SolEthRatioMeanReversionPairsLS

Market-Neutral SOL/ETH Ratio Mean-Reversion Pairs Trade (BINANCE USD-M Perps, Daily Log-Ratio Z-Score, Dollar-Neutral Two-Leg, Correlation-Regime Gated, 3-Parameter)

A DOLLAR-NEUTRAL, TWO-LEG statistical-arbitrage pairs trade on two highly-correlated smart-contract-L1 majors: SOLUSDT.BINANCE (primary/long-or-short leg) and ETHUSDT.BINANCE (opposing leg). This is deliberately market-neutral to attack the single most under-represented dimension in the portfolio (direction: long_only is 84.8% vs a ≤55% target — a 30-point gap, the largest of any bucket), while avoiding every recent graveyard: it is NOT a funding-rate signal (L13), NOT a liquidation-feed strategy (L3), NOT an options structure (L2), NOT a Hyperliquid venue (L11), NOT a cross-venue carry (L1), and NOT an N-asset rank-and-rotate basket (L12 — this is a single focused pair, 2 legs, 4 fills per round trip, not 2N). The signal is intentionally minimal to resist the #1 killer (overfit, 202x): the log of the SOL/ETH price ratio is standardized against its own rolling mean/std over a ~20-day window. When the ratio stretches far from its mean (|z| ≥ 2.5), we bet on convergence — long the relatively-cheap leg, short the relatively-rich leg, sized to equal USD notional so net delta ≈ 0. Both legs are on BINANCE USD-M futures (funding/OHLCV fully catalogued for both symbols — no data-availability wall), and both are top-5-liquidity perps so fills are realistic. Held for days until convergence. 3 parameters only: lookback window, entry z-threshold, correlation-regime floor.

Implements the hypothesis exactly: SOLUSDT primary leg + ETHUSDT opposing leg on BINANCE USD-M futures (both fully catalogued, top-liquidity — no data wall). The decision variable is the raw log-ratio z-score returned continuously every bar (not clipped), thresholded in matching units against entry_z/exit_z/stop_z. Only 3 tuning parameters drive economics (lookback, entry_z, corr_floor) to resist overfit; exit/stop/sizing are conventional risk controls. It is deliberately market-neutral (long+short, one focused pair, 4 fills/round-trip) to attack the portfolio's long-only over-representation, and avoids every recent graveyard (no funding, liquidation, options, Hyperliquid, cross-venue carry, or N-asset rotation). Both legs execute atomically inside calculate_signal via _maybe_act using SOL as the position-state anchor, keeping the two legs opened/closed together with timestamp-aligned price histories (single-unit appends preserve index alignment). Futures venue chosen because the strategy shorts one leg and uses leverage>1, which is wired into position sizing (equity * position_fraction * self.config.leverage) as required. Default entry_z=2.0 (vs the hypothesis's 2.5 ceiling) ensures the sandbox window generates trades while the optimizer can tighten it; a 2-sigma ratio divergence is a large move well above round-trip fee cost.

Genuine no-edge / false-premise result (NOT a data or code problem — data coverage is full: 182 trades, 955 trade-days, metrics_reliable=true). The SOL/ETH log-ratio does not mean-revert — it trends across regimes (SOL's 2021 re-rating vs ETH), so the |z|>=2 convergence bet repeatedly shorts the rich leg into a persistent trend and gets steamrolled: total_return -91.7%, max_drawdown 93.1% (hard-abandon threshold), Sharpe -0.73, PF 0.735, negative expectancy, and losing in every year 2021/2022/2023/2025. The 'dollar-neutral' book is not neutral in practice — skew -12.3, kurtosis 271, a single -49% day, largest_loss -$76.5k vs largest_win $28.1k. This cannot be fixed by the 2-hour parameter sweep: reducing size or retuning z-thresholds only rescales a negative-expectancy, sub-50%-win-rate distribution with losers larger than winners — it cannot flip the sign, and the catastrophic tail is structural to fading a trending ratio. Consistent with the portfolio's 0/53 market-neutral survival. Recommend the Research Lead not re-propose ratio/spread mean-reversion on trending L1 major pairs; if a pairs approach is retried it needs a genuinely cointegrated pair (statistically tested for stationarity of the spread, not merely high return-correlation), which SOL/ETH is not.

Dollar-neutral SOL/ETH ratio mean-reversion pairs trade on BINANCE USD-M perps (daily bars). Standardizes log(SOL/ETH) into a z-score over a ~20-day rolling window; when |z| >= entry_z it opens a market-neutral two-leg pair — long the cheap leg, short the rich leg, equal USD notional per leg (net delta ~0). A correlation-regime floor on SOL/ETH daily log-returns gates entries so it only trades while the ratio is genuinely co-moving. Exits on convergence (|z| <= exit_z), divergence stop (|z| >= stop_z), or a max holding period. Leverage 2x with per-leg notional = 25% of equity x leverage, gross-capped, so gross exposure ~1.0x equity while remaining beta-neutral.

On ETH-leg failure, immediately flatten any opened SOL leg, or only set entry state after both fills confirm.

One-legged entry failure leaves an unhedged, un-timed-out position. The SOL leg submits first, then ETH; if ETH submission raises, the except block resets entry_bar_idx=None and direction=0 but the SOL leg is already open. A naked SOL position with entry_bar_idx=None makes held compute as 0 forever, so the max_holding_bars time-stop can never fire — only a z-based exit can close it. Breaks the dollar-neutral invariant. Unlikely in backtest (in-event fills) but latent for paper/live.

corr_floor != stationarity; consider a half-life/trend filter before spending optimization budget.

Sandbox is catastrophic (-91.7%, Sharpe -4.10, PF 0.67, DD 93%) but this is NOT a code-defect signature per L17: win_rate 0.478 (not 0), |Sharpe| 4.1 (not >5), PF 0.67 (not 0), symmetric 91/91 long/short, beta ~ -0.04 (genuinely delta-neutral), polarity correct. It is an edge/regime failure — mean-reversion faded a persistent SOL-vs-ETH trend (short_win_rate 0.407). Analyst should treat as strong prior toward abandon at BACKTEST_REVIEW.

Restore entry_z=2.5 to match the hypothesis, or let the optimizer tighten it.

Default entry_z=2.0 is below the hypothesis's stated |z|>=2.5. Structural mechanic is correct; only the magnitude differs (documented as a deliberate loosening so the 91-day sandbox trades). Not critical.

No data wall: 182 trades over 955 trade-days, both legs fully catalogued, funding_events_available 7203, metrics_reliable=true — the result is a trustworthy read on the edge

Clean single-pair design (2 legs, no basket/rotation aggregation artifact), minimal 3-parameter signal to resist overfit

Decisively losing: total_return -91.7%, Sharpe -0.73, profit_factor 0.735, expectancy -$393/trade, avg_trade_return_pct -2.05%, information_ratio -0.94

Catastrophic and consistent: max_drawdown 93.1% (hard-abandon per L19), negative in 2021 (-42%), 2022 (-9%), 2023 (-41.5%), 2025 (-9.4%)

'Dollar-neutral' book is not neutral: skew -12.3, kurtosis 271, a single -49% day (2021-02-21), largest_loss -$76.5k vs largest_win $28.1k — the SOL/ETH ratio trends across regimes rather than mean-reverting, so convergence bets get run over

Correlation-regime gate did not prevent the blow-ups; premise (ratio mean-reversion) is false, not a tunable defect

SolEthRatioMeanReversionPairsLS attacked the portfolio's largest quota gap — the 84.8% long-only tilt — with a clean, minimal, market-neutral pairs trade, deliberately steering clear of every recent graveyard (funding, liquidations, options, Hyperliquid, cross-venue carry, N-asset baskets) and using full-history Binance perps so there was no data wall. The backtest was a trustworthy read and it was decisively negative: the SOL/ETH ratio trends rather than mean-reverts, so fading its stretches lost 91.7% with a 93.1% drawdown and a wildly non-neutral tail (skew -12.3, a single -49% day, losers far larger than winners), and the correlation gate did nothing to prevent the blow-ups. The reviewer abandoned it as a false premise, not a tunable defect, noting a sweep could only rescale a losing distribution. The recommendation was explicit: do not re-propose ratio mean-reversion on trending L1 pairs, and any future pairs trade must use a genuinely cointegrated, stationarity-tested spread — which SOL/ETH is not.

High return-correlation is not cointegration: a ratio can be tightly correlated yet still trend across regimes (SOL's 2021 re-rating vs ETH), so ratio/spread mean-reversion on trending L1 major pairs steamrolls convergence bets — any pairs approach needs a spread statistically tested for stationarity, not merely correlated legs, consistent with the portfolio's 0/53 market-neutral survival.

The backtest reviewer returned 'abandon' at the pre-optimization gate: this was a genuine false-premise result rather than a data or code problem — the SOL/ETH log-ratio trends across regimes instead of mean-reverting, so the convergence bet repeatedly shorted the rich leg into a persistent trend, and the 93.1% drawdown alone crossed the hard-abandon threshold. A parameter sweep cannot flip a negative-expectancy, sub-50%-win-rate distribution with losers larger than winners.

A dollar-neutral, two-leg statistical-arbitrage pairs trade on SOLUSDT and ETHUSDT Binance USD-M perps (daily bars) that standardized the log SOL/ETH price ratio into a z-score over a ~20-day window and bet on convergence when it stretched (|z| >= entry threshold) — longing the cheap leg and shorting the rich leg at equal notional — gated by a correlation-regime floor and held for days.

With full data coverage (182 trades over 955 trade-days, metrics_reliable=true), the result was decisively losing: total return -91.7%, Sharpe -0.73, profit factor 0.735, expectancy -$393/trade, avg_trade_return_pct -2.05%, and information ratio -0.94, with a catastrophic 93.1% max drawdown, losses in 2021/2022/2023/2025, and a supposedly neutral book showing skew -12.3, kurtosis 271, a single -49% day, and a largest loss of -$76.5k vs largest win $28.1k. No optimization stage ran.