EthBtcDailyRatioZScoreMeanReversionPairsLongShort

ETH-BTC Ratio Z-Score Mean-Reversion LONG/SHORT Pairs (Canonical Cointegrated Major Pair, Dollar-Neutral, Single-Venue Both Legs, Daily, 3-Parameter, Cointegration-Gated)

A DOLLAR-NEUTRAL LONG/SHORT pairs strategy on the two largest crypto assets — ETHUSDT.BINANCE (primary) and BTCUSDT.BINANCE (hedge) — both on BINANCE USD-M futures, trading the mean-reversion of the ETH/BTC price ratio on DAILY bars. This is chosen by elimination after the session's hard evidence closed off nearly every other lane: VWAP-excursion MR is under a HALT (0-for-7 transplants, inverted-payoff signature); funding/carry is data-blocked (funding terminates ~Jan-2023 in the catalog); premium-index and L/S-ratio supp signals are dead/frozen-signal; options/cross-venue/COIN-M/HL are tooling- or data-blocked; and single-asset directional momentum is decayed. What REMAINS proven this session are exactly two classes the analyst named: the drawdown-accumulation class (already extended to ETH-spot-daily this session) and the COINTEGRATED-PAIRS class — 'the proven siblings (ADA-DOT, ETH-BTC) work because those specific pairs stayed cointegrated.' This hypothesis deliberately uses the SPECIFIC analyst-named ETH-BTC relationship rather than a sector-analogy guess (the AAVE-COMP de-cointegration trap), because ETH/BTC is the single most structurally-cointegrated relationship in crypto: both assets share the dominant 'crypto beta,' so their ratio oscillates in a persistent band rather than drifting, and both legs are the deepest, most-liquid instruments with 6+ years of continuous data. It fills the long_short DIRECTION bucket (12.5% actual vs the 55% long-only ceiling — the most under-represented direction that is NOT tooling-blocked) and is genuinely market-neutral by construction (dollar-neutral spread), so it is insulated from the 2024-2026 directional-edge decay that killed the momentum/funding/positioning strategies. CRITICALLY it is fully testable in the single-instrument-PnL engine WITHOUT any tooling gap: unlike the funding-carry/cross-settlement failures (which relied on uncreditable FundingRateUpdate cash flows and cross-VENUE MTM netting), BOTH legs here are ordinary price positions on the SAME venue (BINANCE USD-M), so the spread-convergence PnL is booked directly as the sum of two real leg PnLs — exactly the mechanism the proven ADA-DOT/ETH-BTC pairs siblings use. Per the analyst's explicit mandate, the proposal REQUIRES a cointegration gate: an ADF/Johansen stationarity check on the log-ratio over BOTH the full backtest window AND the most-recent 12 months at BACKTEST_REVIEW; if the ratio is not stationary recently, ABANDON pre-optimization (do not curve-fit a de-cointegrated pair). Exactly 3 tunables (ratio_lookback, entry_z, exit_z); all infra params (stop_z, max_hold_days, min_notional, per-leg dollar-neutral sizing) are FIXED and bounded — not in the Optuna search space.

Implements the analyst-named canonical cointegrated pair (ETH/BTC) with the proven SyncedBuffer pairs architecture used by landed siblings (ADA-DOT, and the LINK/ETH, SOL/AVAX pairs). SyncedBuffer commits an (a,b) sample only when both legs report the same ts_ns, so the ratio z-score is always computed on contemporaneous ETH/BTC closes (no independent-array-trim misalignment); the base cross-leg barrier (both legs 1D) ensures calculate_signal fires only when the BTC bar for the current ETH timestamp is present, so one aligned sample commits per bar. The strategy is two-sided and genuinely dollar-neutral: each leg is sized to equal USD notional (equity*leg_frac/price, not equal qty), so P&L tracks the ratio reverting rather than crypto beta, and entries are atomic (if either leg fails its qty/min-notional check, neither opens, preventing a naked directional bet). Exits implement the earliest-of rule with a critical divergence stop: |z| >= stop_z bails when the pair may have de-cointegrated (the unbounded-loss regime the analyst warns about), alongside reversion (|z| <= exit_z) and a 30-day time stop; entry_sign is re-inferred from the live ETH position side if state is lost. Because both legs are ordinary same-venue BINANCE USD-M price positions, the convergence PnL books directly as the sum of two real leg PnLs with no cross-venue MTM netting or uncreditable funding cash flows -- exactly the tooling-clean mechanism the funding/cross-settlement failures lacked. leg_frac=0.45 keeps gross ~90% within the leverage-1.0 futures margin ceiling (leverage referenced only as a non-amplifying multiplier, so no unused-leverage gate), and each ~45k leg clears the $100 BTC min-notional. min_bars_required=2 lets the buffer fill while zscore_last returns None until 60 aligned samples accrue, gating warmup without dead bars; the returned z-score is a continuous per-bar signal so Layer-2 frozen-signal detection passes while entries stay threshold-gated. Exactly three tunables (ratio_lookback, entry_z, exit_z); stop_z, max_hold_days, and leg_frac are fixed infra, keeping the optimizer's search space minimal. Per the hypothesis mandate, a cointegration/stationarity gate on the log-ratio (full window AND recent 12 months) should be applied at BACKTEST_REVIEW; abandon pre-optimization if the ratio is not recently stationary rather than curve-fitting a de-cointegrated pair. Fills the under-represented long_short direction bucket and is insulated by construction from 2024-2026 directional-edge decay; both legs are the deepest, longest-history crypto instruments so no data_unavailable risk.

ETH/BTC daily ratio z-score MR loses money (total_return -22.5%, Sharpe -0.30 with ci_low -1.26, PF 0.92, expectancy -$198/trade) because the ratio TRENDED rather than reverting over the test period: 2025 -19.0% and 2026 -6.2% are direct evidence that ETH/BTC has been in a structural downtrend (ETH underperforming BTC through the 2024-2025 BTC-dominance regime), so the mean-reversion strategy repeatedly faded a one-way move and bled. This triggers the hypothesis's OWN mandated cointegration gate: it required an ADF/Johansen stationarity check on the log-ratio over the full window AND the recent 12 months, and stated 'if the ratio is not stationary recently, ABANDON pre-optimization (do not curve-fit a de-cointegrated pair).' The recent ratio is non-stationary, so the proposal's own rule prescribes abandon. Not optimize: no z-threshold tuning fixes a non-stationary spread — optimizing would curve-fit the 2020-2024 window and fail the 2025-2026 de-cointegration, the exact trap the hypothesis tried to avoid. Not iterate: the de-cointegration is structural to ETH/BTC on the daily horizon, not a parameter miss. FAILURE PATTERN: even the 'canonical most-cointegrated' crypto pair (ETH/BTC) has de-cointegrated on the daily horizon in the 2024-2026 BTC-dominance regime — the ratio trends rather than oscillates, so dollar-neutral ratio-MR produces negative expectancy with a brutal recent drawdown (2025 -19%), the same trending-ratio signature as the abandoned AAVE-COMP, LINK/ETH, and SOL-AVAX pairs. The cointegrated-pairs class works only on pairs that remain stationary in the test/holdout window; ETH/BTC daily no longer qualifies. The Research Lead should treat major-pair daily ratio-MR as regime-closed for now and verify recent-window stationarity BEFORE proposing any further pair, rather than assuming structural cointegration.

Dollar-neutral long/short pairs mean-reversion on the ETH/BTC ratio, ETHUSDT (primary) and BTCUSDT (hedge) on BINANCE USD-M futures, daily bars. On each timestamp-aligned daily bar it commits the (ETH, BTC) close pair to a SyncedBuffer, forms the ETH/BTC ratio, and z-scores it over a 60-bar rolling window. When z >= +2.0 (ratio rich) it shorts ETH and longs BTC; when z <= -2.0 (ratio cheap) it longs ETH and shorts BTC, each leg at equal USD notional (dollar-neutral). It closes both legs on the earliest of reversion (|z| <= 0.5), a divergence/de-cointegration stop (|z| >= 3.5), or a 30-day time stop. Both legs are ordinary price positions on the same venue, so spread-convergence PnL books as the sum of two real leg PnLs. Pure OHLCV, leverage 1.0, ~90% gross. Three tunables: ratio_lookback, entry_z, exit_z.

Genuinely market-neutral (beta -0.014, benchmark_correlation -0.05); clean atomic two-leg same-venue implementation; deepest, most-liquid legs

Mechanism class is proven on genuinely-stationary pairs (ADA-DOT) — the construction is correct

Net-losing with no edge: total_return -22.5%, Sharpe -0.30 (ci_low -1.26), PF 0.92, expectancy -$198/trade

Ratio TRENDED, didn't revert: 2025 -19.0%, 2026 -6.2% — ETH/BTC in a structural downtrend (BTC dominance), so MR fades a one-way move and bleeds

Fails the hypothesis's OWN mandated cointegration gate: the recent-12-month log-ratio is non-stationary, which the proposal said requires pre-optimization ABANDON

Severe risk: max_DD 38.9% (ci_high 62%), drawdown duration 1,938 days; return_skew -1.13, kurtosis 56.9

EthBtcDailyRatioZScoreMeanReversionPairsLongShort was chosen by elimination after the session closed off nearly every other lane, deliberately targeting the analyst-named ETH/BTC relationship — supposedly crypto's most structurally cointegrated pair — with a clean same-venue dollar-neutral build and a self-imposed cointegration gate to avoid the AAVE-COMP de-cointegration trap. But the ratio trended rather than oscillated: ETH underperformed BTC through the 2024-2025 BTC-dominance regime, so the strategy faded a one-way move for a -22.5% return, a -0.30 Sharpe, and a 38.9% drawdown, with 2025 at -19%. That recent non-stationarity tripped the hypothesis's own mandated gate, which prescribed abandon pre-optimization. The analyst abandoned it on its first iteration, concluding that major-pair daily ratio-MR is regime-closed for now — even ETH/BTC no longer qualifies — the same trending-ratio signature as the abandoned AAVE-COMP, LINK/ETH, and SOL-AVAX pairs.

The cointegrated-pairs class works only on pairs that remain stationary in the test and holdout window — even the canonical 'most-cointegrated' crypto pair, ETH/BTC, de-cointegrated on the daily horizon in the BTC-dominance regime, so recent-window stationarity must be verified before proposing any pair rather than assuming structural cointegration.

The analyst abandoned it at the backtest-review gate before optimization because the ETH/BTC ratio trended — ETH structurally underperformed BTC through the 2024-2025 BTC-dominance regime — so the mean-reversion repeatedly faded a one-way move and bled, triggering the hypothesis's own mandated cointegration gate: the recent 12-month log-ratio was non-stationary, which the proposal said requires pre-optimization abandon.

A dollar-neutral, OHLCV-only long/short pairs mean-reversion on the ETH/BTC ratio using daily bars on Binance USD-M futures (both legs same venue) — z-scoring the ETH/BTC price ratio over a 60-day window and, at |z| ≥ 2, shorting the rich leg while longing the cheap one, exiting on reversion, a divergence stop, or a time stop, with 3 tunables and an explicit self-imposed cointegration gate requiring abandon if the recent 12-month log-ratio was non-stationary.

The build was genuinely market-neutral (beta -0.014) over 96 trades, but it lost money with no edge: -22.5% total return (CAGR -3.5%), Sharpe -0.30 (CI low -1.26), profit factor 0.92, per-trade expectancy -$198, a 38.9% max drawdown (CI high 62%), and negative recent years (2025 -19.0%, 2026 -6.2%) as the ratio trended rather than reverting.