EthBtcRatioZScoreMeanReversionLongSpreadPair4H

ETH/BTC Ratio Z-Score Mean Reversion Long-Spread Pair (4H)

A DOLLAR-NEUTRAL RELATIVE-VALUE PAIR strategy on Binance USD-M perpetuals: simultaneously LONG ETHUSDT and SHORT BTCUSDT in equal dollar notional, triggered by extreme negative z-scores of the ETH/BTC ratio. CRITICAL DESIGN CHOICES driven DIRECTLY by the session's decisive failure pattern: (a) Long-only OHLCV breakout/momentum on alt-perps has DEFINITIVELY regime-decayed in 2025+ — 4 consecutive abandonments (BTC WRB, ADA Donchian, AVAX Keltner, LTC MACD). The analyst's explicit recommendation: 'REDIRECT new hypothesis generation toward mechanisms that worked through 2025-2026: shorter-timeframe mean reversion, options-based delta-neutral strategies, or cross-asset relative-value pairs — fundamentally different mechanisms that don't depend on directional follow-through.' This proposal implements that exact recommendation. (b) DOLLAR-NEUTRAL by construction — eliminates beta-to-crypto exposure that drove the failed momentum strategies into the regime-decay trap. The trade makes money on RELATIVE movement, not absolute direction. (c) MEAN-REVERSION on a STATIONARY SPREAD — unlike single-asset MR (which failed because absolute crypto prices are non-stationary trending), the ETH/BTC ratio has been historically stationary with strong mean-reversion characteristics across all regimes including 2025-2026 (ETH/BTC has oscillated in a ~0.04-0.08 range for years). (d) Uses ONLY the most liquid pair in crypto — ETH and BTC have the deepest order books, narrowest spreads, and most reliable execution. (e) MAXIMALLY SIMPLE — only 2 decision parameters (z-score lookback, entry threshold). (f) ZERO supplementary data dependency.

Implements the analyst's explicit redirect toward cross-asset relative-value pairs after four long-only momentum/breakout abandonments. Uses the repo's SyncedBuffer helper (the same primitive behind landed pairs strategies) so the ETH/BTC ratio is always computed on contemporaneous closes -- it commits an (a,b) sample only when both legs report the same ts_ns, eliminating the independent-array-trim alignment bug. The base cross-leg barrier (both legs 4-HOUR) ensures calculate_signal fires only when the BTC bar for the current ETH timestamp is present, so the buffer commits one aligned sample per bar; min_bars_required is 2 so the buffer starts filling immediately and zscore_last returns None until lookback_bars aligned samples accrue, gating warmup without dead bars. The strategy is one-sided per the hypothesis (long-spread only): it enters LONG ETH / SHORT BTC exactly when z <= -entry_z (ratio cheap) and exits both legs when z reverts to >= exit_z or the time stop fires. Entry is dollar-neutral (each leg sized to equal USD notional = equity*leg_frac/price, not equal qty) which makes the position delta-neutral so P&L tracks ratio reversion rather than either coin's direction, and it is atomic -- if either leg fails its qty/min-notional check, neither leg opens, preventing a naked directional bet. leg_frac=0.45 keeps gross ~90% within the leverage-1.0 futures margin ceiling (leverage not set >1, so the unused-leverage gate does not apply). All order management is centralized in calculate_signal with the should_enter/should_exit/position_size hooks inert because the base entry path manages only the primary instrument. The returned z-score is a genuinely continuous per-bar signal so Layer-2 frozen-signal detection passes while entries remain threshold-gated. ETH and BTC are the two most liquid crypto perps with full multi-year 4H history, so no leg risks data_unavailable, and only two decision parameters keep the optimization surface minimal per the low-parameter guidance.

Do not optimize — the strategy is a consistent money-loser whose core premise is falsified by its own backtest. Over 126 trades it returns -23.5% with profit_factor 0.90, Sharpe -0.63 (sharpe_ci_low -1.62), omega 0.71, Sortino -0.53, expectancy -$198/trade, and negative annual returns in 5 of 7 years (2020 -9%, 2022 -2%, 2023 -10%, 2025 -8%, 2026 -6%). The hypothesis bets that the ETH/BTC ratio is stationary and mean-reverting (~0.04-0.08), but the loss pattern — a one-sided long-spread book with long_win_rate 0.65 yet a negative overall result and a fat left tail (return_skew -3.5, largest_loss -$21k vs largest_win $16.6k) — is the signature of a NON-stationary, trending spread: ETH structurally underperformed BTC across 2021-2025, so 'buy ETH / short BTC when the ratio is cheap' repeatedly caught a falling knife. The one-sided design compounds this: trading only the long-spread on negative z can profit only if the ratio rises, with no defense against the secular downtrend. This is a clean, correctly-sized, dollar-neutral build (beta 0.03, exposure 65%, no blow-up) and trade count is healthy, so it is NOT a sizing bug or sparsity — it is a genuine absence of edge because the spread is not mean-reverting over the test period, and no change to lookback or entry_z makes a trending ratio revert. The pairs-MR mechanism may work on a genuinely cointegrated pair, but ETH/BTC over this window is not one; this run should be abandoned rather than optimized (and a future attempt would need a cointegration-tested, two-sided spread, not a one-sided long-only ETH/BTC bet).

Dollar-neutral ETH/BTC ratio z-score mean-reversion pair on Binance USD-M futures, 4H bars, long-spread only. On each timestamp-aligned 4H bar it commits the (ETH, BTC) close pair to a SyncedBuffer, computes the ETH/BTC ratio, and z-scores it over a rolling lookback window. When the z-score is <= -entry_z (ratio unusually cheap) it goes LONG ETHUSDT and SHORT BTCUSDT at equal USD notional (delta-neutral), and closes both legs when the z-score reverts to >= exit_z (~0) or a time stop fires. The trade profits from relative ETH-vs-BTC movement, not market direction, so it carries no crypto beta. Pure OHLCV, no supplementary data, leverage 1.0 with ~90% gross (45% per leg). Two decision parameters: lookback_bars and entry_z.

Clean dollar-neutral construction (beta 0.03, exposure 65%, avg_position 47%) with NO sizing/exposure bug — a correctly-built pair, unlike the recent multi-instrument blow-ups

Most-liquid pair, healthy frequency (126 trades), pure OHLCV

Consistent money-loser: total_return -23.5%, Sharpe -0.63 (sharpe_ci_low -1.62), profit_factor 0.90, omega 0.71, Sortino -0.53, expectancy -$198/trade; negative in 5 of 7 years

Premise falsified: the ETH/BTC ratio is NOT stationary over the test period — it trended down (ETH structurally underperformed BTC), so z-score mean-reversion keeps buying a declining spread (return_skew -3.5, fat left tail, largest_loss -$21k)

One-sided long-spread design can only profit if the ratio rises — no defense against the secular ETH/BTC downtrend

No parameter change (lookback, entry_z) makes a trending/non-stationary spread mean-revert

EthBtcRatioZScoreMeanReversionLongSpreadPair4H was an explicit implementation of the analyst's recommendation to pivot away from regime-decayed long-only momentum toward dollar-neutral relative-value, betting that the historically range-bound ETH/BTC ratio would mean-revert. The execution was sound — a correctly-sized, beta-neutral pair with 126 trades and no sizing bug — but it lost 23.5% with a profit factor of 0.90 and a fat left tail, because the ratio was non-stationary and trended down as ETH underperformed BTC. The one-sided long-spread design could only profit if the ratio rose, leaving it defenseless against the secular downtrend. The analyst abandoned the run at the backtest-review gate, noting the mechanism might work on a genuinely cointegration-tested, two-sided spread but not on this one-sided ETH/BTC bet.

A pairs mean-reversion strategy requires a genuinely cointegrated, stationary spread, and a one-sided long-spread design has no defense against a secular downtrend — even a clean dollar-neutral build that removes crypto beta cannot profit when the chosen ratio trends rather than reverts, so cointegration must be tested and the spread traded two-sided.

The analyst abandoned it at the backtest-review gate before optimization because its core premise was falsified by its own data: the ETH/BTC ratio was not stationary over the test period — it trended down as ETH structurally underperformed BTC across 2021-2025 — so the one-sided 'buy the cheap spread' logic repeatedly caught a falling knife, with no parameter change able to make a trending spread mean-revert.

A dollar-neutral, OHLCV-only relative-value pair on Binance USD-M 4H perps — going long ETHUSDT and short BTCUSDT in equal dollar notional when the ETH/BTC ratio's z-score fell to an extreme negative (the ratio looked unusually cheap), exiting on reversion to the mean or a time stop, as a one-sided long-spread bet on the ETH/BTC ratio's supposed stationarity, with just two decision parameters.

The build was clean and correctly sized (beta 0.03, exposure 65%, no blow-up) over a healthy 126 trades (63 long, 63 short), but it lost money: -23.5% total return (CAGR -4.5%), Sharpe -0.63 (CI low -1.62), profit factor 0.90, per-trade expectancy -$198, a fat left tail (skew -3.5, largest loss -$21k vs largest win $16.6k), and negative returns in 5 of 7 years.