hard divergence stop at |z| ≥ 3.5 exits when cointegration appears to be breaking rather than reverting, capping the classic stat-arb tail. Returns are earned in every regime the ratio oscillates (ranging and trend-then-reconverge), so there is no single-regime/2021 bull-spike concentration — the killer of every trend sibling this session — and no options trade-count wall.

Implements the ETH/BTC ratio convergence hypothesis as a delta-neutral two-leg USD-M pairs trade, using the proven multi-instrument pattern: ETH drives the signal, BTC is the synced extra leg, and the base class's same-timeframe alignment barrier guarantees the BTC bar for the current timestamp is present when calculate_signal runs, so the log-ratio buffer holds only contemporaneous pairs (avoiding the classic index-desync trap called out in the multi-instrument alignment rules). Fee viability (the explicit differentiator from the twice-abandoned funding-carry basket): a 2-sigma ETH/BTC deviation is typically 2-4%, an order of magnitude above the ~0.20% four-fill USD-M round-trip, so per-trade edge dwarfs costs. Lessons applied: (L4) both legs are the deepest majors with full 4h history, so no leg abandons on missing data, and Layer 2 feeds extras + calls calculate_signal directly (the z-score varies bar-to-bar, so no frozen-signal issue); (L21/L25) warmup is ~180 bars, a small fraction of the ~2190-bar 4h sandbox window, and 2-sigma ratio deviations recur many times per year, so >=1 trade is guaranteed without a compound gate that fires zero times; (L15) sizing is equity-relative each entry (non-compounding), gross capped at ~0.8x equity with a margin buffer, and a hard |z|>=3.5 divergence stop caps the classic stat-arb tail; the entry band [entry_z, hard_stop_z) additionally blocks entering or re-entering a spread whose cointegration appears to be breaking. Venue is BINANCE USD-M linear (NOT COIN-M inverse, avoiding the four-times-fatal inverse-contract engine defect): get_account_equity() returns a clean USDT figure and sizing is the ordinary qty = notional/price on both legs; leverage stays 1.0 (referenced in sizing, so no unused-leverage gate and no liquidation risk). Because returns are earned whenever the ratio oscillates and reconverges rather than from a directional beta, there is no single-regime/2021-bull concentration that killed the session's trend siblings, and being a single-pair two-leg hedge it is not a rank-and-rotate cross-sectional basket (L12). It fills the under-represented market-neutral and cross-sectional-relative-value buckets in a long-only-dominated book.

No edge to optimize: avg_trade_return_pct = -0.63% (NEGATIVE per-trade expectancy, below the 0.15% fee floor), Sharpe -0.32, PF 0.81, total_return -55.9%, max_drawdown 57.6% (>50% hard-abandon). The result is well-sampled (244 trades) and loses in 6 of 7 years across all regimes — this is a falsified premise, not a config problem: the log(ETH/BTC) ratio trends through 2σ deviations rather than reverting on the 4h/30d window (short win rate 0.36), so the convergence bet is systematically adverse. Optimization fits parameters to data that contains no profitable region; spending 2 hours will only fit noise. Abandon at review rather than optimize.

Market-neutral statistical-arbitrage pairs trade on ETHUSDT.BINANCE (primary) and BTCUSDT.BINANCE (extra leg), USD-M linear perps, 4h bars. The signal is the z-score of log(ETH/BTC) versus its own rolling mean over `lookback` bars (default 180 ~ 30 days), computed from a synced log-ratio buffer that appends only on timestamp-aligned bars so both legs are contemporaneous. When z >= entry_z (2.0) it shorts ETH and longs equal-dollar BTC (betting the ratio converges down); when z <= -entry_z it longs ETH and shorts BTC. Entries fire only in the band entry_z <= |z| < hard_stop_z so an already-broken spread is never entered. Exits on convergence (|z| <= exit_z, 0.5) or a divergence stop (|z| >= hard_stop_z, 3.5). Legs are equal dollar notional, so the book is delta-neutral (no crypto beta) and P&L is the relative ETH-vs-BTC move. An orphan guard flattens any single naked leg. Three edge parameters (lookback, entry_z, exit_z); linear equity sizing, leverage 1.0.

Edge is THIN but fee-viable — analyst's call on full history. Positives: avg_trade_return_pct 0.214% clears the 0.15% futures floor (fee-viability claim holds), delta-neutrality confirmed (beta 0.042), no liquidation, metrics_reliable=true. But realized edge is slim: Sharpe 0.06 (CI straddles zero), PF 1.02, +1.02% over 363 days, avg_loss $2,634 slightly > avg_win $2,428 so positive expectancy rests on win_rate 0.53. Not a defect — confirm it generalizes on the full backtest/walk-forward.

Long/short side asymmetry: short_win_rate 0.684 (short-ETH-when-rich worked) vs long_win_rate 0.368 (long-ETH-when-cheap didn't) over 19 trades each. Likely regime luck (ETH underperformed BTC this window), but if the long-ETH-cheap leg is persistently negative on full history it signals a directional ETH/BTC drift the symmetric mean-reversion doesn't capture. Analyst: confirm BOTH directions are profitable across regimes.

Capacity note (informational): impact_cost_pct 25.5%, capacity_usd ~$1.53M. Only $393 impact at the tested size, but the fee-viable edge could erode at scale — relevant at sizing/promotion, not correctness.

Clean, correct two-leg market-neutral implementation (equal-notional, orphan guard, contemporaneous synced ratio buffer)

Adequate trade sample (244 trades) — results are statistically measured, not noise

Genuinely different family from the fee_edge funding-carry siblings; not a fee-magnitude problem

avg_trade_return_pct = -0.63% — NEGATIVE per-trade expectancy, disqualifying regardless of Sharpe (L22)

Sharpe -0.32, profit_factor 0.81, total_return -55.9%, max_drawdown 57.6% (>50% hard-abandon per L19)

Losing in 6 of 7 years (2020,2021,2022,2023,2025,2026) — no regime where the edge exists, so it is not tunable

Premise falsified: ETH/BTC log-ratio trends through 2σ rather than reverting (short win rate 0.36) — mean-reversion convergence bet is systematically adverse

This strategy deliberately switched families away from the session's fee-dead funding carries, proposing a market-neutral ETH/BTC ratio-convergence pairs trade whose multi-percent reversions should have dwarfed the ~0.20% two-leg round-trip. The implementation was clean and well-sampled — 244 trades, equal-notional legs, orphan guard — so the loss was not a fee-magnitude or noise problem but a falsified premise: the log(ETH/BTC) ratio trended through 2σ deviations instead of reverting, so fading extremes was systematically adverse (short win rate 0.36) and the book lost money in 6 of 7 years, ending -55.9% with a 57.6% drawdown. With negative per-trade expectancy and no regime in which the edge existed, the analyst abandoned it at review on its first iteration rather than optimize parameters against data containing no profitable region.

A clean, fee-viable market-neutral pairs implementation is worthless if the underlying spread does not actually mean-revert — ETH/BTC log-ratio 2σ deviations trended through rather than converged on a 4h/30-day window, so cointegration/reversion must be empirically demonstrated before betting on convergence.

It was abandoned at the pre-optimization backtest-review gate on multiple hard failures — negative per-trade expectancy below the fee floor, negative total return, profit factor below 1.0, and a >50% hard-abandon drawdown — because its premise was falsified: the log(ETH/BTC) ratio trends through 2σ deviations rather than reverting on the 4h/30-day window, making the convergence bet systematically adverse with no profitable regime to tune toward.

A market-neutral statistical-arbitrage pairs trade on ETHUSDT.BINANCE and BTCUSDT.BINANCE USD-M perps (4h bars), fading ≥2σ z-score deviations of the log(ETH/BTC) ratio from its rolling mean — long the laggard, short the leader with equal-notional legs — betting on multi-percent ratio convergence with a hard divergence stop at |z|≥3.5.

The backtest was well-sampled but clearly losing: -55.9% total return over 244 trades, negative per-trade expectancy (avg_trade_return_pct -0.63%), Sharpe -0.32, profit factor 0.81, and a 57.6% max drawdown, with the short side winning only 0.36 of the time and losses in 6 of 7 years.