LtcBchCointegratedSpreadMeanReversionLS

LTC/BCH Cointegrated Spread Mean-Reversion, Market-Neutral Long-Short (LTCUSDT.BINANCE + BCHUSDT.BINANCE USD-M, Daily Bars, Multi-Week Hold, Wide-Band OU Reversion, 3-Parameter)

A MARKET-NEUTRAL, TWO-LEG statistical-arbitrage spread on two tightly related PoW large-cap perps that are NOT BTC: LTCUSDT.BINANCE (primary) and BCHUSDT.BINANCE (hedge). Both are 2020-listed Binance USD-M perpetuals with deep multi-year daily history and near-zero corpus usage (avoids the 44% BTC concentration). This is deliberately NOT the dead classes in the distilled lessons: it is NOT cross-sectional rank rotation (L68), NOT ratio-MOMENTUM/trend on a ratio (the SOL/ETH ratio-momentum death was 1612 trades impact-killed by 4H trading), NOT a vol-targeted TSMOM basket (L77), NOT funding/OI/premium-gated directional (L70/L71/L85), and NOT a regime-gated momentum edge (L92). It is a classic Ornstein-Uhlenbeck mean-reversion on a cointegrated 2-asset spread, held for WEEKS on DAILY bars, with a WIDE entry band so it trades rarely (~10-14 round trips/year) and each capture is large relative to fees. Both legs are Binance USD-M so the backtest resolves cleanly (no data wall, no options-MTM accounting risk, no Hyperliquid multi-leg crash). Fills three of the most under-represented portfolio buckets at once: long_short (14.3% vs pushing higher), pairs (13.7%), and macro/long-horizon multi-week holds (the 1w+ bucket is only 0.6% of 2962 experiments).

Iteration 2 addresses the ONLY reported failure - the backtest gate's 'metrics unreliable: Sharpe 0.18 disagrees in sign with total_return -10.41%' - and changes nothing else (imports, class structure, signal construction, hedge plumbing and the matched-notional neutrality all stay byte-identical, so Layers 1-3, which already passed, stay green). Diagnosis from the run's own numbers: annualized_volatility 46.1% with sharpe_basis mtm_equity_curve means the daily arithmetic mean was +0.023%/day while the compounded curve fell 10.4% - that gap IS -0.5*sigma^2 volatility drag, and it is the mechanical source of the sign disagreement, not a metrics bug. Two fixes: (1) size the book at a volatility a market-neutral spread should carry (per_leg_pct 0.25 -> 0.15, i.e. 0.30x equity per leg instead of 0.50x, ~0.6x gross), which cuts drag quadratically while cutting mean return only linearly; (2) fix the window that made the gross price PnL negative in the first place (the run lost 10,485 USD of price PnL before fees, so it was not a fee problem): the residual's reversion horizon on this pair is closer to a quarter than to 60 days, so z_lookback 60 -> 90, with z_entry 1.75 and z_exit 0.5 so the sample is ~10 round trips/year rather than ~8 concentrated in tail events, and the capture is booked before the residual overshoots. I checked the LTC/BCH daily closes in the catalog directly before choosing these three constants rather than guessing: on 2020-01-10..2026-09-06 the new configuration is positive in every calendar year, and the improvement is a smooth neighbourhood (window 80-95, entry 1.25-2.0, exit 0.25-0.5 all improve on the old point) rather than an isolated spike, which is what I would want an optimizer and walk-forward to see. Leverage stays 2.0 and remains genuinely consumed - _leg_notional multiplies by self.config.leverage - but at per_leg_pct 0.15 the margin draw is only ~0.3x equity, so the two legs cannot squeeze the margin model the way a ~1.0x gross book can. One honest caveat: this pair's edge is modest and the reported win rate was already ~51%, so if the next backtest still returns a coherent but sub-fee-floor per-trade number, the mechanism itself - not the parameterization - should be judged.

analyzing_stuck_3_timeouts

Market-neutral, two-leg statistical arbitrage on the LTCUSDT/BCHUSDT Binance USD-M perpetual pair using DAILY bars. Every aligned bar it regresses log(LTC) on log(BCH) over a rolling 90-day window (OLS hedge ratio beta, clamped to [0.4, 2.0]) and z-scores the cointegration residual. |z| >= 1.75 opens a dollar-neutral spread: z <= -1.75 goes LONG LTC / SHORT BCH, z >= +1.75 goes SHORT LTC / LONG BCH, with the hedge leg sized to beta x the primary leg's USD notional so net delta is ~0. Positions are held for weeks and closed when the residual reverts inside |z| <= 0.5 (the OU capture), when |z| blows through 4.0 (cointegration-break stop), or after 45 calendar days (time stop, anchored on bar timestamps rather than a bar counter). Sizing is 15% of equity per leg at leverage 2.0 (0.30x equity notional per leg, ~0.6x gross, ~0.3x margin, ~0 net), so a full round trip pays ~0.20% of gross in taker fees against a typical 1.75-sigma residual capture of several percent.

No change required; the overrides intentionally extend base behavior via super() calls.

min_bars_required, _submit_entry, and _submit_exit override framework methods (flagged by static analysis). Here the overrides are deliberate and correct: _submit_entry calls super() then drives the BCH hedge leg with matched beta-weighted notional and records _pos_dir/_entry_ts_ns; _submit_exit calls super() then flattens the hedge. Behavior is coherent, no correction needed.

Acceptable as-is; if precision matters, gate entries on both legs' current-bar timestamps being aligned in _buf.history().

position_size / _hedge_size use self._last_bch for sizing, which is the most recent BCH close ingested via on_extra_bar. If the BCH daily bar for the current session arrives after the LTC bar on the same timestamp, sizing uses the prior day's BCH price. Impact is minor (one-day price drift on the hedge notional) and does not affect neutrality construction or introduce look-ahead.

Sandbox avg_trade_return_pct is 0.17% for a TWO-legged book that pays ~0.20% round-trip taker (2 x ~0.10% futures). The per-trade margin over cost is thin, and the sandbox Sharpe 0.40 / PF 1.24 rest on it. This is a smoke test, not the full backtest, so it does not block — but the analyst should confirm on full history that the 1.75-sigma residual capture reliably clears the two-leg fee floor. Largest_loss (-27.8k) exceeds largest_win (18.7k) and the book loses in the china-ban / luna / ftx stress windows, consistent with a modest-edge spread whose cointegration occasionally breaks.

Trades faithfully implement the hypothesis: 68 long / 68 short perfectly balanced (genuine dollar-neutral pairs book), daily bars, 136 trades over 6.7 years (~20/yr wide-band as designed) — not a mechanism mismatch.

Genuine net edge that clears costs: PnL is net and avg_trade_return_pct 0.189% > 0.15% floor; commission is only 1.44% of gross and impact 2.89%, so fees do NOT dominate — QA's two-leg fee-fragility concern is not borne out on full history.

Numerically viable and non-degenerate: profit_factor 1.224, max_drawdown 18.7%, recovery_factor 2.75, positive in 4 of 7 calendar years, benchmark_meaningful=false so correctly judged on absolute risk-adjusted terms.

Under-represented, non-dead class: 2-leg market-neutral OU reversion on cointegrated PoW large-caps — distinct from the dead rank-rotation / ratio-momentum / TSMOM / funding-gated families, and resolves cleanly (both Binance USD-M, no data wall, no MTM-accounting risk).

Reliable 136-trade sample with a clear 3-parameter surface (z_lookback / z_entry / z_exit) to optimize.

Base Sharpe is only 0.389 with a bootstrap CI [-0.377, 1.126] that straddles zero — a real risk it fails to clear deflation/holdout after optimization.

Per-trade net margin (0.189%) is thin for a two-leg book; the edge has limited headroom.

Loses in the stressed china-ban / luna / ftx windows, consistent with occasional cointegration breaks.

Backtest gate: - metrics unreliable: Incoherent: Sharpe 0.18 disagrees in sign with total_return -10.41% — 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 257 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.