BtcEthDailyEngleGrangerCointegrationResidualMRPairs1D

BTC-ETH Daily Engle-Granger Cointegration Residual Mean Reversion Pairs Trade with 90-Day Rolling Window

A dollar-neutral, daily-bar pairs trading strategy on BTCUSDT.BINANCE and ETHUSDT.BINANCE perpetual futures using PROPER ENGLE-GRANGER COINTEGRATION (not the failed 4H log-ratio z-score). This directly addresses the analyst's explicit recommendation from a prior failure log entry: 'Daily-bar pairs trades with longer lookbacks (60-120 days) have shown more durable cointegration in academic crypto literature than 4H/30-bar versions. A future hypothesis proposing daily-bar pairs with cointegration-based signals (Engle-Granger residual mean reversion) would be structurally distinct and worth a fresh attempt.' The mechanism (Engle 1987, Engle-Granger 1987 Nobel-prize-cited): (1) on each daily bar close, run a rolling 90-day OLS regression of ETH_close vs BTC_close to estimate the time-varying hedge ratio β; (2) compute the RESIDUAL series ε[t] = ETH[t] - β × BTC[t] over the last 90 days; (3) z-score the latest residual against its 90-day mean and std; (4) when z > +2 (ETH richly valued vs BTC), go SHORT ETH + LONG BTC (sized by current β to be dollar-neutral on the regression-adjusted spread); when z < -2, go LONG ETH + SHORT BTC. Hold until |z| < 0.5 (reversion confirmed) or max 21 days. Critically different from the failed 4H ratio z-score: (a) Engle-Granger ESTIMATES the hedge ratio dynamically rather than assuming a fixed 1:1 ratio — this handles the documented 2022-2025 ETH-BTC decoupling because β itself rebases as cointegration weakens; (b) DAILY bars give the 90-day rolling regression sufficient observations (90 vs 30 in the 4H/30-bar failure) for statistical power; (c) the longer cointegration window matches the timescale at which crypto majors actually mean-revert (multi-week, not multi-hour). Uses ONLY BINANCE USD-M perp daily bars — no supplementary data, no spot, no options. Avoids EVERY confirmed failure mode. Fills FOUR under-represented buckets: long-short direction (14.6% → toward 45%), pairs scope (extends 18.6% pairs category), proven-academic mechanism, and OHLCV-only data dependency. Position sizing: 15% per leg, with the leg sizes adjusted by β to be dollar-neutral on the spread (not on the individual instruments). Risk per trade: 1.5% of equity capped by hard stop.

Implements the hypothesis's Engle-Granger residual mean-reversion exactly and fixes the prior 4H log-ratio failure mode. (1) PROPER COINTEGRATION: instead of assuming a 1:1 log-ratio, it regresses ETH price level on BTC price level via OLS every bar to estimate a dynamic beta+alpha and z-scores the regression RESIDUAL — so beta rebases automatically as ETH-BTC cointegration strengthens/weakens (handles 2022-2025 decoupling). (2) DAILY bars with a 90-day window give 90 regression observations for statistical power vs the failed 30-bar/4H version, matching the multi-week timescale at which crypto majors mean-revert. (3) DOLLAR-NEUTRAL ON THE SPREAD: ETH leg = 15% of equity notional, BTC coins = beta x ETH coins, so the book's P&L is q_eth*d(eps) — neutral to the common BTC factor, not equal-dollar per instrument. Because beta*(P_btc/P_eth) ~ 1 for ETH-BTC, the beta-scaled BTC notional naturally lands near the ETH notional (no runaway), and a beta band guard (0<beta<5) blocks degenerate hedges. (4) RISK: reversion exit at |z|<0.5, time stop at 21 days, and a hard z_stop=4.5 divergence stop serving as the cointegration-break / 1.5%-risk cap. (5) FEE-VIABLE: 2-sigma residual entries with multi-week holds target spread moves far above the ~0.10% round-trip cost per leg; expected per-trade edge clears futures fees. (6) VENUE: futures venue is required because the strategy goes short on one leg every trade; leverage=1.0 since gross exposure is ~30% and dollar-neutral. (7) DATA: BTCUSDT/ETHUSDT.BINANCE daily are the most liquid, fully-backfilled USD-M perps — no data_unavailable risk. Implementation reuses the proven ADADOTPairsSpreadMR multi-leg execution skeleton (SyncedBuffer alignment, base-class on_bar contract, hedge leg driven via _submit_entry/_submit_exit overrides) to avoid alignment and lifecycle bugs.

Correctly-implemented pairs trade with a negative, recently-decayed edge — not a coding or tuning problem. The strategy is genuinely dollar-neutral (beta -0.01, vol 7.6%) and trades exactly as hypothesized (50 balanced long+short Engle-Granger entries, dynamic beta hedge), but the BTC-ETH residual does not mean-revert profitably: profit_factor 0.84, Sharpe -0.60, expectancy -$139/trade, total_return -15.7%, win_rate exactly 0.50. Decisively, the losses concentrate in the recent regime (2021 -9.1%, 2025 -9.7%, 2026 -1.4%), confirming the ETH-BTC cointegration the hypothesis relies on has weakened over 2022-2025 — the dynamic-beta fix the hypothesis bet on did not restore an edge. With only 50 pair entries over 6 years against 6 tunable parameters, optimization would fish a noise corner that fails walk-forward/holdout (a 15-day holdout on the negative recent window would fail), and the fat negative tail (skew -8.76, kurtosis 122, largest_loss -$7,979) is a cointegration-break risk no threshold tuning removes. FAILURE PATTERN: BTC-ETH daily Engle-Granger residual mean-reversion has negative edge over 2020-2026 because the pair has de-cointegrated in the recent regime — even proper dynamic-beta cointegration (the structurally-distinct upgrade over the failed 4H log-ratio) produces PF<1 with the worst losses in the most recent years, and 50 entries is too sparse to optimize without overfitting.

Dollar-neutral daily-bar pairs trade on ETHUSDT.BINANCE (primary) and BTCUSDT.BINANCE (hedge) USD-M perpetual futures using proper Engle-Granger cointegration. Each daily bar runs a rolling 90-day OLS regression of ETH_close on BTC_close to estimate a time-varying hedge ratio beta and intercept alpha, forms the residual series eps = ETH - (alpha + beta*BTC), and z-scores the latest residual against the window. Enters SHORT ETH + LONG BTC when z >= +2 (ETH rich), LONG ETH + SHORT BTC when z <= -2 (ETH cheap). The BTC leg is sized at beta x ETH_coins so the combined position is dollar-neutral on the regression spread (P&L tracks the residual, neutral to the common BTC factor). Exits when |z| reverts inside 0.5, the spread diverges past a hard z_stop of 4.5 (cointegration-break / risk cap), or after 21 days. Pure OHLCV, no supplementary data. Uses SyncedBuffer to keep the two legs timestamp-aligned, and the base class's cross-leg alignment barrier so the signal uses contemporaneous ETH/BTC closes.

Correctly implemented: genuinely dollar-neutral (beta -0.01, benchmark_correlation -0.15, annualized vol 7.6%), 50 balanced long+short pair entries, dynamic OLS hedge ratio, clean entry gate (0 dropped)

Trades match the hypothesis mechanism exactly — Engle-Granger residual z-score entries, β-sized BTC hedge leg

Low fee drag (commission_pct_of_gross 2.04%) — costs are not the cause of the loss

Negative edge: profit_factor 0.84, Sharpe -0.60, expectancy -$139/trade, total_return -15.7%, win_rate exactly 0.50 (coin flip)

Premise has decayed: the worst years are the recent ones (2021 -9.1%, 2025 -9.7%, 2026 -1.4%) — BTC-ETH cointegration is weakest in the current regime, the opposite of what a deployable edge needs

Fat negative tail: return_skew -8.76, kurtosis 122, largest_loss -$7,979 vs avg_win $1,482 — divergence-stop losses dwarf the average win

Thin sample (50 pair entries over 6 years) vs 6 tunable parameters — optimization would overfit noise and fail walk-forward/holdout

Dynamic beta was the hypothesis's explicit fix for ETH-BTC decoupling, and the backtest shows it did not produce a positive edge

This strategy was the deliberate, structurally-distinct retry of a failed 4H ratio pairs trade: a dollar-neutral BTC-ETH daily Engle-Granger residual mean-reversion with a rolling 90-day OLS hedge ratio. It was implemented exactly as designed — genuinely market-neutral, balanced long/short, low fee drag — but had a negative edge (profit factor 0.84, Sharpe -0.60, -15.7%, coin-flip win rate) with a fat negative tail and losses concentrated in the most recent years. The analyst abandoned it at the backtest-review gate on the first iteration, concluding the pair has de-cointegrated in the current regime so even the dynamic-β fix produced PF<1, and that the thin 50-entry sample against 6 parameters would overfit — so it never advanced to optimization or risk review.

A proper, academically-grounded upgrade (dynamic-β Engle-Granger cointegration over a fixed log-ratio) still fails when the underlying relationship has decayed — BTC-ETH cointegration is weakest in the recent regime, exactly where a deployable edge is needed — and 50 pair entries against 6 parameters is too sparse to optimize without curve-fitting; cointegration-break tail risk (skew -8.76) cannot be tuned away.

The backtest-review analyst issued an 'abandon' verdict: the trade was correctly implemented and genuinely market-neutral, so this is a negative, recently-decayed edge rather than a bug — the BTC-ETH residual does not mean-revert profitably and the pair has de-cointegrated over 2022-2025, so the dynamic-β fix the hypothesis bet on did not restore an edge, while 50 entries against 6 tunable parameters would only overfit and fail holdout.

A dollar-neutral daily-bar pairs trade on BTCUSDT/ETHUSDT Binance futures using proper Engle-Granger cointegration — a rolling 90-day OLS of ETH on BTC to estimate a time-varying hedge ratio β, z-scoring the residual, and fading extremes (short ETH/long BTC at z≥+2, long ETH/short BTC at z≤-2; exit at |z|<0.5, hard stop, or 21 days) — pitched as the structurally distinct, dynamic-β upgrade over a previously failed 4H fixed-ratio z-score.

Across 100 leg-trades (50 balanced long/short pair entries) it returned -15.7% with profit factor 0.84, Sharpe -0.60, expectancy -$139/trade, an exactly-coin-flip 50% win rate and a 19.4% max drawdown; it was genuinely dollar-neutral (beta -0.01, 7.6% vol, 2.04% fee drag) but carried a fat negative tail (skew -8.76, kurtosis 122, largest loss -$7,979) with losses concentrated in recent years (2021 -9.1%, 2025 -9.7%, 2026 -1.4%).