51bitquant MartingleSpotStrategyV2: Faithful ETC Spot Drawdown-and-Bounce Accumulation

Implementation fix only; the trading rules are unchanged. Likely cause of the single 3033-day position: the take-profit SELL was sized from self.current_pos, a float summed fill by fill, and rounded down to the size increment, so it could leave one increment of ETC unsold. That leftover (<11 USDT) let the next cycle start on top of it, and the engine position never reached zero. Fixes: (1) The basket SELL now uses the engine's open position quantity (cache.positions_open(...)[0].quantity) as is, without rounding down, so a filled sell closes the position. I expect no leftover from fees because this CASH venue should charge commission in USDT, not ETC; I have not checked this. Any leftover from a partly filled sell that an addition later cancels is added to the next basket and sold with it. (2) On every bar, current_pos and avg_price are taken from the engine position (quantity, avg_px_open). The cycle state (avg_price, entry_lowest, last_entry_price, current_increase_pos_times) resets when the engine is flat. It also resets in on_order_filled when a SELL leaves less than the source's 11 USDT min_notional, matching the source's on_trade / new-cycle reset. (3) on_order_denied and on_order_rejected now log a warning and count refusals. Platform adaptations are unchanged: 1% of cycle-start equity replaces 1000 USDT, and orders are close-priced resting limits filled only by later market events. The source's trading_fee is used only for the diagnostic total_profit. The new code passes the static analyzer with no errors, but I did not run a backtest, so the expected count of closed cycles (hundreds) is a projection from the ~2,350 fills in the previous run, not a measurement.

IMPLEMENTATION FIX — the hypothesis is unchanged: make the code do what it already says. Do not add or change trading rules. Implementation fix only; do not change the hypothesis or any fixed contract value. The engine position for ETCUSDT.BINANCE_SPOT must go fully flat at the end of every accumulation cycle, so each cycle is reported as one closed position. In the backtest it opened 2018-06-21 and never closed (one 3033-day position, about 2,350 fills). Fixes: (1) Size the take-profit SELL from the engine's actual open position quantity (cache.positions_open(instrument_id)[0].quantity, net of any base-asset commission), not from self.current_pos. Make sure the quantity is not truncated below the held amount. If the held quantity is not a multiple of size_increment, sell the full held quantity; any sub-increment dust left after the fill must be flattened, or must not count as an open cycle. (2) On every bar and in on_order_filled, reconcile self.current_pos with the engine position quantity. After a SELL fill brings the position to zero or dust, reset avg_price, entry_lowest, last_entry_price and current_increase_pos_times, so the next cycle starts clean, as in the source. (3) Log or assert if a reduce-only SELL is denied or rejected. (4) Re-run and confirm: closed positions number in the hundreds or more, avg_holding is hours to weeks, and the final open position at the backtest end is at most one cycle's inventory. Then avg_trade_return_pct, PF and win rate can be judged against the 0.25% spot floor. ## Library refinements (from the knowledge library; test them, do not assume them) The library holds the exact MartingleSpotStrategyV2 source. Its rules support the analyst's fix: exit the whole basket from the real position and reset the cycle when flat. The author's own V3 and two FMZ DCA/grid design articles say the risky parts are no trend gate, no basket stop and no time exit, plus a win rate measured only on completed cycles. Refinements 2-5 change the rules, so they belong in a variant hypothesis, not in the faithful port. 1. [exit] Basket exit sized from the engine position, cycle reset on flat: Implementation only, rules unchanged. The take-profit SELL quantity is the engine's open position quantity, cache.positions_open(instrument_id)[0].quantity, not self.current_pos. Do not floor it below the held amount. On every bar and in on_order_filled, set current_pos = engine quantity and avg_price = the engine position's avg_px_open. When the position is flat, or holds less than 11 USDT of notional (the source's min_notional), reset avg_price=0, entry_lowest=0, last_entry_price=0, current_increase_pos_times=0 and clear the buy/sell order lists. Log any denied or rejected reduce-only SELL. Judge the re-run on mark-to-market equity, not on the win rate of completed cycles. — The backtest shows 1 position over 3033 days, 6 'trades' and a 100% win rate. Cycles never close, so per-trade metrics mean nothing. The source sells abs(current_pos) and runs the next cycle only when current_pos*close < min_notional (V2 on_bar/on_trade), so local and real holdings must match. D-Man V3 exits all filled layers as one basket at the exchange's average position price, to avoid 碎片持仓 and orphan orders. The Kelly-grid article freezes new risk when 持仓与本地账本不一致. It also warns that a completed-grid win rate (its example: 96.8%) hides open losses, so it measures the strategy on mark-to-market equity. (source: MartingleSpotStrategyV2 p.1; 用 Rust 在 FMZ 实现 D-Man V3 布林带反转DCA策略 p.1; 网格不该决定仓位：用分数凯利给只做多网格加一道资金预算 p.1) 2. [regime] Trend gate on new cycles only: Variant only. Start a new cycle (the initial BUY) only if, at the bar close, EMA(3000) >= EMA(12000) on 1m closes (about 50h and 200h) AND close >= EMA(12000). Additions and exits inside an open cycle are unchanged. Raise warmup_bars from 4320 to 12000. — V2 buys every 4% drop from the 2-day high whatever the trend. On ETC that buys into long bear legs and builds up to 8 layers of inventory. The long-only Kelly grid allows new grid entries only when FastEMA >= SlowEMA and Close >= SlowEMA. Its reasoning: 正向漂移都是待验证假设 (positive drift is an assumption to test, not a law). The anchor must not chase price down while a position is open. (source: 网格不该决定仓位：用分数凯利给只做多网格加一道资金预算 p.1) 3. [stop] Basket stop at 2% of cycle-start equity plus a time exit: Variant only. Exit the whole basket at the bar close if the unrealised loss is at least 2.0% of the equity captured at cycle start: (avg_price - close) * position_qty >= 0.02 * cycle_start_equity. Also exit if the cycle has been open more than 43200 bars (30 days). After a stop, wait 1440 bars before a new cycle can start. — Fully loaded, the cycle holds 1% * (1 + 1.3 + ... + 1.3^7) ≈ 23.9% of equity, and nothing ever exits it at a loss. This backtest's 19.5% max drawdown came from one inventory held for years. D-Man V3 exits the basket on a stop of 3 × ScaleRatio and after MaxHoldingBars. The Kelly grid has a latched RISK_HALT on a hard stop or an equity drawdown. A 2%-of-equity stop also meets the platform's risk-control option (a). (source: 用 Rust 在 FMZ 实现 D-Man V3 布林带反转DCA策略 p.1; 网格不该决定仓位：用分数凯利给只做多网格加一道资金预算 p.1) 4. [exit] Author's V3 trailing take-profit: Variant only. Replace the fixed 2% take-profit limit. Track entry_highest = max(high) since the last fill. Exit the basket at the close when close/avg_price - 1 >= 0.02 AND entry_highest/close - 1 >= 0.01. Keep exit_profit_pct=0.02 so each trade still clears the 0.25% spot floor plus about 0.2% round-trip spot fees. — In MartingleSpotStrategyV3, 51bitquant replaced V2's fixed take-profit with 利润超过1%，且最高价回调1%后平仓 (exit_profit_pct=0.01, exit_pull_back_pct=0.01), so winners can run past the target. D-Man V3 also uses a trailing take-profit, armed at 0.8 × ScaleRatio. We keep the 2% profit floor rather than V3's 1% because spot fees on our venue are 0.1% per side. (source: MartingleSpotStrategyV3 p.1; 用 Rust 在 FMZ 实现 D-Man V3 布林带反转DCA策略 p.1) 5. [parameter] Volatility-scaled add spacing with a cost floor, frozen per cycle: Variant only. At cycle start, freeze dump_down_pct = clamp(max(0.04, 3 * ATR(14 hourly, built from 840 1m bars)/close, 2*(0.001+0.0005)*2), 0.04, 0.12). Set bounce_back_pct = 0.25 * dump_down_pct, with a minimum of 0.01. Both stay fixed for the whole cycle. — A fixed 4% add step is tight when ETC is volatile, so all 7 additions fill early in a crash, and loose when it is calm. The Kelly grid sets spacing = max(MinGridPct, ATR/Price × AtrMultiplier, 2 × (fee+slippage) × CostFloorMultiple). It notes that filled layers must not have their targets moved by later ATR changes. D-Man V3 also scales its DCA distances by band half-width, clamped between 0.2% and 8%. (source: 网格不该决定仓位：用分数凯利给只做多网格加一道资金预算 p.1; 用 Rust 在 FMZ 实现 D-Man V3 布林带反转DCA策略 p.1)