股票交易中的非线性回归模型·综合运用
（3/3）· 当线性指标失效后，怎样把微分方程与自适应比率接进真实盘口并管住回撤
很多人把 RSI、MACD 这类七八十年代的线性工具直接套在分形特征明显的市场上，以为调调参数就能预测下一步。历史回测漂亮、实盘却越漂越远，往往不是行情变了，而是模型假设就错了。这一篇把前面两篇铺好的非线性思路收进可用系统，不再谈虚的。
从咖啡馆里的反向信号到自适应仓位
一套趋势模型曾在欧元兑美元看似平稳上扬时弹出空单信号，人为直觉觉得荒谬，但40分钟后欧元跌了35个点——模型抓到的是价格结构里的微变，肉眼很难察觉。这条经验指向一个铁律：别跟系统争辩，尤其外汇和贵金属这种高波动品种，人为情绪往往是亏损源头。 错过信号和内存故障乱开仓的事故，逼出一个轻量警报模块。关键在 _emergency_notification：遇到内存异常或连环开仓，系统发短信并自动停交易，等人工介入才恢复，避免账户被单边扫穿。 仓位管理上，固定0.1手像穿芭蕾鞋走钢丝。后来改成按余额和回撤自适应：基础风险取余额1%，回撤超5%风险减半，再叠ATR做波动率修正。_normalize_volume最麻烦——不同经纪商最小步长不同，有的允0.010手，有的只收整数手，得逐家配参数。 高波动日（美联储讲话、黑五之类）价格像醉酒乱撞。直接关系统太粗暴，改接经济日历API：重要事件前30分钟自动隐身、后30分钟重启，等于给资金加了道缓冲阀。
<span class="keyword">def</span> notify_signal(self, signal_type, message): <span class="keyword">try</span>: <span class="comment"># Format the message</span> timestamp = class="type">class="kw">datetime.now().strftime(<span class="class="type">class="kw">string">&class="macro">#x27;%Y-%m-%d %H:%M:%S&class="macro">#x27;</span>) formatted_msg = <span class="class="type">class="kw">string">f"[<span class="subst">{timestamp}</span>] <span class="subst">{signal_type}</span>: <span class="subst">{message}</span>"</span> <span class="comment"># Send to Telegram</span> <span class="keyword">if</span> self.use_telegram <span class="keyword">and</span> self.telegram_token: self.telegram_bot.send_message( chat_id=self.telegram_chat_id, text=formatted_msg, parse_mode=<span class="class="type">class="kw">string">&class="macro">#x27;HTML&class="macro">#x27;</span> ) <span class="comment"># Local logging</span> <span class="keyword">with</span> <span class="built_in">open</span>(self.log_file, <span class="class="type">class="kw">string">&class="macro">#x27;a&class="macro">#x27;</span>, encoding=<span class="class="type">class="kw">string">&class="macro">#x27;utf-class="num">8&class="macro">#x27;</span>) <span class="keyword">as</span> f: f.write(<span class="class="type">class="kw">string">f"<span class="subst">{formatted_msg}</span>\n"</span>) <span class="comment"># Check critical signals</span> <span class="keyword">if</span> signal_type <span class="keyword">in</span> [<span class="class="type">class="kw">string">&class="macro">#x27;ERROR&class="macro">#x27;</span>, <span class="class="type">class="kw">string">&class="macro">#x27;MARGIN_CALL&class="macro">#x27;</span>, <span class="class="type">class="kw">string">&class="macro">#x27;CRITICAL&class="macro">#x27;</span>]: self._emergency_notification(formatted_msg) <span class="keyword">except</span> Exception <span class="keyword">as</span> e: <span class="comment"># If the notification failed, send the message to the console at the very least</span> <span class="built_in">print</span>(<span class="class="type">class="kw">string">f"Error sending notification: <span class="subst">{str(e)}</span>\n<span class="subst">{formatted_msg}</span>"</span>) <span class="keyword">def</span> calculate_position_size(self): <span class="class="type">class="kw">string">"""Calculating the position size taking into account volatility and drawdown"""</span> <span class="keyword">try</span>: <span class="comment"># Take the total balance and the current drawdown</span> account_info = mt5.account_info() current_balance = account_info.balance drawdown = (account_info.equity / account_info.balance - <span class="number">class="num">1</span>) * <span class="number">class="num">100</span> <span class="comment"># Basic risk - class="num">1% of the deposit</span> base_risk = current_balance * <span class="number">class="num">0.01</span> <span class="comment"># Adjust for current drawdown</span> <span class="keyword">if</span> drawdown < -<span class="number">class="num">5</span>: <span class="comment"># If the drawdown exceeds class="num">5%</span> risk_factor = <span class="number">class="num">0.5</span> <span class="comment"># Slash the risk in half</span> <span class="keyword">else</span>: risk_factor = <span class="number">class="num">1</span> - <span class="built_in">abs</span>(drawdown) / <span class="number">class="num">10</span> <span class="comment"># Smooth decrease</span> <span class="comment"># Take into account the current ATR</span> atr = self.calculate_atr()
「仓位与市况的风控闸门」
这段逻辑把「下多少手」和「现在能不能做」拆成了两道独立闸门。前者先取点值，用 (base_risk * risk_factor) 除以 (atr * pip_value) 算出原始手数，再交给 _normalize_volume 按经纪商可开手数取整；异常时退回 0.1 手保底，并抛出错误信号。 市况检查更硬：先用 _is_high_impact_news_time 拦掉高影响新闻窗口；再用 ATR(5) 对比 ATR(20)，若短周期波动超过常态两倍直接放弃交易，并回传具体数值便于复盘。 最后一道是点差过滤，当前 spread 大于 max_allowed_spread 同样不放行。外汇与贵金属杠杆高，这类双重拦截能显著降低滑点与事件跳空的概率，建议直接把代码丢进 MT5 的 Python 桥或重写进 EA 实跑验证。
pip_value = self.get_pip_value() # Volume calculation rounded to available lots raw_volume = (base_risk * risk_factor) / (atr * pip_value) class="kw">return self._normalize_volume(raw_volume) except Exception as e: self.notify_signal(&class="macro">#x27;ERROR&class="macro">#x27;, f"Volume calculation error: {str(e)}") class="kw">return class="num">0.1 # Minimum safety volume def check_market_conditions(self): """Checking the market status before a deal""" # Check the calendar of events if self._is_high_impact_news_time(): class="kw">return False # Calculate volatility current_atr = self.calculate_atr(period=class="num">5) # Short period normal_atr = self.calculate_atr(period=class="num">20) # Normal period # Skip if the current volatility is class="num">2+ times higher than the norm if current_atr > normal_atr * class="num">2: self.notify_signal( &class="macro">#x27;INFO&class="macro">#x27;, f"Increased volatility: ATR(class="num">5)={current_atr:.5f}, " f"ATR(class="num">20)={normal_atr:.5f}" ) class="kw">return False # Check the spread current_spread = mt5.symbol_info(self.symbol).spread if current_spread > self.max_allowed_spread: class="kw">return False class="kw">return True
◍ 用 ATR 给止损留活路
早期用固定止损跑实盘，一周就被噪音扫掉一半仓位，市场专治各种想当然。后来改用平均真实波幅（ATR）动态定止损：波动大就放宽，平静期就收紧，系统对行情的适应能力明显上来。 入场逻辑尽量做减法。新手最容易犯的是信号频发——买、卖、再买来回打脸。加一道 15 分钟交易间隔和持仓过滤，这种可笑的简单改动反而救了回测曲线。 风险线只有一条死规矩：单笔亏损不超账户 1%。按 ATR 50 点算，舒适仓位是 0.2 手；欧盘 EURUSD 真实流动时表现最好，重要新闻前停手更划算，模型跟不上那种混乱。 下面这段开仓函数把 ATR 直接嵌进止损止盈：BUY 时 SL 放在 entry - atr，TP 只给 atr/3，赔率偏保守但触发逻辑清晰。 def open_position(self): try: atr = self.calculate_atr() # 取当前 ATR 值，用于动态间距 predicted_price = self.get_model_prediction() # 模型给出的预测价 current_price = mt5.symbol_info_tick(self.symbol).ask # 实时卖一价（做多参照） signal = "BUY" if predicted_price > current_price else "SELL" # 预测价高于现价看多，否则看空 if signal == "BUY": entry = mt5.symbol_info_tick(self.symbol).ask # 多单入场用 ask sl_level = entry - atr # 止损退 atr 点，随波动放大 tp_level = entry + (atr / 3) # 止盈仅 atr 三分之一，快进思路 else: entry = mt5.symbol_info_tick(self.symbol).bid # 空单入场用 bid sl_level = entry + atr tp_level = entry - (atr / 3) request = { "action": mt5.TRADE_ACTION_DEAL, # 市价成交 "symbol": self.symbol, "volume": self.lot_size, "type": mt5.ORDER_TYPE_BUY if signal == "BUY" else mt5.ORDER_TYPE_SELL, "price": entry, "sl": sl_level, "tp": tp_level, "deviation": 20, # 允许 20 点滑点 "magic": 234000, "comment": f"pred:{predicted_price:.6f}", # 备注写预测价便于复盘 "type_filling": mt5.ORDER_FILLING_FOK, # 成交或全撤 } result = mt5.order_send(request) if result.retcode != mt5.TRADE_RETCODE_DONE: raise ValueError(f"Error opening position: {result.retcode}") print(f"Position opened {signal}: price={entry:.5f}, SL={sl_level:.5f}, " f"TP={tp_level:.5f}, ATR={atr:.5f}") return result.order except Exception as e: print(f"Position opening failed: {str(e)}") return None 外汇与贵金属杠杆高，ATR 止损也只是降低噪音误杀概率，不等于控住所有跳空风险。把上面代码丢进 MT5 的 Python 环境，改 symbol 和 lot_size 就能看自己的品种怎么反应。
def open_position(self): try: atr = self.calculate_atr() predicted_price = self.get_model_prediction() current_price = mt5.symbol_info_tick(self.symbol).ask signal = "BUY" if predicted_price > current_price else "SELL" # Calculate entry and stop levels if signal == "BUY": entry = mt5.symbol_info_tick(self.symbol).ask sl_level = entry - atr tp_level = entry + (atr / class="num">3) else: entry = mt5.symbol_info_tick(self.symbol).bid sl_level = entry + atr tp_level = entry - (atr / class="num">3) # Send an order request = { "action": mt5.TRADE_ACTION_DEAL, "symbol": self.symbol, "volume": self.lot_size, "type": mt5.ORDER_TYPE_BUY if signal == "BUY" else mt5.ORDER_TYPE_SELL, "price": entry, "sl": sl_level, "tp": tp_level, "deviation": class="num">20, "magic": class="num">234000, "comment": f"pred:{predicted_price:.6f}", "type_filling": mt5.ORDER_FILLING_FOK, } result = mt5.order_send(request) if result.retcode != mt5.TRADE_RETCODE_DONE: raise ValueError(f"Error opening position: {result.retcode}") print(f"Position opened {signal}: price={entry:.5f}, SL={sl_level:.5f}, " f"TP={tp_level:.5f}, ATR={atr:.5f}") class="kw">return result.order except Exception as e: print(f"Position opening failed: {str(e)}") class="kw">return None
把数学交给Python，把下单交给MQL5
把整套数学框架硬塞进MQL5是个坑。Python配科学计算库做参数寻优顺手，MQL5跑交易逻辑才高效，没必要拿螺丝刀当锤子使。折中方案是用Python跑Nelder-Mead迭代挑出七个系数，MQL5只负责按系数算预测值和下单。 这七个系数就是模型核心：0.2752466、0.01058082、0.55162082、0.03687016、0.27721318、0.1483476、0.0008025，背后是数周优化和数千次迭代。预测方程几乎原样搬进MQL5，靠t-1、t-2两个价格点的线性、二次、差值、正弦项加常数拼出结果。 锁仓账户回测自2015年起收益约40%，回撤仅0.82%，月均超4%，期间累计成交量7800手、额外贡献约1.5%收益。测试用的最近一年数据做比率优化，这类系统最好零杠杆跑，收益略高于债基和美元存款就够。 别把正态当圣经 DCA时间加权仓位平均在这套EA里是核心机制，外汇和贵金属品种上风险极高。锁仓模式保守参数回测漂亮，但实盘可能一次性清空账户，碰这类策略先想清楚爆仓距离。 让小布替你跑这套 新K线才触发逻辑、目标盈利平掉所有仓、可同持多仓加量——这些都在三百行内写完。想换参数组合，Python重算后改g_coeffs数组即可，交易模块不用动。
class="type">class="kw">double g_coeffs[class="num">7] = {class="num">0.2752466, class="num">0.01058082, class="num">0.55162082, class="num">0.03687016, class="num">0.27721318, class="num">0.1483476, class="num">0.0008025}; class="type">class="kw">double GetPrediction(class="type">class="kw">double price_t1, class="type">class="kw">double price_t2) { class="kw">return g_coeffs[class="num">0] * price_t1 + class=class="str">"cmt">// Linear t-class="num">1 g_coeffs[class="num">1] * MathPow(price_t1, class="num">2) + class=class="str">"cmt">// Quadratic t-class="num">1 g_coeffs[class="num">2] * price_t2 + class=class="str">"cmt">// Linear t-class="num">2 g_coeffs[class="num">3] * MathPow(price_t2, class="num">2) + class=class="str">"cmt">// Quadratic t-class="num">2 g_coeffs[class="num">4] * (price_t1 - price_t2) + class=class="str">"cmt">// Price change g_coeffs[class="num">5] * MathSin(price_t1) + class=class="str">"cmt">// Cyclic g_coeffs[class="num">6]; class=class="str">"cmt">// Constant } class="type">void OpenPosition(class="type">bool buy_signal, class="type">class="kw">double lot) { class="type">MqlTradeRequest request; class="type">MqlTradeResult result; ZeroMemory(request); request.action = TRADE_ACTION_DEAL; request.symbol = Symbol(); request.volume = lot; request.type = buy_signal ? ORDER_TYPE_BUY : ORDER_TYPE_SELL; request.price = buy_signal ? SymbolInfoDouble(Symbol(), SYMBOL_ASK) : SymbolInfoDouble(Symbol(), SYMBOL_BID); class=class="str">"cmt">// ... other parameters } if(total_profit >= ProfitTarget) { CloseAllPositions(); class="kw">return; } class="type">bool isNewBar() { class="type">class="kw">datetime lastbar_time = class="type">class="kw">datetime(SeriesInfoInteger(Symbol(), PERIOD_CURRENT, SERIES_LASTBAR_DATE)); if(last_time == class="num">0) { last_time = lastbar_time; class="kw">return(false); } if(last_time != lastbar_time) { last_time = lastbar_time; class="kw">return(true); } class="kw">return(false); }
「最后一句大实话」
附件里那两个文件就是这套思路的落脚点：MarketSolver.py 是 12.61 KB 的 Python 选优与在线交易脚本，MarketSolver.mq5 是 16.62 KB 的 MQL5 实盘 EA，前者挑比率、后者跑交易，二者靠同一组比率参数咬合。 评论区有人拿原始 py 接了 Qt5 界面筛货币对做实验，也有人坚持零栏 tick 加第一根 K 线就够了——说明落地姿势本就多样，但都没脱离「选优+执行」这条主线。 系统永远没有完美态，外汇和贵金属又是高杠杆高风险场地，你能做的是让 EA 足够聪明去抓概率优势，又足够简单不在极端行情里自爆。深夜里写到这里，剩下的路，开 MT5 把那 16.62 KB 的 EA 拖进图表自己跑一遍最实在。