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)

Hypothesis

Decline optimization of this tested LTC/BCH operationalization: reliable mean net leg return is -0.673% versus the 0.15% venue floor, with leg profit factor 1.019 and paired-episode profit factor 1.046. The previously requested frozen-equilibrium repair is present, and no further specific implementation repair is supported. This does not reject pairs trading generally or draw conclusions about the withheld period after 2025-06-16.

Implementation

Market-neutral OU mean reversion on the LTC/BCH log-price cointegration residual (LTCUSDT + BCHUSDT Binance USD-M, daily bars). Enter long LTC/short BCH when the rolling 90-day residual z <= -z_entry (reverse when >= +z_entry), legs at beta-matched USD notional. While in a position, the residual is measured against the hedge ratio, mean and sd FROZEN at entry; exit when that anchored z reverts inside +-z_exit, breaks beyond z_stop, or after max_hold_days. Market orders at the signal bar's close (24/7 venue: equivalent to next bar's open).

Iteration 3 addresses 'per-trade edge below costs' (0.046% per leg). The previous exit used the ROLLING z-score, whose window contains the dislocation itself; the rolling mean drifted toward price so z re-entered the exit band largely because the equilibrium moved, closing trades after little actual spread convergence. The exit/stop are now judged against the OU equilibrium frozen at entry (beta, residual mean, residual sd kept as restart-safe scalars), so each closed reversion trade captures roughly (z_entry - z_exit) entry-time residual sds of real price spread, which should lift per-trade return well above the round-trip cost. Entry rule, parameters, sizing, legs and config are unchanged; z_entry/z_exit/z_lookback remain the tunable signal parameters for optimization.

Verification Results

Optional: rebuild _pos_dir, the entry timestamp and the frozen beta/mean/sd from the open positions on restart, instead of flattening.

should_exit returns True when _pos_dir == 0, which flattens any position that has no strategy state after a paper restart. Positions stay unprotected only if the scalars are not restored. This does not affect the backtest.

Optional: check self.legs_balanced() and log or flatten when the hedge did not fill.

If _hedge_size() returns 0, _submit_entry returns early after the primary order: no hedge is placed, _pos_dir stays 0, and the next bar's should_exit flattens the primary. That leaves one bar of naked LTC exposure. position_size already refuses the entry when the hedge would be below min notional, so this path should be practically unreachable. A hedge order refused by the venue is not detected either.

The sandbox logged 316 fills over about 2400 daily bars. Counting both legs, that is roughly 24 round trips a year, against the 10-14 the hypothesis expects. With z_entry=1.75 on a 90-day rolling window the band is not very wide. The analyst should check on the full history that the average capture per round trip (the frozen-equilibrium exit should raise it versus iteration 2's 0.046% per leg) clears about 0.10% round-trip taker fees on each of the two legs, and should check how many trades end at the 45-day time stop or the z_stop rather than by reversion. If time-stop exits dominate, the spread is not mean-reverting on this horizon.

Backtest Review

Sharpe

0.16

Total return

5.71%

Max drawdown

31.20%

Trades

170

Win rate

52.4%

Profit factor

1.02

The ledger shows 71 complete opposing-leg episodes, approximately 13 per year, consistent with the intended frequency.

Closed-position PnL totals +5,311.72 USDT; portfolio return is positive at 4.78%.

Mean net leg return is -0.673%, below the 0.15% venue floor. Leg profit factor is 1.019; paired-episode profit factor is 1.046, with 33 winning episodes out of 71.

The frozen-equilibrium repair is already implemented but has not restored adequate per-trade economics. QA's economic concern is borne out; its excessive-frequency concern is not.

Median holding is 9 days; only 9 of 71 episodes last at least 45 days. Time stops do not dominate. Other exit reasons cannot be determined from this ledger.

Positive dollar PnL and negative average notional return reflect different weighting. Neither supports claiming negative aggregate portfolio PnL.

Analysis

Per-trade edge below costs: avg_trade_return_pct 0.046% vs the 0.15% fee floor over 110 trades. Optimization cannot fix this; change the mechanism, holding period or timeframe so the per-trade return clears the floor, or recommend abandoning. Go ahead with optimization. Suggested plan: tune z_entry, z_exit and z_lookback. These are the three declared signal parameters, and the right entry/exit band and lookback window are genuinely uncertain. Keep z_stop, max_hold_days, per_leg_pct and the beta clamps fixed. Note for the gate: the unweighted avg_trade_return_pct (0.046%) understates the net return per leg weighted by notional (1.21%). Judge the economics on the reconciled figure.

Outcome Summary

Matching the intended low trading frequency and freezing the exit equilibrium did not establish adequate per-trade economics for this LTC/BCH implementation.

After three iterations, backtest review returned abandon: mean net leg return was below the 0.15% venue floor, with leg profit factor 1.019 and paired-episode profit factor 1.046 despite the implemented frozen-equilibrium repair. Optimization and subsequent analyst and risk stages were not reached.

Trade daily LTC/BCH perpetual spread mean reversion with opposing legs, wide entry bands and a frozen entry equilibrium intended to capture multi-week convergence.

The backtest reported 5.71% total return, Sharpe 0.164, 170 closed leg positions and 31.20% maximum drawdown. Separately, the review reported 71 complete paired episodes, 4.78% portfolio return and a mean net leg return of -0.673%.