EthBtcRatioTrendMarketNeutralLS

ETH/BTC Ratio Trend, Market-Neutral Long-Short Pair (Long ETHUSDT.BINANCE / Short BTCUSDT.BINANCE USD-M When the ETH-over-BTC Ratio is Trending Up, Reverse When Trending Down, Daily Bars, Zero-Cross with Deadband, 2-Parameter)

A DOLLAR-NEUTRAL, TWO-INSTRUMENT relative-value TREND strategy that trades the ETH/BTC price RATIO as its own instrument on daily bars, using BINANCE USD-M perpetuals so both legs can be shorted. This is deliberately NOT another single-name absolute-momentum clone (the one surviving family, now saturated across BTC/ADA/SOL/AVAX/NEAR/etc.) and NOT a cointegration/mean-reversion pair (LtcBchCointegratedSpread died): it is the crypto complex's single most persistent RELATIVE trend — capital rotating between the two largest assets ('alt-season' vs 'BTC-dominance' regimes). The ratio R = P(ETH)/P(BTC) trends for weeks at a time; we ride the trend in the SPREAD, not the direction of crypto. Because the book is beta-neutral to the whole complex, it should make money in exactly the vol-stressed / crash windows where directional trend books lose (a genuine regime-diversifier for the portfolio, whose entire long-short survivor set is single-name directional). Only 2 parameters (lookback, deadband) to resist the overfit that killed 380 prior hypotheses. Primary = ETHUSDT.BINANCE, hedge = BTCUSDT.BINANCE, dollar-matched legs (equal USD notional each side => ~zero net crypto beta by construction; residual beta from ETH's ~1.1 beta to BTC is small and further trimmed by the neutral sizing).

Iteration 2 targets the incoherent-metrics failure (Sharpe +0.31 vs total_return -11.24%): a positive mean daily mark-to-market return with a negative compounded equity curve is volatility drag, and the drag source here was a structural leg-desynchronisation bug. The old should_enter required the BTC hedge to already appear in cache.positions_open before allowing the ETH entry, but the hedge market order is submitted on the same bar and only fills on the NEXT bar - so every entry ran a fully NAKED directional BTC leg for a whole daily bar before the ETH leg went on. That injected raw crypto beta and its variance into a book that is supposed to be beta-neutral, inflating mtm volatility (drag) while the realized equity curve bled. Fix: track the submitted-but-unfilled hedge as a pending direction (_hedge_pending, expiring after 2 bars if it never fills) and accept it in should_enter, so both legs are submitted on the same bar and fill together; plus a should_exit safety that flattens the ETH leg if the mirrored hedge ever goes missing, so the book can never carry a one-legged directional position (lesson L59). Leg_fraction was trimmed 0.35 -> 0.30 (gross 1.2x equity at leverage 2.0, which the sizing consumes via self.config.leverage) to further cut compounding drag. Signal logic, imports, class structure and the deadband state machine that passed Layers 1-3 are unchanged.

analyzing_stuck_3_timeouts

Dollar-neutral ETH/BTC ratio-trend pair on Binance USD-M daily bars. Signal = 100*ln(R_t/R_{t-lookback}) where R = P(ETHUSDT)/P(BTCUSDT), computed every bar. A zero-cross-with-deadband state machine goes long-spread (long ETH / short BTC) when the ratio trend exceeds +deadband, short-spread when it falls below -deadband, and flat on the zero cross. Both legs are sized to identical USD notional (equity * leg_fraction * leverage) so the book carries ~zero net crypto beta and expresses only the ETH-over-BTC rotation. Two signal parameters (lookback, deadband).

Optionally require both same-ts closes to be present before computing/acting on the signal for that bar, to guarantee ETH and BTC are always sampled at the identical instant.

On a shared close-timestamp, if the ETH primary bar is dispatched before the BTC extra bar, calculate_signal falls back to _hedge_last (the prior day's BTC close), so the ratio mixes today's ETH close with yesterday's BTC close for that bar. This is conservative (never future-peeking) but introduces a one-day staleness in the ratio depending on intra-timestamp dispatch order. The lookback-trend signal is robust to this, so it is not a correctness failure.

Verify the base template calls the subclass min_bars_required for warmup gating; rename if the override is not honored.

min_bars_required overrides a framework method (static-analysis base_shadow warning). Appears intentional (returns lookback+2 to gate warmup) and behaves correctly, but confirm the base template honors the override rather than its own implementation.

Sandbox is a smoke test but two things are worth carrying to the analyst: (1) regime attribution shows the 'normal' vol tercile at -30.4% return / -0.02 Sharpe while all PnL comes from calm (+28.5%) and stressed (+38.9%) terciles — the market-neutral diversifier thesis holds in stress (good) but the book bleeds in the middle regime, so full-history edge is regime-lumpy; (2) max_drawdown 49.6% with a 1381-day drawdown-duration and Sharpe CI [-0.40, 1.06] straddling zero — the point estimate (Sharpe 0.33, PF 1.04) is marginal and not yet skill-distinguishable. Also note leg sizing is equal-USD-notional (dollar-neutral), which is what the hypothesis specifies, but ETH carries ~1.1 beta to BTC so a small residual long-beta remains by construction — the hypothesis acknowledges this. None of this blocks; the optimizer/full backtest decides.

avg_trade_return_pct 0.537% is ~3.5x the 0.15% Binance USD-M fee floor on a two-leg basis; commission is only 1.71% of gross and impact 12.8% of gross, so this is NOT a fee-fragility case — the gross edge clears costs.

Adequate, decisive sample: 198 trades (99 long-spread / 99 short-spread, balanced) over 6.7 years of daily bars; entry_diag shows 99 signaled = 99 submitted with both legs synchronized on the same bar.

Code faithfully implements the stated hypothesis — dollar-neutral ETH/BTC ratio-trend with a zero-cross+deadband state machine, benchmark_meaningful correctly false, positive expectancy (+$100/trade).

Genuinely differentiated mechanism (relative-value ratio trend, market-neutral) with only 2 signal parameters and a large unexplored region; not a saturated single-name momentum clone (L82/L93) nor a cross-sectional dollar-neutral basket (L45). Regime attribution confirms the diversifier thesis in stress (calm +20.6%, stressed +16.8%).

profit_factor 1.029 is thin and sits in the fee/impact drag band; payoff ratio only 1.07 (avg_win 7198 vs avg_loss 6716) with 0.49 win rate is a weak trend-follower distribution — a real trend book wants big asymmetric winners.

Sharpe 0.343 with CI [-0.44, 1.08] straddling zero — point estimate not yet skill-distinguishable.

max_drawdown 49.5% with a 1381-day (3.8yr) drawdown duration on a book that is supposed to be a diversifier; whole positive return is carried by 2022 and 2025 spikes.

Regime-lumpy and time-lumpy: loses in the 'normal' vol tercile (-12.9%) and in 3 of 7 calendar years, including back-to-back 2023 (-21%) and 2024 (-26%). QA's regime-lumpiness and marginal-point-estimate concerns are borne out on full history.

Backtest gate: - metrics unreliable: Incoherent: Sharpe 0.31 disagrees in sign with total_return -11.24% — the Sharpe is computed from the mtm_equity_curve series while the return comes from the equity curve. Judge on the equity curve.

Benjamini-Hochberg at q=0.10 over 253 programme candidates keeps 4. A candidate that does not survive here is not distinguishable from the programme's own noise, however good its individual statistics look.