Leeks Reaper Port (FMZ #34388 by Zero, after richox/okcoin-leeks-reaper): Price-Burst Chase + Gradual Rebalance-to-Neutral, Long-Short on BTCUSD_PERP.BINANCE_CM 1H, Burst Threshold Rescaled from 280ms Sampling by sqrt(time)

Port of the FMZ strategy '移植 OKCoin 韭菜收割机' by Zero (a port of richox/okcoin-leeks-reaper). Credit: the strategy is by those authors; this is our own reimplementation of the rules, not their code. CLASSIFICATION: (a) concrete trading idea. Its two parts are: (1) TREND: when a depth-weighted mid price breaks out of the recent sample range by more than a small percentage threshold, go all-in in the burst direction ('chase rises, kill falls'), with size scaled down when recent traded volume is thin and further scaled down just after a previous burst trade; (2) BALANCE: whenever the book is away from the neutral 50% position (which for the author meant zero net exposure, because the account was balanced with borrowed coin), keep shaving small slices off it until it is back to neutral, so gains are banked and the end of the trend is not held. HONEST VIABILITY NOTE: the author of the original says it stopped working once OKCoin started charging fees. A literal port with the original numbers is fee-dead here and cannot be backtested: the burst threshold is 0.005% of price (5e-5), checked every 280 ms, against a ~0.10% taker round trip. It also needs 3-level order-book snapshots, which we do not have as multi-year history. The user asked for a faithful port but also allowed adapting a non-viable idea, so ONLY TIME SCALE and VENUE are changed. Every rule is kept (burst test shape, 0.7/0.3 volume EMA, volume scaling, 0.8x/0.64x post-trade damping, all-in sizing, flip on opposite burst, 0.48/0.52 neutral band, rebalance in small clips). The one numeric change is to the threshold, and it follows a stated rule rather than a fit: a random-walk move grows with the square root of time, so a 5e-5 move per 280 ms sample becomes 5e-5 * sqrt(3600/0.28) = 0.567% per 1-hour sample. The volume threshold (10 BTC per 280 ms poll on 2016 OKCoin) has no meaning for a 2026 contract venue, so it becomes a relative test: the volume EMA compared with its own trailing 30-day median. VENUE: BTCUSD_PERP.BINANCE_CM (inverse COIN-M perp, 2020-08 to now on 1H, about 6 years). The original is a BTC strategy, which is why BTC is kept even though BTC is already heavily used in the corpus. Putting it on COIN-M moves the run off the 68%-share BINANCE USD-M bucket into the under-quota COIN-M bucket (3.1% vs the >=5% target). The inverse BTC-margined account also matches the original's coin-plus-cash balanced account, where the neutral point meant no fiat P&L swing. The trigger uses price and volume only, with no supplementary feed, so the lesson-160 COIN-M problem (no funding/OI collector) does not apply. Flat state: the strategy is flat whenever the rebalance has brought exposure back inside the neutral band. Expected time in market is ~35-45% (about 2-4 bursts a week, each unwound over about 24 hours).

Iteration 2 fixes the QA critical: sizing used the static USDT balance, which never moves on an inverse account because PnL and fees settle in BTC. The result was constant ~$100k notional, 187% average exposure and liquidation. Sizing now uses mark-to-market USD equity = stable balance + (BTC balance - BTC baseline captured on the first bar + unrealized BTC PnL) * close. Contracts are hard-capped at equity*alloc_frac*leverage/multiplier, which is <=1x equity at leverage 1, so exposure shrinks as equity falls. The neutral band is corrected to 0.02 (the 0.48/0.52 band is +/-2% of assets). After a restart the damping clock is rebuilt from the open position's ts_opened, and the docstring is corrected. The burst threshold is kept at the hypothesis's pre-stated sqrt(time) value of 0.567% rather than fitted, so burst frequency stays whatever that rule gives. I agree with QA's concern that on 1H BTC this threshold sits near ordinary hourly noise (~2 bursts/day, not 2-4/week). If the corrected backtest confirms fee drag, the honest reading is that the sqrt-time rescaling does not preserve the original's 'rare burst' character. Signal logic, polarity and the rebalance mechanics are unchanged.

Verification failed (Layer 4 — QA review) [class=code_defect]: - [edge_concern] The strategy fires far more often than the hypothesis expects. It stated 2-4 bursts a week and 35-45% time in market; the sandbox shows 659 positions in 342 days (~2 a day), 88.65% exposure, and commission equal to 22.56% of gross (every 1/24 rebalance clip pays the taker fee). PF 0.75 over 659 trades is close to the lesson-125 polarity-suspicion zone. The polarity checks out (a break above the prior max is a BUY), so the likelier explanation is that the sqrt(time)-rescaled 0.567% threshold sits inside ordinary 1H BTC noise (hourly sigma is about 0.4-0.7%). That turns the 'burst' into a noise trigger with a 24-clip fee tail. After fixing the sizing, check how often bursts fire per week against the hypothesis's 2-4 before spending the optimizer on it. - [critical] The 'all-in 1x equity' sizing is not equity-relative on this venue. _target_contracts() sizes as get_account_equity() * alloc_frac * leverage. On BINANCE_CM, resolve_account_equity() returns the venue's USDT balance, which the runner seeds at 100,000 USDT. Inverse PnL and commissions settle in BTC (the runner also seeds 50 BTC for that), so the USDT balance never moves. Every burst and every refill therefore sizes a fixed ~$100k notional (1000 contracts), no matter how much has been lost. The sandbox shows the result: avg_position_pct 187% on a configured leverage of 1.0, liquidated=true on 2025-05-01, total_return -100%, DD 100%. The hypothesis specifies all-in at 1x equity with post-trade damping. What actually ran is a constant-notional book whose real leverage rises as equity falls, with no stop. That is the compounding-blowup pattern from lessons 148/156, and it breaks the stated sizing. (line 131) — fix: Size from mark-to-market equity in USD. For example: equity_usd = USDT balance + (BTC balance - seeded BTC) * close, or track realized plus unrealized PnL of the inverse position in USD inside the strategy (pos.unrealized_pnl / realized_pnl are in BTC, so multiply by close). Clamp the target so contracts * multiplier <= equity_usd * alloc_frac * leverage, and assert it is <= 1.0x equity at leverage 1. Check the sandbox avg_position_pct afterwards: it should be <= ~100%, with no liquidation. - [warning] The neutral band is 4% of equity (neutral_band_frac=0.04). The original's 0.48/0.52 band is +/-2% of total assets around the 50% ratio. Measured as notional distance from neutral that is 2%, not 4%, unless the 'all-in' leg is defined as 50% of assets. The effect is small (the book flattens one or two clips early), but the mapping should be stated or corrected. (line 176) — fix: Use 0.02, or document why the full band width is used. - [warning] Every same-direction burst calls _mark_burst(), which resets _last_burst_ts and sets _entry_contracts to max(target, held). Refills and damping therefore restart on every consecutive burst bar. This matches the original's per-tick behaviour, but at the sandbox burst frequency (~2 a day) it keeps the book near full size most of the time (88.65% exposure) instead of unwinding over ~24h. (line 150) — fix: Behaviour follows the port, so no change is required. If the burst rate is recalibrated, re-check exposure against the hypothesis's 35-45%. - [warning] The docstring says per-position scalar state is 'restart-safe/persisted'. However, _entry_contracts, _pos_dir and _last_burst_ts are plain attributes. A live warmup replay runs calculate_signal only, so after a restart they are 0. The clip base then falls back to the held size, which is correct, but damping is lost for up to 10 bars. Minor. (line 60) — fix: Rebuild _last_burst_ts from the open position's ts_opened on start, or accept this and fix the docstring.