EthBtcRatioSpreadMeanReversionMarketNeutralLS

ETH-BTC Ratio Spread Mean-Reversion (Carry-Free Relative-Value, Market-Neutral Long-Short, Binance USD-M, 4H, De-Cointegration Stopped)

A MARKET-NEUTRAL, two-leg RELATIVE-VALUE spread mean-reversion on the ETH/BTC ratio, both traded as Binance USD-M perpetuals (ETHUSDT.BINANCE primary, BTCUSDT.BINANCE hedge). This is a deliberate pivot AWAY from the directional-momentum family that the analyst has now declared repeatedly dead this regime (cross-sectional AND absolute trend on crypto majors both abandoned), toward the explicitly-recommended alternative: a carry-free relative-value structure that profits from REVERSION rather than trend. The mechanism: ETH and BTC are the two most economically-tethered, highest-correlation, longest-listed crypto assets; their log-price spread (ETH vs a rolling hedge-ratio * BTC) oscillates around a slowly-moving mean. When the spread stretches to a statistical extreme, bet on reversion; when it normalizes, exit. CRITICAL DESIGN NOTES separating this from every closed family: (1) it is NOT same-asset basis (ETH and BTC are DIFFERENT assets, so the two legs carry genuine, non-cancelling relative price PnL the engine credits — unlike the delta-neutral same-asset basis/funding pairs that cancel to ~0 and lose on fees); (2) it is NOT cross-venue / HYPERLIQUID — both legs are Binance USD-M with 6+ years of deep history, so walk-forward has real statistical power (no 2.3-year / 7-month data wall); (3) it is NOT funding/basis CARRY — funding is never relied on as a cash flow; (4) it is NON-momentum — it profits precisely from the tightened-correlation chop regime that killed the momentum family (tighter correlation = harder spread reversion); (5) the spread crosses its bands many times per year across all regimes, so walk-forward OOS windows are densely populated — the opposite of the OOS-all-zero event-sparsity that killed the drawdown-accumulation class. Intentionally minimal: 3 core tunables (z_lookback, z_entry, z_exit) plus a fixed de-cointegration blowout stop, to resist overfitting.

Implements the analyst-recommended carry-free relative-value pivot using the proven SyncedBuffer pairs architecture (as in the landed SOL/AVAX and LINK/ETH pairs): SyncedBuffer commits an (ETH, BTC) sample only when both legs report the same ts_ns, and the base cross-leg barrier (both 4H) fires calculate_signal only when the BTC bar for the current ETH timestamp is present, so the rolling OLS hedge ratio and z-score are computed on contemporaneous prices with no stale-leg lag. The spread is the hypothesis's log(ETH)-beta*log(BTC) with beta a rolling OLS slope (clamped to [0.3,2.0] for stability), and the BTC leg is sized at beta x the ETH leg so the position is neutral to the common BTC/ETH beta and isolates the relative ETH/BTC mispricing -- genuine non-cancelling PnL, distinct from a delta-neutral same-asset basis that nets to ~0 and bleeds fees. Entries are atomic (both legs viable or neither, so the pair is never one-legged into a naked directional bet) and dollar-sized off equity*leg_frac with a 95% gross cap, keeping the book within the leverage-1.0 futures margin ceiling (leverage referenced only as a non-amplifying multiplier, so no unused-leverage gate). Exits close both legs on reversion (|z|<=z_exit) or a fixed de-cointegration blowout stop (|z|>=z_stop) -- the essential pairs risk control that cuts the trade when the relationship structurally breaks rather than riding it to ruin. min_bars_required=2 lets the buffer fill immediately while _spread_z returns None until z_lookback aligned samples accrue, gating warmup without dead bars; the returned z is a continuous per-bar signal so Layer-2 frozen-signal detection passes while entries stay threshold-gated. Both legs are the deepest, longest-listed Binance USD-M majors (6+ years 4H history) so walk-forward has real power and no data wall; should_enter/should_exit/position_size are inert because the base entry path manages only the primary instrument.

The ETH/BTC ratio-spread mean-reversion premise is falsified by its own backtest: the spread TRENDS, it does not oscillate around a stationary mean. Over a dense 564-trade sample (6.5 years) it loses -78.1% (CAGR -21.9%) with Sharpe -1.86 and the ENTIRE CI below zero (-2.45 to -1.18), PF 0.71, probabilistic_sharpe 0.0, a 79% max drawdown, and losses in 6 of 7 years (only a flat 2026). The payoff geometry is a falling knife: return_skew -5.94, kurtosis 64, avg_win $710 < avg_loss $933 — fading 'rich' ETH (short ETH/long BTC) and 'cheap' ETH (long ETH/short BTC) gets run over because ETH/BTC trends persistently (ETH outperformed BTC into 2021, then declined versus BTC through 2022-2025), so the de-cointegration blowout stops simply realize the trend losses. This is a premise problem, not a tuning problem: the ETH/BTC relationship is NOT cointegrated, so no z_lookback/z_entry/z_exit/z_stop setting converts a reversion strategy on a trending spread into a winner, and a best-of-225 sweep would fit noise against an all-years-negative -78% base. FAILURE PATTERN: pairs/ratio mean-reversion on crypto majors is a dead family because the majors trend and de-cointegrate rather than oscillate around a stationary ratio — the ETH/BTC ratio is a persistent trend, not a mean-reverting spread, so fading its z-extremes catches falling knives (negative-skew, large-loss tails) and bleeds to a -78% catastrophe; a 'de-cointegration stop' cannot rescue a relationship that was never cointegrated. The hypothesis's own framing ('toward the explicitly-recommended relative-value alternative') is the trap: relative-value reversion presumes cointegration, which these assets lack.

Market-neutral relative-value mean-reversion on the ETH/BTC log-spread, both legs Binance USD-M perpetuals (ETHUSDT primary, BTCUSDT hedge), 4-hour bars. Each aligned bar it fits a rolling OLS hedge ratio beta of log(ETH) on log(BTC) over z_lookback samples, forms the spread log(ETH)-beta*log(BTC), and z-scores it. When z >= +2 (ETH rich) it shorts ETH and longs BTC; when z <= -2 (ETH cheap) it longs ETH and shorts BTC, sizing the BTC leg at beta x the ETH leg so the book is neutral to common beta. It closes both legs when |z| reverts to <= 0.5 or on a de-cointegration blowout stop at |z| >= 4. Pure OHLCV, leverage 1.0, gross capped at 95% of equity. Three core tunables plus a fixed stop.

Clean two-leg market-neutral construction (atomic both-legs-or-none, rolling OLS hedge ratio, de-cointegration stop), genuinely non-cancelling relative PnL, deep full-history data, low beta (-0.05)

Dense 564-trade sample on confirmed-data majors — not a sparsity/data-wall artifact

Catastrophic: total return -78.1%, CAGR -21.9%, Sharpe -1.86 with entire CI below zero (-2.45 to -1.18), PF 0.71, max DD 79%, probabilistic_sharpe 0.0

Negative in 6 of 7 years (only a flat 2026); the ratio trends rather than mean-reverts

Falling-knife payoff: return_skew -5.94, kurtosis 64, avg_win $710 < avg_loss $933 — de-cointegration stops realize large trend losses

Premise falsified: ETH/BTC is NOT cointegrated/oscillating; it trends persistently, so fading the spread is the dead mean-reversion family

>0 to be viable

negative 6/7 years

majority positive

EthBtcRatioSpreadMeanReversionMarketNeutralLS pivoted away from the dead momentum family toward the explicitly-recommended relative-value alternative, fading statistical extremes in the ETH/BTC log-spread as a market-neutral, carry-free pair on deep full-history Binance data. The construction was clean — atomic two-leg execution, rolling OLS hedge ratio, de-cointegration stop, genuinely non-cancelling relative PnL — but the result was a -78% catastrophe with a Sharpe whose entire CI sat below zero, a 79% drawdown, and losses in six of seven years. The analyst abandoned it at the backtest-review gate on its first iteration, diagnosing a premise failure: ETH/BTC trends persistently (ETH outperforming BTC into 2021, then declining versus BTC through 2022-2025) rather than mean-reverting, so fading the spread is a falling-knife trade no parameter can rescue. The recorded conclusion was that the hypothesis's own framing — assuming relative-value reversion implies cointegration — was the trap, since these majors lack the stationary ratio the mechanism requires.

Pairs/ratio mean-reversion on crypto majors is a dead family because the majors trend and de-cointegrate rather than oscillate around a stationary ratio — relative-value reversion presumes cointegration, which ETH/BTC lacks, so a 'de-cointegration stop' cannot rescue a relationship that was never cointegrated, and the genuinely non-cancelling two-leg construction is precisely what makes the trend losses real.

The analyst issued an 'abandon' verdict at the pre-optimization backtest-review gate, so optimization, analysis, and risk review never ran: the premise was falsified by its own backtest — the ETH/BTC ratio TRENDS persistently rather than oscillating around a stationary mean, so fading its z-extremes catches falling knives, and no z-parameter setting converts a reversion strategy on a trending, non-cointegrated spread into a winner.

A market-neutral, two-leg relative-value spread mean-reversion on the ETH/BTC ratio (ETHUSDT primary, BTCUSDT hedge, both Binance USD-M 4H perps) that fit a rolling OLS hedge ratio, z-scored the log-spread, and bet on reversion — shorting ETH/longing BTC when the spread stretched rich and the reverse when cheap, exiting on normalization or a fixed de-cointegration blowout stop — a deliberate pivot away from the declared-dead directional-momentum family toward a carry-free relative-value structure.

On a dense 564-trade sample over 6.5 years it failed catastrophically: total return -78.1%, CAGR -21.9%, Sharpe -1.86 with the entire CI below zero (-2.45 to -1.18), profit factor 0.71, probabilistic Sharpe 0.0, 79% max drawdown, negative in 6 of 7 years, with a falling-knife payoff geometry (return skew -5.94, kurtosis 64, avg win $710 < avg loss $933) as the de-cointegration stops realized large trend losses.