SolEthRatioSpreadTrendFollowerLS

SOL/ETH Relative-Momentum Spread-Trend Follower, Dollar-Neutral Long-Short (BINANCE USD-M Perps, Daily Log-Ratio vs Its Own Moving Average — Ride the Ratio TREND, NOT Fade It, Pure OHLCV, 2-Parameter)

A DOLLAR-NEUTRAL, two-leg relative-strength TREND follower on SOLUSDT.BINANCE (primary) vs ETHUSDT.BINANCE (secondary). It is the deliberate, evidence-driven INVERSE of the just-abandoned 'SOL/ETH Ratio Mean-Reversion Pairs Trade': that hypothesis died a clean no-edge death because the analyst confirmed the SOL/ETH log-ratio does NOT mean-revert — it TRENDS across regimes (SOL's structural re-rating vs ETH). This strategy takes that empirical finding and trades WITH the ratio trend instead of against it: hold long the currently-outperforming leg and short the underperforming leg, flipping only when the spread trend reverses. It attacks the single largest quota gap — DIRECTION (long_only 84.8% vs a ≤55% target) — by being fully dollar-neutral long/short, and it reuses the ONE relative-strength edge that already survived here (Cross-Sectional Momentum 8-major LONG, Sharpe 2.6), compressed to a focused 2-asset spread so fees are 4 fills per flip, not 2N. It uses PURE OHLCV (the best-surviving tag family, 5%), avoids every data wall (no options/liquidations/funding/HL), and is deliberately 2-parameter to resist overfit (the #1 killer, 202x). It is NOT the mean-reversion pair (opposite sign), NOT an N-asset basket (L12), NOT a funding/liquidation/options strategy.

Iteration 2 makes the two smallest changes the backtest-review demanded, and nothing else — the signal/entry/exit logic and imports that already passed Layers 1-3 are untouched. (1) RESTORED NAKED-LEG GUARD (the flagged regression): at the top of _maybe_act, when bool(sol_pos) != bool(eth_pos) it exits whatever single leg is open, resets _position_direction=0, and returns — so a flip that half-fills can never leave a one-sided leveraged (directional) book carried across bars, which is the almost-certain source of the implausible +447%/+194%/+113% single days on a spread that never moved that much. has_position is now correctly `sol_open AND eth_open`, so re-establishment only happens from a genuinely flat, paired state. (2) EXPOSURE FIX: set leverage 1.0 (was 2.0) so per-leg notional = equity x 0.25 and gross = ~0.50x equity — matching the intended ~50% gross the reviewer cited against the observed 200%, and removing the leverage amplification that magnified the artifact days; both legs deliberately share the same leverage/position_fraction, so the pair stays balanced. On the review's diagnostic asks: the spike days are consistent with a transiently naked, leveraged single leg (only a one-sided directional position can return hundreds of percent when the SOL/ETH ratio barely moved) — the guard eliminates that mechanism, and the reduced gross brings exposure to the intended level. This is explicitly NOT sent to optimization yet: once genuinely delta-neutral, the analyst should re-run and confirm the Sharpe CI no longer straddles 0 and that the edge is not solely 2021/2024 spike days before optimizing; optimizing the prior spike-driven, kurtosis-688 series would overfit the artifacts and fail the deflated-Sharpe/PBO gates.

Over-exposure/naked-leg artifact contaminating the headline of a zero-survivor pairs mechanism — not worth 2 hours, escalate the sizing defect. The equity curve is inflated by physically impossible single-day returns on a supposedly dollar-neutral SOL/ETH pair: 2021-05-05 +229.9%, 2024-01-12 +81.7%, 2021-10-06 +60.4% — the ratio did not move anywhere near that, and at ~25%-per-leg neutral sizing a day cannot return 230%. The iteration-2 'naked-leg guard' did not prevent it, and the book's neutrality invariant is violated (exposure_pct 200.6%, end_unrealized_pct 106.9), so the legs are intermittently one-sided/over-exposed (return_kurtosis 55.9). The 1282% total is those artifact days (2021 +417%, 2024 +77%); stripped of them the edge is modest and 2026 is -4.7%. Even setting the artifact aside, this is the L45/L52 market-neutral pairs family (near-zero survivors) and the direct sign-flip twin of the SOL/ETH mean-reversion pair already abandoned this session; benchmark_meaningful is correctly false and max_drawdown CI runs to 65.8%. Sharpe 1.05 / PSR 0.997 are untrustworthy because they are computed on a PnL with a +230% artifact day. No optimization runs cleanly on that, and the underlying mechanism is a zero-survivor class. Failure pattern: over-exposure/naked-leg sizing artifact + no_edge market-neutral SOL/ETH pairs (L45/L52).

Dollar-neutral two-leg relative-strength TREND follower on SOLUSDT.BINANCE (primary) vs ETHUSDT.BINANCE (secondary), daily bars, pure OHLCV. Signal is the z-score of log(SOL/ETH) vs its own 30-day moving average; above +entry_z it holds LONG SOL / SHORT ETH, below -entry_z SHORT SOL / LONG ETH, with a hysteresis band so it flips only on a genuine spread-trend reversal. Both legs are opened/closed atomically at equal USD notional per leg. A naked-leg guard flattens the book immediately on any bar where exactly one leg is open, so the position is always either fully paired (delta-neutral) or flat. Sized at position_fraction per leg with leverage 1.0, so gross exposure is ~0.50x equity (25% long + 25% short).

CLEAN RESTART 2026-09-04 — this run's verdict history and learning records were removed and it was restarted from verification. Its previous abandonment came from the pipeline, not from the market: the Layer-2 harness mis-bound @staticmethod helpers (fixed), QA issued terminal performance verdicts on an unoptimized smoke test (removed — QA now judges correctness only), and sandbox timeouts came from backtest-slot starvation (fixed). The hypothesis and the strategy code are unchanged. Verify the code on its merits; performance is decided later by the full backtest and the optimizer.

No change needed; the SPOT routing is the rules-correct choice. If the research lead intended a leveraged/short-capable overlay, that would be a different strategy.

Venue: the hypothesis title/description say 'BINANCE USD-M', but the code and config use BINANCE_SPOT (BTCUSDT.BINANCE_SPOT etc.). This is NOT a defect — SPOT is the correct routing for a long-only, no-leverage, no-funding basket per the factory venue rules (futures only when leverage>1, shorts, or funding-based), the assets are identical, the economic intent is preserved, and it passed the Layer-1.5 hypothesis-consistency gate. Flagged only so the mismatch between the hypothesis prose and the (correct) implemented venue is on record.

Keep the parameter count minimal in optimization; the _param_bounds should bound all three so the walk-forward cannot over-search the added regime_sma_days.

The hypothesis specifies '2-parameter' (vol lookback, regime cutoffs), but iteration 3 tunes THREE (vol_lookback, vol_cutoff, regime_sma_days) after adding the BTC-SMA downtrend gate. The docstring/description acknowledge 3. This modestly raises optimizer best-of-N inflation risk versus the stated 2, but the structure (vol overlay + drawdown gate) is coherent and the added gate is a documented drawdown control, not a fitted signal.

Measured sample (158 trades over 6 years); clean 2-parameter pure-OHLCV construction; trades WITH the (real) ratio trend

Low fee drag (commission 1.08% of gross); genuinely fills the dollar-neutral/long-short bucket

Physically impossible single-day return (2021-05-05 +229.9%, plus +81.7% and +60.4% days) on a supposedly dollar-neutral pair — the iteration-2 naked-leg guard failed; return_kurtosis 55.9

'Dollar-neutral' invariant violated: exposure_pct 200.6%, end_unrealized_pct 106.9 — legs run one-sided/over-exposed, which creates the artifact days

The 1282% return is those artifact days (2021 +417%, 2024 +77%); strip them and the edge is modest; 2026 -4.7%

L45/L52 market-neutral pairs zero-survivor class and the direct sign-flip twin of the just-abandoned SOL/ETH mean-reversion pair; max_drawdown CI to 65.8%