ianzeng123 FMZ Gold Multi-Confluence System on BTC USD-M

Iteration 2 fixes the QA defects. (1) Stop-loss and take-profit are now resting venue orders, placed as a market entry plus a linked STOP_MARKET/LIMIT bracket (one-updates-other) at close -/+ 1.5*ATR_smooth and close +/- 4.5*ATR_smooth. They now fill intrabar at their levels, as Pine strategy.exit does, instead of at the bar close. The trailing stop moves on the bar close, and modify_order shifts the resting stop leg, matching Pine's per-bar strategy.exit(stop=trail_stop) update. should_exit now handles only the structure-break strategy.close. (2) Sizing keeps the source's risk_qty = 1% equity / stop_distance but adds a notional cap that does not depend on ATR: min(max_notional_mult=2.5, leverage) * equity / price. Config leverage drops from 20 to 3. The source sizes without a cap, so this cap is a platform risk adaptation; it changes no fixed plan parameter. (3) min_bars_required rises to 2*ema_filter (400), so EMA200 and the other EMAs are primed from 400 buffered bars before trading. Smoothed RSI is now seeded with the first valid RSI, as Pine's custom ema() does. Pine's custom ema() also seeds with the first source value, so the close-seeded EMAs already match it. (4) A session_tz_offset_hours parameter (default 0 = UTC, Binance's exchange timezone) is exposed for the session windows. As in the source, the spread filter is still never applied and the BB squeeze is still always false. Caveat: the earlier sandbox's profit factor of 0.86 suggests the 7-of-10 confluence may have no edge on BTC 1h even with correct exits.

Verification failed (Layer 4 — QA review) [class=code_defect]: - [edge_concern] Even with stops fixed, the sandbox showed PF 0.86 over 1255 trades. The average win was only 1.54x the average loss, although the design targets 3R. Commissions were about 17% of gross. At an average of about 4x notional, the 0.10% round trip costs about 0.4% of equity per trade, against a 1% risk budget. After the stop fix, check whether the payoff ratio moves toward the 2-3R the design implies. If it does not, the 7-of-10 confluence has no edge on BTC 1h. - [critical] Stops, take-profits and trailing stops are detected from the bar's high/low but exited with a market order at the bar CLOSE (should_exit -> _submit_exit). In the Pine source, strategy.exit stop/limit orders fill intrabar at their price (process_orders_on_close only affects market orders). Two consequences follow. (a) A take-profit touched intrabar is exited at the close, which can be far below the target or even a loss, so the 3R winners the strategy depends on are cut short. The sandbox's average win/average loss ratio is 1.54, not about 3. (b) A stop that is crossed is exited at the close, which can be several stop distances beyond the stop. The hypothesis asks to preserve the source's code behaviour; this changes the exit mechanics that set both the payoff and the risk per trade. (line 236) — fix: When the entry fills (on_order_filled / on_position_opened), submit resting reduce-only orders at the computed levels: a stop_market at stop_loss and a limit at take_profit, OCO or cancelled on the other's fill. Replace the stop order when the trailing stop ratchets. Keep the bearish/bullish structure-break exit as a signal exit in should_exit(). - [critical] Sizing is risk_quantity = equity*1%/(1.5*ATR_smoothed), capped only by equity*leverage/price with leverage 20. The ATR denominator is inverse-volatility and shrinks in calm hours (L156). With 1h BTC ATR around 0.1-0.3%, the stop is about 0.15-0.45% and notional runs to 2-7x equity, up to the 20x clamp. The sandbox's average position was 395% of equity. The 1%-of-equity risk only holds if the stop fills at its price. Because exits fill at the close (issue above), one adverse hourly bar of about 3% on 6x notional loses about 18% of equity. Realised loss is therefore unbounded relative to the stated risk_per_trade. The sandbox was liquidated (-100%, liquidation 2026-07-08, max DD 100%), against the plan's own max_drawdown_pct constraint of 20%. This is a sizing/risk defect (L148/L156), not an edge outcome. (line 260) — fix: Enforce the stop with a resting stop order so that size*stop_distance is about 1% of equity. In addition, put an explicit notional cap independent of ATR, e.g. min(risk_quantity, max_notional_mult*equity/price) with max_notional_mult around 3. Lower config.leverage to match that cap rather than 20. - [warning] EMAs (including EMA200), the MACD EMAs and the signal line are seeded from the first close in the buffer, and min_bars_required is only 35. EMA200 and ema_trend are therefore unconverged for the first ~200-600 bars, and those early scores differ from Pine, which seeds ta.ema with an SMA. This only matters early in the window and after resets: the optimization plan uses warmup_bars 500, which mostly covers it. (line 66) — fix: Return max(..., ema_filter) or a larger value such as 3*ema_filter from min_bars_required, or seed each EMA with the SMA of its first N closes. - [warning] Session windows are evaluated in UTC. The FMZ Pine runtime may evaluate time(session) in the platform's exchange/chart timezone, which may not be UTC (FMZ commonly uses UTC+8). If so, each session window is shifted 8 hours. Session is one score point, so the effect is limited. (line 133) — fix: Confirm the source timezone and add a session_tz_offset_hours parameter if it is not UTC. - [warning] should_enter mutates entry_price, stop_loss, take_profit and trailing state before position_size runs. If the order is dropped for minimum notional or zero size, the state is stale. This is harmless today, because should_exit only runs with an open position and the next should_enter overwrites the state. It becomes a bug if fills differ from close (stops are anchored to the signal close, not to the fill price). (line 205) — fix: Recompute the stop/TP from the actual avg_px_open in on_position_opened.