SolAvaxHlCointGatedStatArb1D

Cointegration-Gated Stat-Arb: SOL/AVAX Spread Mean Reversion on HYPERLIQUID (Market-Neutral, Daily, ADF + Half-Life Gated)

Dollar-neutral statistical-arbitrage pairs trade on two SAME-SECTOR HYPERLIQUID L1 perps (SOL and AVAX) whose RELATIVE value oscillates within alt-rotation cycles. This is a deliberate reframe of a prior ETH/BTC z-fade that failed because ETH/BTC is NOT a stationary spread (ETH structurally lost ground to BTC for years, so a fixed-window z-fade repeatedly shorted the winner into a trend). The fix is NOT a new narrative pair — it is an ADAPTIVE ADMISSION GATE that only allows positions while the spread is empirically stationary. Pipeline per bar: (1) compute log-spread = log(SOL) - beta*log(AVAX) where beta is a rolling OLS hedge ratio over the lookback (NOT a fixed 1:1 ratio); (2) run a rolling Engle-Granger / ADF stationarity test on the spread over the lookback window — if the spread is NOT stationary (ADF p-value >= 0.05), STAND ASIDE and force-flatten any open position; (3) estimate the Ornstein-Uhlenbeck half-life of the spread, and only admit trades when half-life is in a tradeable band (e.g. ~2-25 days) — reject sluggish or non-reverting regimes; (4) when stationary AND half-life valid, compute the spread z-score and fade extremes. Both legs trade on HYPERLIQUID (~0.045% taker/side, ~0.18% RT for the pair) on DAILY bars, so the HL ~5000-candle limit is a non-issue. Simultaneously advances three of the most-violated portfolio quotas: HYPERLIQUID venue (toward >=20%), market-neutral direction (toward reducing the 91% long-only tilt), and pairs scope.

Implements the REVISED hypothesis exactly, addressing the analyst's core finding that fixed-window z-fades fail on non-stationary spreads. (1) INSTRUMENT: pivots from ETH/BTC (majors that trended apart for years) to SOL/AVAX, two same-sector L1 perps whose relative value oscillates within alt-rotation cycles, giving a genuinely mean-reverting residual. (2) MECHANISM: adds the two fixes the analyst prescribed — a rolling Engle-Granger/ADF STATIONARITY GATE (implemented as a proper ADF t-stat with lag augmentation, compared to the ~5% critical value; trades only when stationary, force-flat otherwise) and an OU HALF-LIFE FILTER with a matched time stop — converting 'assume reversion' into 'measure reversion, stand aside otherwise' and removing the fade-a-trend failure mode. (3) HEDGE RATIO: replaces the fixed 1:1 ratio with a dynamic rolling OLS beta in log space, the statistically correct construction for a cointegrated spread. KEPT the salvageable core: dollar-neutral z-score reversion, daily bars, HYPERLIQUID venue — filling the three most-violated quotas (HL venue, market-neutral direction, pairs scope). Venue: HYPERLIQUID futures is correct because the strategy is long+short (market-neutral), so it cannot run on spot. leverage=1.0 (each leg 0.35x equity => ~0.7x gross), so leverage is not referenced in sizing — avoiding the leverage-unused gate. Daily bars sidestep the HL ~5000-candle history limit. Residual risk (thin trade count from selective gating) is mitigated with a 5% ADF threshold and a wide 2-40 day half-life band rather than over-tight gates; the ADF/half-life logic was unit-tested to correctly pass a stationary AR(1) series and reject a random walk.

Not worth optimizing: only 4 pair trades (8 legs) over 3 years, all clustered in a single ~2-month window (Feb–Apr 2024) — far too small a sample for a 3-window walk-forward, and the OOS windows would be effectively empty. The scarcity is structural, not tunable: the ADF + half-life gate (the strategy's entire thesis) correctly stood aside ~97% of the time because SOL/AVAX is only fleetingly stationary — the very non-stationarity failure the hypothesis set out to escape (ETH/BTC), now reappearing on the new pair. Even in the tradeable window the edge does not survive costs: gross expectancy is only +$20/trade (PF 1.03), commission $250 and market impact turn total_return negative (-0.45, Sharpe ~0.08), and impact_cost_pct is 50% of gross with capacity only ~$400k (position_fraction 0.35/leg is too large for daily HL alt liquidity). The only ways to raise trade count are to loosen the gate — which re-admits exactly the non-stationary spreads it exists to reject — or to select a genuinely cointegrated pair, which is a hypothesis-level retarget for the Research Lead, not a developer parameter iteration. FAILURE PATTERN: an adaptive cointegration gate is only as good as the pair it is pointed at; a well-built ADF/half-life stat-arb on a pair that is rarely stationary (SOL/AVAX) correctly stands aside almost always, yielding a sample too small to optimize and a thin gross edge that fees + 50% market impact erase. The gate working as designed is the diagnosis, not a bug — the pair, not the code, is the problem.

Market-neutral, cointegration-gated statistical-arbitrage pairs trade on SOL/AVAX HYPERLIQUID perps using 1-DAY bars. Each bar: (1) estimate a rolling OLS hedge ratio beta from log(SOL)=alpha+beta*log(AVAX) over a 60-day window and build the log-spread = log(SOL) - beta*log(AVAX); (2) run an Augmented Dickey-Fuller stationarity test on the spread and only admit trades while it is empirically stationary (ADF t-stat <= -2.86, ~5% critical), force-flattening any open pair when stationarity breaks; (3) estimate the Ornstein-Uhlenbeck half-life and only admit when it falls in a tradeable band (2-40 days); (4) z-score the spread and fade extremes (|z|>=2.0). z>0 (SOL rich) => SHORT SOL / LONG AVAX; z<0 (SOL cheap) => LONG SOL / SHORT AVAX, dollar-neutral 0.35x equity per leg. Exits on reversion (|z|<=0.5), regime break (|z|>=3.5), gate failure, or 25-bar time stop. Both legs submitted atomically from the acting layer; the returned signal is the continuous spread z-score.

Genuinely differentiated and well-engineered: dollar-neutral pairs, rolling OLS hedge ratio, ADF stationarity gate + OU half-life band, atomic two-leg submission — a clean market-neutral construction that advances the HYPERLIQUID and market-neutral quotas

The adaptive admission gate behaves correctly, standing aside when the spread is non-stationary rather than blindly fading a trend

Only 4 pair trades (8 legs) in 3 years — far too small a sample for a 3-window walk-forward; optimization would fit noise on near-empty OOS windows

All trades cluster in a single ~2-month window (Feb–Apr 2024); the gate admits SOL/AVAX almost never, meaning the pair is only fleetingly cointegrated — the same non-stationarity failure the hypothesis tried to escape, on a new pair

Edge does not survive costs: total_return -0.45, Sharpe ~0.08 (CI low -11.9), gross expectancy only +$20/trade against $250 commission

impact_cost_pct 50% — market impact eats half the gross PnL — with capacity only ~$400k; position_fraction 0.35/leg is too large for daily HL alt liquidity

Fixing trade frequency requires either loosening the gate (which reintroduces the exact non-stationary-spread failure mode) or choosing a different pair (a hypothesis-level change, not developer tuning)

Mechanism class (dollar-neutral ratio mean-reversion / statistical arbitrage) is genuinely unrefuted — absent from the failure patterns, and fills three under-represented buckets (HYPERLIQUID venue, market-neutral direction, pairs scope).

Code is sound: same-timeframe BTC hedge leg with a cross-leg alignment barrier (same-timestamp z-score, not stale T-1), synced single ratio buffer, continuous z-score signal thresholded in should_enter (avoids the gated-to-0.0 anti-pattern), dollar-neutral equal-notional legs, spot-correct long-short semantics.

benchmark_meaningful=FALSE correctly flagged (market-neutral), so judged on absolute metrics rather than penalized vs buy-hold.

Honest base config LOSES money: Sharpe -0.29, profit_factor 0.96, total_return -4.4%, negative expectancy (-$1,686/trade) over 72 round trips — ETH/BTC ratio did not profitably mean-revert.

Optimized config is a degenerate overfit: win_rate 8.9%, expectancy STILL NEGATIVE (-$3,613/trade), return_kurtosis 440, headline carried by one end-of-sample day (2026-05-20 +30.18%, the last bar) with end_unrealized_pct 5.73 (un-exited MTM position).

Deflated Sharpe = 0.0 (is_significant FALSE); optimized Sharpe 0.86 far below the expected-max luck bar of 10.27 over 225 trials.

PBO = 0.9656 (~97% probability of backtest overfitting — extreme).

Holdout FAILED with a NEGATIVE Sharpe -1.56 (ratio -0.481).

Sensitivity FAILED with 4 cliffs on the CORE signal params (z_exit, z_entry, lookback, dd_stop_pct).

Capacity tiny ($31M) with impact_cost_pct 5.67% — edge (if any) exists only at toy scale.

Walk-forward is_overfitted=FALSE is contradicted by OOS windows [-2.27, 11.998, 0.0] (one outlier carries it), PBO 0.97, and the negative holdout.

The Research Lead REVISED this hypothesis because the previous premise did not work. Implement the REVISED hypothesis (see the hypothesis description) from scratch — do not reuse the prior strategy. Revision rationale: WHAT I CHANGED vs the failed ETH/BTC hypothesis and WHY: (1) INSTRUMENT — pivoted from ETH/BTC (two majors with DIFFERENTIATED drivers that trended apart for years) to SOL/AVAX, two same-sector high-throughput L1s whose relative value oscillates within alt-rotation cycles, giving an actual mean-reverting residual. (2) MECHANISM — the analyst's core finding was that the PREMISE 'the ratio is stationary' was empirically false, and no parameter tuning can rescue a non-stationary spread. So I added the exact fixes the analyst prescribed: a rolling Engle-Granger/ADF STATIONARITY GATE (trade only when the spread is currently stationary, force-flat when it isn't) and an OU HALF-LIFE FILTER with a half-life-matched time stop. This converts the strategy from 'assume reversion' to 'measure reversion, stand aside otherwise' — directly removing the failure mode of fading a trending spread. (3) HEDGE RATIO — replaced the fixed dollar-neutral ratio with a dynamic rolling OLS beta, which is the statistically correct construction for a cointegrated spread. WHAT I KEPT (salvageable core): the dollar-neutral z-score reversion mechanism, daily bars, and the HYPERLIQUID venue — these still fill the three most-violated quotas (HL venue, market-neutral, pairs) exactly as intended, and daily bars sidestep the HL ~5000-candle history limit. I did NOT re-propose a fixed-window z-fade on a major-vs-major ratio, per the explicit instruction. The residual risk the Analyst should monitor is trade count: selective gating on a young HL daily history may yield a thin sample, in which case the fallback is to broaden the same-sector candidate pool rather than relax the stationarity test.

SolAvaxHlCointGatedStatArb1D was a well-engineered, genuinely differentiated market-neutral pairs trade — a deliberate reframe of a failed ETH/BTC z-fade — that added an ADF stationarity test and OU half-life band so it would only fade the SOL/AVAX spread while that spread was empirically mean-reverting, simultaneously advancing the Hyperliquid, market-neutral, and pairs quotas. The gate behaved exactly as intended, but that was the problem: SOL/AVAX proved only fleetingly stationary, so the strategy admitted just 4 pair trades in three years, all bunched into Feb–Apr 2024, for a net-losing result (-0.45 return, Sharpe ~0.08) whose thin +$20/trade gross edge was wiped out by commissions and ~50% market impact at a tiny ~$400k capacity. The reviewer abandoned it at the backtest-review gate after two iterations, and the analyst recommended revising the hypothesis toward a truly cointegrated pair, concluding the pair — not the code — was the failure.

An adaptive cointegration gate is only as good as the pair it points at; a correctly-built ADF/half-life stat-arb on a rarely-stationary pair will properly stand aside almost always, yielding a sample too sparse to validate and a gross edge too thin to survive fees and market impact — the gate working as designed is the diagnosis, not a bug.

It was abandoned at the pre-optimization backtest-review gate for a structurally untunable trade drought: the ADF + half-life gate correctly stood aside almost always because SOL/AVAX is only fleetingly cointegrated, leaving far too small a sample to optimize and a thin gross edge that fees and 50% market impact erased. The analyst's verdict was revise_hypothesis — the fix is a genuinely cointegrated pair, a hypothesis-level retarget rather than a parameter change.

A dollar-neutral, market-neutral SOL/AVAX statistical-arbitrage pairs trade on Hyperliquid daily bars that used a rolling OLS hedge ratio and an adaptive admission gate — a rolling ADF stationarity test plus an Ornstein-Uhlenbeck half-life band — to fade the spread z-score only while the spread was empirically stationary, standing aside otherwise.

The base backtest was net-losing and statistically empty: total return -0.45, Sharpe ~0.08 (CI low -11.9), only 8 legs across 4 pair trades in 3 years (50% win rate, profit factor 1.03), all clustered in a single ~2-month window (Feb–Apr 2024). Gross expectancy was just +$20/trade against $250 commission, with market impact eating ~50% of gross and capacity only ~$400k.

EthBtcRatioMeanReversionPairsHyperliquidDaily