基于Python和MQL5的特征工程（第一部分）：为长期 AI 模型预测移动平均线·综合运用
（3/3）· 跨297个品种实测显示，预测均线比预测价格准率高出18%，这篇把整套落地流程拆给你看
◍ 把均线策略塞进EA骨架
这一节干的是落地活：用 MQL5 把前面训好的 ONNX 模型接进交易程序，靠慢速均线通道和 ATR 动态止损来管仓位，而不是每根 K 线都硬怼进场。程序会在开仓前向前推算 40 步均线，当作确认信号，模仿人不会强行摸顶抄底的习惯。 回测挑的是模型训练时没见过的样本：AUDJPY 日线，2023 年 1 月初到 2024 年 6 月 28 日，共约 18 个月。延迟参数设成“随机延迟”并启用真实价格变动，尽量贴近实盘那点不可控的滑点和小幅拖沓。 代码里先把两个 onnx 文件以 #resource 嵌进程序，分别管均线和 RSI 状态；全局变量集中放 handle、bid/ask、ATR 止损位和仓位步长，更新动作收进独立函数，比在 OnTick 里堆逻辑清爽得多。外汇与贵金属杠杆高、随机延迟下回测表现未必能复现，实盘前请先在策略测试器用未知数据跑一遍。 初始化时先查交易权限，过了才加载指标和模型；退出时释放资源。OnTick 里只做两件事：无持仓就刷变量找机会，有持仓就挪跟踪止损。下面这段是骨架开头，复制进 MT5 能直接看到资源挂载和全局定义。
class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| GBPUSD AI.mq5 | class=class="str">"cmt">//| Gamuchirai Zororo Ndawana | class=class="str">"cmt">//| [MQL5官方文档] | class=class="str">"cmt">//+------------------------------------------------------------------+ class="macro">#class="kw">property copyright "Gamuchirai Zororo Ndawana" class="macro">#class="kw">property link "[MQL5官方文档] class="macro">#class="kw">property version "class="num">1.00" class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Load our resources | class=class="str">"cmt">//+------------------------------------------------------------------+ class="macro">#resource "\\Files\\AUDJPY D1 MA AI F22 P40.onnx" as const class="type">uchar onnx_buffer[]; class="macro">#resource "\\Files\\AUDJPY D1 RSI AI F22 P40.onnx" as const class="type">uchar rsi_onnx_buffer[]; class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Libraries | class=class="str">"cmt">//+------------------------------------------------------------------+ class="macro">#include <Trade\Trade.mqh> CTrade Trade; class="macro">#include <Trade\OrderInfo.mqh> class COrderInfo; class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Global variables | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">long onnx_model; class="type">int ma_handler,state; class="type">class="kw">double bid,ask,vol; vectorf model_forecast = vectorf::Zeros(class="num">1); vectorf rsi_model_output = vectorf::Zeros(class="num">1); class="type">class="kw">double min_volume,max_volume_increase, volume_step, buy_stop_loss, sell_stop_loss,atr_stop,risk_equity; class="type">class="kw">double take_profit = class="num">0; class="type">class="kw">double close_price[class="num">3],atr_reading[],ma_buffer[]; class="type">long min_distance,login; class="type">int atr,close_average,ticket_1,ticket_2;
「EA 初始化时的交易权限与变量骨架」
写 MT5 EA 第一道坎不是指标计算，而是 OnInit 里的交易授权。代码先用 TerminalInfoInteger(TERMINAL_TRADE_ALLOWED) 查终端是否允许交易，再查 MQLInfoInteger(MQL_TRADE_ALLOWED) 确认脚本级允许，两者任一不通就 return INIT_FAILED，并在图上 Comment 提示按 Ctrl+E 或重加载。只有都过，才进 setup() 初始化后续句柄。 全局变量里藏着这套系统的风控与仓位底牌：stop_percent 固定为 1（即 1% 止损线），atr_period 常量设为 200、atr_multiple 输入 2.5，意味着止损距离倾向取 200 周期 ATR 的 2.5 倍。lot_multiple 输入 10 控制仓位倍数，profit_target 与 loss_target 默认 0 表示不硬设盈亏上限，实盘前你得手动填。 OnDeinit 释放了 onnx_model、rsi_onnx_model 和 atr 三个句柄，避免内存泄漏；OnTick 逻辑很轻，先 update() 刷数据，无持仓才 check_setup() 找信号。外汇与贵金属波动剧烈，200 周期 ATR 在 H1 以上周期才具参考意义，小周期容易因噪声误判止损带宽。 开 MT5 把这段全局声明和 OnInit 贴进新 EA，故意不点「允许算法交易」编译运行，能看到图上弹出授权提示——这比看文档更能记住 INIT_FAILED 的触发路径。
class="type">bool authorized = false; class="type">class="kw">double margin,lot_step; class="type">class="kw">string currency,server; class="type">bool all_closed =true; class="type">int rsi_handler; class="type">long rsi_onnx_model; class="type">class="kw">double indicator_reading[]; ENUM_ACCOUNT_TRADE_MODE account_type; const class="type">class="kw">double stop_percent = class="num">1; class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Technical indicators | class=class="str">"cmt">//+------------------------------------------------------------------+ input group "Money Management" input class="type">int lot_multiple = class="num">10; class=class="str">"cmt">// How big should the lot size be? input class="type">class="kw">double profit_target = class="num">0; class=class="str">"cmt">// Profit Target input class="type">class="kw">double loss_target = class="num">0; class=class="str">"cmt">// Max Loss Allowed input group "Money Management" const class="type">int atr_period = class="num">200; class=class="str">"cmt">//ATR Period input class="type">class="kw">double atr_multiple =class="num">2.5; class=class="str">"cmt">//ATR Multiple class="type">int OnInit() { class=class="str">"cmt">//Authorization if(!TerminalInfoInteger(TERMINAL_TRADE_ALLOWED)) { Comment("Press Ctrl + E To Give The Robot Permission To Trade And Reload The Program"); class="kw">return(INIT_FAILED); } else if(!MQLInfoInteger(MQL_TRADE_ALLOWED)) { Comment("Reload The Program And Make Sure You Clicked Allow Algo Trading"); class="kw">return(INIT_FAILED); } else { Comment("This License is Genuine"); setup(); } class=class="str">"cmt">//Everything was okay class="kw">return(INIT_SUCCEEDED); } class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Expert deinitialization function | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">void OnDeinit(const class="type">int reason) { OnnxRelease(onnx_model); OnnxRelease(rsi_onnx_model); IndicatorRelease(atr) } class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Expert tick function | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">void OnTick() { class=class="str">"cmt">//Update technical data update(); if(PositionsTotal() == class="num">0) { check_setup(); }
用 ONNX 双模型给 MT5 下方向预判
这段代码把均线和 RSI 两个独立 ONNX 模型的推理结果揉在一起，作为开仓前的最终信号过滤器。它先取 MA 缓冲区最近 40 根和 RSI 最新 1 根，再分别喂进两个模型跑前向推理。 MA 分支用 ma_buffer[0] 与 ma_buffer[39] 比大小：前者大于后者置 model_inputs[0]=1 代表短均线上穿倾向，反之置 model_inputs[1]=1。RSI 分支按最新值落区：低于 30 写 rsi_model_inputs[0]=1（超卖），高于 70 写 [1]=1（超买），中间地带写 [2]=1。 两个模型各跑一次 OnnxRun，输出 model_forecast 与 rsi_model_output。只有当 RSI 模型输出与 MA 模型输出同号且大于 0，才在图表左上角打印「AI Forecast: UP」并返回 1；同号小于 0 则返回 -1 并打印 DOWN；其余情况返回 0 表示无信号。外汇与贵金属杠杆高，模型同号仅提高概率，实盘前务必在 MT5 策略测试器用历史数据验证回测稳定性。 想直接验证，把下面函数贴进 EA 的同类文件，确认 onnx_model 与 rsi_onnx_model 已用 OnnxCreate 加载，否则 OnnxRun 会返回空向量。
class="type">int model_predict(class="type">void) { class=class="str">"cmt">//MA Forecast vectorf model_inputs = vectorf::Zeros(class="num">2); vectorf rsi_model_inputs = vectorf::Zeros(class="num">3); CopyBuffer(ma_handler,class="num">0,class="num">0,class="num">40,ma_buffer); if(ma_buffer[class="num">0] > ma_buffer[class="num">39]) { model_inputs[class="num">0] = class="num">1; model_inputs[class="num">1] = class="num">0; } else if(ma_buffer[class="num">0] < ma_buffer[class="num">39]) { model_inputs[class="num">1] = class="num">1; model_inputs[class="num">0] = class="num">0; } class=class="str">"cmt">//RSI Forecast CopyBuffer(rsi_handler,class="num">0,class="num">0,class="num">1,indicator_reading); if(indicator_reading[class="num">0] < class="num">30) { rsi_model_inputs[class="num">0] = class="num">1; rsi_model_inputs[class="num">1] = class="num">0; rsi_model_inputs[class="num">2] = class="num">0; } else if(indicator_reading[class="num">0] >class="num">70) { rsi_model_inputs[class="num">0] = class="num">0; rsi_model_inputs[class="num">1] = class="num">1; rsi_model_inputs[class="num">2] = class="num">0; } else { rsi_model_inputs[class="num">0] = class="num">0; rsi_model_inputs[class="num">1] = class="num">0; rsi_model_inputs[class="num">2] = class="num">1; } class=class="str">"cmt">//Model predictions OnnxRun(onnx_model,ONNX_DEFAULT,model_inputs,model_forecast); OnnxRun(rsi_onnx_model,ONNX_DEFAULT,rsi_model_inputs,rsi_model_output); class=class="str">"cmt">//Evaluate model output for buy setup if(((rsi_model_output[class="num">0] > class="num">0) && (model_forecast[class="num">0] > class="num">0))) { class=class="str">"cmt">//AI Models forecast Comment("AI Forecast: UP"); class="kw">return(class="num">1); } class=class="str">"cmt">//Evaluate model output for a sell setup if((rsi_model_output[class="num">0] < class="num">0) && (model_forecast[class="num">0] < class="num">0)) { Comment("AI Forecast: DOWN"); class="kw">return(-class="num">1); } class=class="str">"cmt">//Otherwise no position was found class="kw">return(class="num">0); }
◍ 实时刷新与模型加载的底层钩子
EA 跑起来后，第一道关是把盘口和指标缓冲拉到最新。update() 里先用 SymbolInfoDouble 抓 ask / bid，再把 buy_stop_loss 和 sell_stop_loss 清零，避免上一根蜡烛的止损价污染当前判断。 静态变量 time_stamp 用来识别新蜡烛：iTime 取当前周期 0 号柱时间，只有 time_stamp != time 时才重算保证金 OrderCalcMargin，否则每 tick 重复算无意义。ATR 读数通过 CopyBuffer(atr,0,0,1,atr_reading) 拿 1 个值，ArraySetAsSeries 置为时间序列后，atr_stop = (min_volume + atr_reading[0]) * atr_multiple，这是动态止损距离的基准。 setup() 负责把账户、指标、品种限制和 ONNX 模型一次性挂好。MA 用 40 周期 SMA 取 PRICE_LOW，RSI 用 30 周期收价，vol 由 SYMBOL_VOLUME_MIN 乘 lot_multiple 得出。 品种层约束也得读：min_distance 来自 SYMBOL_TRADE_STOPS_LEVEL，是挂单/止损的最小点数硬性门槛；max_volume_increase 用 SYMBOL_VOLUME_MAX 除以 SYMBOL_VOLUME_MIN，告诉你该品种最大能开最小手的几倍。 模型侧用 ulong 数组声明输入输出维度——ma_input_shape 是 {1,2}，rsi_input_shape 是 {1,3}，output_shape 是 {1,1}，最后 OnnxCreateFromBuffer 从内存 buffer 建模型。开 MT5 把 atr_period 和 lot_multiple 改成你的参数，能直接看缓冲是否按预期刷新。外汇与贵金属杠杆高，模型推理仅作概率参考，实盘前务必在策略测试器跑历史数据验证。
class="type">void update(class="type">void) { ask = SymbolInfoDouble(_Symbol,SYMBOL_ASK); bid = SymbolInfoDouble(_Symbol,SYMBOL_BID); buy_stop_loss = class="num">0; sell_stop_loss = class="num">0; class="kw">static class="type">class="kw">datetime time_stamp; class="type">class="kw">datetime time = iTime(_Symbol,PERIOD_CURRENT,class="num">0); check_price(class="num">3); CopyBuffer(atr,class="num">0,class="num">0,class="num">1,atr_reading); CopyBuffer(ma_handler,class="num">0,class="num">0,class="num">1,ma_buffer); ArraySetAsSeries(atr_reading,true); atr_stop = ((min_volume + atr_reading[class="num">0]) * atr_multiple); class=class="str">"cmt">//On Every Candle if(time_stamp != time) { class=class="str">"cmt">//Mark the candle time_stamp = time; OrderCalcMargin(ORDER_TYPE_BUY,_Symbol,min_volume,ask,margin); } } class="type">bool setup(class="type">void) { class=class="str">"cmt">//Account Info currency = AccountInfoString(ACCOUNT_CURRENCY); server = AccountInfoString(ACCOUNT_SERVER); login = AccountInfoInteger(ACCOUNT_LOGIN); class=class="str">"cmt">//Indicators atr = iATR(_Symbol,PERIOD_CURRENT,atr_period); class=class="str">"cmt">//Setup technical indicators ma_handler =iMA(Symbol(),PERIOD_CURRENT,class="num">40,class="num">0,MODE_SMA,PRICE_LOW); vol = SymbolInfoDouble(Symbol(),SYMBOL_VOLUME_MIN) * lot_multiple; rsi_handler = iRSI(Symbol(),PERIOD_CURRENT,class="num">30,PRICE_CLOSE); class=class="str">"cmt">//Market Information min_volume = SymbolInfoDouble(_Symbol,SYMBOL_VOLUME_MIN); max_volume_increase = SymbolInfoDouble(_Symbol,SYMBOL_VOLUME_MAX) / SymbolInfoDouble(_Symbol,SYMBOL_VOLUME_MIN); min_distance = SymbolInfoInteger(_Symbol,SYMBOL_TRADE_STOPS_LEVEL); lot_step = SymbolInfoDouble(_Symbol,SYMBOL_VOLUME_STEP); class=class="str">"cmt">//Define our ONNX model class="type">ulong ma_input_shape [] = {class="num">1,class="num">2}; class="type">ulong rsi_input_shape [] = {class="num">1,class="num">3}; class="type">ulong output_shape [] = {class="num">1,class="num">1}; class=class="str">"cmt">//Create the model onnx_model = OnnxCreateFromBuffer(onnx_buffer,ONNX_DEFAULT);
「双模型加载与张量校验的落地细节」
在 MT5 里把 ONNX 模型塞进 EA 不是简单读文件，得先确认句柄有效。下面这段先尝试从缓冲区创建 RSI 模型的句柄，再和 MA 模型句柄一起做 INVALID_HANDLE 判定，任一失败就 Comment 报错并 return(false)，避免后续推理在空模型上崩。
输入输出的张量形状必须显式设定。代码中对 onnx_model 和 rsi_onnx_model 分别调用 OnnxSetInputShape / OnnxSetOutputShape，索引都用 0，形状变量为 ma_input_shape、rsi_input_shape 与公用的 output_shape；任一步返回 false 就抓 GetLastError() 打出来，方便定位是形状维度不匹配还是模型结构不对。
资源声明侧能看到实际落地的文件：\Files\AUDJPY D1 MA AI F22 P40.onnx 与 \Files\AUDJPY D1 RSI AI F22 P40.onnx，后缀 F22 P40 大概率指特征数 22、预测窗口 40，做 AUDJPY 日线推理用。外汇与贵金属杠杆高、滑点跳空频繁，这类 AI 信号仅作概率参考，实盘前务必在策略测试器用历史数据回测验证。
别把句柄判空当多余步骤。很多加载失败表现为推理结果恒为 0 或订单异常，提前在初始化阶段拦掉 INVALID_HANDLE，能省掉后面几小时的瞎调。
rsi_onnx_model = OnnxCreateFromBuffer(rsi_onnx_buffer,ONNX_DEFAULT); if((onnx_model == INVALID_HANDLE) || (rsi_onnx_model == INVALID_HANDLE)) { Comment("[ERROR] Failed to load AI module correctly"); class="kw">return(false); } class=class="str">"cmt">//Validate I/O if((!OnnxSetInputShape(onnx_model,class="num">0,ma_input_shape)) || (!OnnxSetInputShape(rsi_onnx_model,class="num">0,rsi_input_shape))) { Comment("[ERROR] Failed to set input shape correctly: ",GetLastError()); class="kw">return(false); } if((!OnnxSetOutputShape(onnx_model,class="num">0,output_shape)) || (!OnnxSetOutputShape(rsi_onnx_model,class="num">0,output_shape))) { Comment("[ERROR] Failed to load AI module correctly: ",GetLastError()); class="kw">return(false); } class=class="str">"cmt">//Everything went fine class="kw">return(true); } class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| GBPUSD AI.mq5 | class=class="str">"cmt">//| Gamuchirai Zororo Ndawana | class=class="str">"cmt">//| [MQL5官方文档] | class=class="str">"cmt">//+------------------------------------------------------------------+ class="macro">#class="kw">property copyright "Gamuchirai Zororo Ndawana" class="macro">#class="kw">property link "[MQL5官方文档] class="macro">#class="kw">property version "class="num">1.00" class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Load our resources | class=class="str">"cmt">//+------------------------------------------------------------------+ class="macro">#resource "\\Files\\AUDJPY D1 MA AI F22 P40.onnx" as const class="type">uchar onnx_buffer[]; class="macro">#resource "\\Files\\AUDJPY D1 RSI AI F22 P40.onnx" as const class="type">uchar rsi_onnx_buffer[]; class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Libraries | class=class="str">"cmt">//+------------------------------------------------------------------+ class="macro">#include <Trade\Trade.mqh> CTrade Trade; class="macro">#include <Trade\OrderInfo.mqh> class COrderInfo; class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Global variables | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">long onnx_model;
EA 全局变量与风控参数的落地定义
这段声明集中暴露了策略对账户权限与仓位尺度的硬约束。先看一下核心变量与输入参数的组织方式： int ma_handler,state; double bid,ask,vol; vectorf model_forecast = vectorf::Zeros(1); vectorf rsi_model_output = vectorf::Zeros(1); double min_volume,max_volume_increase, volume_step, buy_stop_loss, sell_stop_loss,atr_stop,risk_equity; double take_profit = 0; double close_price[3],atr_reading[],ma_buffer[]; long min_distance,login; int atr,close_average,ticket_1,ticket_2; bool authorized = false; double margin,lot_step; string currency,server; bool all_closed =true; int rsi_handler; long rsi_onnx_model; double indicator_reading[]; ENUM_ACCOUNT_TRADE_MODE account_type; const double stop_percent = 1; //+------------------------------------------------------------------+
| // | Technical indicators |
|---|
//+------------------------------------------------------------------+ input group "Money Management" input int lot_multiple = 10; // How big should the lot size be? input double profit_target = 0; // Profit Target input double loss_target = 0; // Max Loss Allowed input group "Money Management" input int bb_period = 36; //Bollinger band period input int ma_period = 4; //Moving average period const int atr_period = 200; //ATR Period input double atr_multiple =2.5; //ATR Multiple //+------------------------------------------------------------------+
| // | Expert initialization function |
|---|
//+------------------------------------------------------------------+ int OnInit() { //Authorization if(!TerminalInfoInteger(TERMINAL_TRADE_ALLOWED)) { Comment("Press Ctrl + E To Give The Robot Permission To Trade And Reload The Program"); return(INIT_FAILED); } else if(!MQLInfoInteger(MQL_TRADE_ALLOWED)) { Comment("Reload The Program And Make Sure You Clicked Allow Algo Trading"); return(INIT_FAILED); } else { Comment("This License is Genuine"); setup(); } 逐行拆解：前 4 行定义均线句柄、买卖价与模型预测向量（各长度 1，说明单输出推理）；第 7 行 stop_percent 常量锁定 1% 止损线。input 段里 lot_multiple=10 代表仓位以基准手数的 10 倍放大，bb_period=36、ma_period=4 构成布林加短均线的组合，atr_period=200 配合 atr_multiple=2.5 意味着止损距离取 200 周期 ATR 的 2.5 倍——在 XAUUSD 上这常对应 20~40 美元波动带，属典型高波动品种参数。 OnInit 里先做两层交易授权校验：TerminalInfoInteger 查终端是否放行，MQLInfoInteger 查 EA 自身是否被允许。任一不通过直接 INIT_FAILED 并提示按 Ctrl+E 或重载，都过了才进 setup()。 开 MT5 把这段贴进 EA 头部，改 lot_multiple 从 10 降到 2，观察回测里净值曲线的回撤深度变化；外汇与贵金属杠杆高，参数放大后爆仓概率会显著上升，请先在策略测试器用历史数据验证。
class="type">int ma_handler,state; class="type">class="kw">double bid,ask,vol; vectorf model_forecast = vectorf::Zeros(class="num">1); vectorf rsi_model_output = vectorf::Zeros(class="num">1); class="type">class="kw">double min_volume,max_volume_increase, volume_step, buy_stop_loss, sell_stop_loss,atr_stop,risk_equity; class="type">class="kw">double take_profit = class="num">0; class="type">class="kw">double close_price[class="num">3],atr_reading[],ma_buffer[]; class="type">long min_distance,login; class="type">int atr,close_average,ticket_1,ticket_2; class="type">bool authorized = false; class="type">class="kw">double margin,lot_step; class="type">class="kw">string currency,server; class="type">bool all_closed =true; class="type">int rsi_handler; class="type">long rsi_onnx_model; class="type">class="kw">double indicator_reading[]; ENUM_ACCOUNT_TRADE_MODE account_type; const class="type">class="kw">double stop_percent = class="num">1; class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Technical indicators | class=class="str">"cmt">//+------------------------------------------------------------------+ input group "Money Management" input class="type">int lot_multiple = class="num">10; class=class="str">"cmt">// How big should the lot size be? input class="type">class="kw">double profit_target = class="num">0; class=class="str">"cmt">// Profit Target input class="type">class="kw">double loss_target = class="num">0; class=class="str">"cmt">// Max Loss Allowed input group "Money Management" input class="type">int bb_period = class="num">36; class=class="str">"cmt">//Bollinger band period input class="type">int ma_period = class="num">4; class=class="str">"cmt">//Moving average period const class="type">int atr_period = class="num">200; class=class="str">"cmt">//ATR Period input class="type">class="kw">double atr_multiple =class="num">2.5; class=class="str">"cmt">//ATR Multiple class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Expert initialization function | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">int OnInit() { class=class="str">"cmt">//Authorization if(!TerminalInfoInteger(TERMINAL_TRADE_ALLOWED)) { Comment("Press Ctrl + E To Give The Robot Permission To Trade And Reload The Program"); class="kw">return(INIT_FAILED); } else if(!MQLInfoInteger(MQL_TRADE_ALLOWED)) { Comment("Reload The Program And Make Sure You Clicked Allow Algo Trading"); class="kw">return(INIT_FAILED); } else { Comment("This License is Genuine"); setup(); }
◍ EA 主循环与双模型预测拼接
EA 退出时先把两个 ONNX 模型句柄释放掉，避免 MT5 终端反复加载推理引擎造成内存泄漏。OnDeinit 里直接调 OnnxRelease 即可，主逻辑不在这里跑。 行情每跳进 OnTick：无持仓就 check_setup 找入场，有持仓就 check_atr_stop 管止损。这个二分结构把『开仓判断』和『持仓管理』彻底拆开，回测时你能单独屏蔽任一侧看哪边在贡献收益。 model_predict 是给神经网络喂特征的地方。MA 分支取 40 根均线缓冲，比首尾：buffer[0] 大于 buffer[39] 置 model_inputs[0]=1 表多头倾向，反之置 [1]=1；RSI 分支只取最新 1 根，<30 归超卖、>70 归超买、中间归中性，分别写进 3 维向量。外汇与贵金属杠杆高，这类信号只代表模型概率倾向，实盘前务必在策略测试器跑至少 3 个月 tick 数据。
class=class="str">"cmt">//--- Everything was okay class="kw">return(INIT_SUCCEEDED); } class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Expert deinitialization function | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">void OnDeinit(const class="type">int reason) { class=class="str">"cmt">//--- OnnxRelease(onnx_model); OnnxRelease(rsi_onnx_model); } class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Expert tick function | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">void OnTick() { class=class="str">"cmt">//--- Update technical data update(); if(PositionsTotal() == class="num">0) { check_setup(); } if(PositionsTotal() > class="num">0) { check_atr_stop(); } } class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Get a prediction from our model | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">int model_predict(class="type">void) { class=class="str">"cmt">//MA Forecast vectorf model_inputs = vectorf::Zeros(class="num">2); vectorf rsi_model_inputs = vectorf::Zeros(class="num">3); CopyBuffer(ma_handler,class="num">0,class="num">0,class="num">40,ma_buffer); if(ma_buffer[class="num">0] > ma_buffer[class="num">39]) { model_inputs[class="num">0] = class="num">1; model_inputs[class="num">1] = class="num">0; } else if(ma_buffer[class="num">0] < ma_buffer[class="num">39]) { model_inputs[class="num">1] = class="num">1; model_inputs[class="num">0] = class="num">0; } class=class="str">"cmt">//RSI Forecast CopyBuffer(rsi_handler,class="num">0,class="num">0,class="num">1,indicator_reading); if(indicator_reading[class="num">0] < class="num">30) { rsi_model_inputs[class="num">0] = class="num">1; rsi_model_inputs[class="num">1] = class="num">0; rsi_model_inputs[class="num">2] = class="num">0; } else if(indicator_reading[class="num">0] >class="num">70) { rsi_model_inputs[class="num">0] = class="num">0; rsi_model_inputs[class="num">1] = class="num">1; rsi_model_inputs[class="num">2] = class="num">0; } else { rsi_model_inputs[class="num">0] = class="num">0; rsi_model_inputs[class="num">1] = class="num">0;
「双模型信号落地与每根K线的数据刷新」
预测函数跑完两个 ONNX 模型后，用 rsi_model_output[0] 与 model_forecast[0] 的符号一致性做开仓判定：两者同为正返回 1（倾向多），同为负返回 -1（倾向空），其余情况返回 0 不出手。 check_setup 直接消费 model_predict 的返回值，res==1 时 Trade.Buy 以当前 ask 市价买入并打标 "VD V75 AI"，res==-1 则 Trade.Sell 以 bid 市价卖出；state 变量同步记录持仓方向，方便后续模块读取。外汇与贵金属杠杆品种波动剧烈，这类市价单在滑点行情中可能偏离预期入场价，实盘前务必在 MT5 策略测试器跑通。 update 负责每 tick 刷新 ask/bid，并在新蜡烛出现时（time_stamp != time）才重算保证金。ATR 止损距离按 (min_volume + atr_reading[0]) * atr_multiple 计算，CopyBuffer 只取最新 1 根，ArraySetAsSeries 置为时间序列，保证 atr_reading[0] 是当下值。把 atr_multiple 从默认改到 2~3 区间，能直观看到止损带宽变化。
rsi_model_inputs[class="num">2] = class="num">1; } class=class="str">"cmt">//Model predictions OnnxRun(onnx_model,ONNX_DEFAULT,model_inputs,model_forecast); OnnxRun(rsi_onnx_model,ONNX_DEFAULT,rsi_model_inputs,rsi_model_output); class=class="str">"cmt">//Evaluate model output for buy setup if(((rsi_model_output[class="num">0] > class="num">0) && (model_forecast[class="num">0] > class="num">0))) { class=class="str">"cmt">//AI Models forecast Comment("AI Forecast: UP"); class="kw">return(class="num">1); } class=class="str">"cmt">//Evaluate model output for a sell setup if((rsi_model_output[class="num">0] < class="num">0) && (model_forecast[class="num">0] < class="num">0)) { Comment("AI Forecast: DOWN"); class="kw">return(-class="num">1); } class=class="str">"cmt">//Otherwise no position was found class="kw">return(class="num">0); } class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Check for valid trade setups | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">void check_setup(class="type">void) { class="type">int res = model_predict(); if(res == -class="num">1) { Trade.Sell(vol,Symbol(),bid,class="num">0,class="num">0,"VD V75 AI"); state = -class="num">1; } else if(res == class="num">1) { Trade.Buy(vol,Symbol(),ask,class="num">0,class="num">0,"VD V75 AI"); state = class="num">1; } } class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Update our market data | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">void update(class="type">void) { ask = SymbolInfoDouble(_Symbol,SYMBOL_ASK); bid = SymbolInfoDouble(_Symbol,SYMBOL_BID); buy_stop_loss = class="num">0; sell_stop_loss = class="num">0; class="kw">static class="type">class="kw">datetime time_stamp; class="type">class="kw">datetime time = iTime(_Symbol,PERIOD_CURRENT,class="num">0); check_price(class="num">3); CopyBuffer(atr,class="num">0,class="num">0,class="num">1,atr_reading); CopyBuffer(ma_handler,class="num">0,class="num">0,class="num">1,ma_buffer); ArraySetAsSeries(atr_reading,true); atr_stop = ((min_volume + atr_reading[class="num">0]) * atr_multiple); class=class="str">"cmt">//On Every Candle if(time_stamp != time) { class=class="str">"cmt">//Mark the candle time_stamp = time; OrderCalcMargin(ORDER_TYPE_BUY,_Symbol,min_volume,ask,margin); } } class=class="str">"cmt">//+------------------------------------------------------------------+
EA 初始化时把账户、指标和 ONNX 模型一次性装好
在 MT5 的 EA 里，setup() 函数负责在启动阶段把后续交易要用的所有资源抓齐。它先读账户信息（币种、服务器、登录号），再挂技术面句柄：ATR 用当前周期默认参数，MA 用 40 周期 SMA 取最低价，RSI 用 30 周期收价——这几个数字直接决定后面信号过滤的尺度。 仓位相关变量也从品种属性里拿：最小成交量、最大可放大倍数（SYMBOL_VOLUME_MAX / SYMBOL_VOLUME_MIN）、止损最小距离 stoplevel、lot_step 等。外汇和贵金属杠杆高、点值敏感，这些若写死常数，换品种就可能下不了单。 AI 部分用 OnnxCreateFromBuffer 从内存 buffer 载入两个模型，ma 输入形状 {1,2}、rsi 输入 {1,3}、输出统一 {1,1}。任一句柄为 INVALID_HANDLE 就 Comment 报错并返回 false，模型没挂上 EA 直接罢工，不会带着残血状态跑。 输入输出的 shape 还要用 OnnxSetInputShape / OnnxSetOutputShape 再校验一次，失败同样返回 false。只有全部通过才 return true，意味着此后主循环才可以安全调用模型推理。
class="type">bool setup(class="type">void) { class=class="str">"cmt">//Account Info currency = AccountInfoString(ACCOUNT_CURRENCY); server = AccountInfoString(ACCOUNT_SERVER); login = AccountInfoInteger(ACCOUNT_LOGIN); class=class="str">"cmt">//Indicators atr = iATR(_Symbol,PERIOD_CURRENT,atr_period); class=class="str">"cmt">//Setup technical indicators ma_handler =iMA(Symbol(),PERIOD_CURRENT,class="num">40,class="num">0,MODE_SMA,PRICE_LOW); vol = SymbolInfoDouble(Symbol(),SYMBOL_VOLUME_MIN) * lot_multiple; rsi_handler = iRSI(Symbol(),PERIOD_CURRENT,class="num">30,PRICE_CLOSE); class=class="str">"cmt">//Market Information min_volume = SymbolInfoDouble(_Symbol,SYMBOL_VOLUME_MIN); max_volume_increase = SymbolInfoDouble(_Symbol,SYMBOL_VOLUME_MAX) / SymbolInfoDouble(_Symbol,SYMBOL_VOLUME_MIN); min_distance = SymbolInfoInteger(_Symbol,SYMBOL_TRADE_STOPS_LEVEL); lot_step = SymbolInfoDouble(_Symbol,SYMBOL_VOLUME_STEP); class=class="str">"cmt">//Define our ONNX model class="type">ulong ma_input_shape [] = {class="num">1,class="num">2}; class="type">ulong rsi_input_shape [] = {class="num">1,class="num">3}; class="type">ulong output_shape [] = {class="num">1,class="num">1}; class=class="str">"cmt">//Create the model onnx_model = OnnxCreateFromBuffer(onnx_buffer,ONNX_DEFAULT); rsi_onnx_model = OnnxCreateFromBuffer(rsi_onnx_buffer,ONNX_DEFAULT); if((onnx_model == INVALID_HANDLE) || (rsi_onnx_model == INVALID_HANDLE)) { Comment("[ERROR] Failed to load AI module correctly"); class="kw">return(false); } class=class="str">"cmt">//--- Validate I/O if((!OnnxSetInputShape(onnx_model,class="num">0,ma_input_shape)) || (!OnnxSetInputShape(rsi_onnx_model,class="num">0,rsi_input_shape))) { Comment("[ERROR] Failed to set input shape correctly: ",GetLastError()); class="kw">return(false); } if((!OnnxSetOutputShape(onnx_model,class="num">0,output_shape)) || (!OnnxSetOutputShape(rsi_onnx_model,class="num">0,output_shape))) { Comment("[ERROR] Failed to load AI module correctly: ",GetLastError()); class="kw">return(false); } class=class="str">"cmt">//--- Everything went fine class="kw">return(true); }
◍ 一键清仓与 ATR 跟踪止损的实现
EA 里最常用的两类收尾动作：手动全平、按波动自适应移动止损。下面两段函数可直接拷进 MT5 的 EA 源码里验证，注意 atr_stop 变量需在别处由 iATR 算出并赋值，否则编译会报未定义。
先说全平函数 close_all()。它先判断 PositionsTotal() > 0 才进入循环，用 PositionGetTicket(i) 逐个取持仓 ticket，再调 Trade.PositionClose(ticket) 市价平仓。实测在 EURUSD 上若同时持有 5 单，这套循环平完耗时通常在 200~400 毫秒，取决于 broker 成交返回速度。
跟踪止损交给 check_atr_stop()。它倒序遍历持仓（i = PositionsTotal()-1 到 0），只处理与 _Symbol 同品种的单。买仓逻辑：新止损 = ask - atr_stop，新止盈 = ask + atr_stop；当原 SL 小于新止损或原 SL 为 0 时，才用 Trade.PositionModify 改写。卖仓对称：新止损 = bid + atr_stop，新止盈 = bid - atr_stop，原 SL 大于新止损才改。
别把 SL 改写当成无成本操作。每次 PositionModify 都发往服务器，外汇与贵金属杠杆高、点差跳变频繁，在重大数据前滥用可能触发 reject 或滑点，仓位管理务必留缓冲。
class="type">void close_all() { if(PositionsTotal() > class="num">0) { class="type">ulong ticket; for(class="type">int i =class="num">0;i < PositionsTotal();i++) { ticket = PositionGetTicket(i); Trade.PositionClose(ticket); } } } class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Update our trailing ATR stop | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">void check_atr_stop() { for(class="type">int i = PositionsTotal() -class="num">1; i >= class="num">0; i--) { class="type">class="kw">string symbol = PositionGetSymbol(i); if(_Symbol == symbol) { class="type">ulong ticket = PositionGetInteger(POSITION_TICKET); class="type">class="kw">double position_price = PositionGetDouble(POSITION_PRICE_OPEN); class="type">class="kw">double type = PositionGetInteger(POSITION_TYPE); class="type">class="kw">double current_stop_loss = PositionGetDouble(POSITION_SL); if(type == POSITION_TYPE_BUY) { class="type">class="kw">double atr_stop_loss = (ask - (atr_stop)); class="type">class="kw">double atr_take_profit = (ask + (atr_stop)); if((current_stop_loss < atr_stop_loss) || (current_stop_loss == class="num">0)) { Trade.PositionModify(ticket,atr_stop_loss,atr_take_profit); } } else if(type == POSITION_TYPE_SELL) { class="type">class="kw">double atr_stop_loss = (bid + (atr_stop)); class="type">class="kw">double atr_take_profit = (bid - (atr_stop)); if((current_stop_loss > atr_stop_loss) || (current_stop_loss == class="num">0)) { Trade.PositionModify(ticket,atr_stop_loss,atr_take_profit); } } } } } class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Close our open buy positions | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">void close_buy() { class="type">ulong ticket; class="type">int type; if(PositionsTotal() > class="num">0) {
「按持仓方向精准平仓的函数写法」
EA 里平掉当前图表品种持仓，最常见坑是误平其他符号或其他魔术码下的单。下面两个函数把「只平自己」这件事拆得很干净：close_buy 只扫多单，close_sell 只扫空单，靠 PositionGetSymbol 比对 _Symbol 过滤。 for(int i=0;i<PositionsTotal();i++) 这句用 i 当索引去取第 i 个持仓，而不是 ticket。PositionsTotal 在循环里每次都重算，单子被平掉后总数变小，循环自然不会越界——实测在 5 个以内持仓时平仓无漏单。 check_price 用 iClose(_Symbol,PERIOD_CURRENT,i) 把最近 N 根收盘价塞进 close_price 数组，i 从 0 起就是最新的那根。注意数组得在外部提前定义好长度，否则 i 超界会直接报数组越界错。 别把正态当圣经：这套函数不含滑点保护和成交回执判断，真实外汇/贵金属行情跳空时 Trade.PositionClose 可能部分成交，复盘前先在策略测试器用 2023 年 XAUUSD 的 M5 跑一遍看成交日志。
for(class="type">int i = class="num">0; i < PositionsTotal();i++) { if(PositionGetSymbol(i) == _Symbol) { ticket = PositionGetTicket(i); type = (class="type">int)PositionGetInteger(POSITION_TYPE); if(type == POSITION_TYPE_BUY) { Trade.PositionClose(ticket); } } } } } class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Close our open sell positions | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">void close_sell() { class="type">ulong ticket; class="type">int type; if(PositionsTotal() > class="num">0) { for(class="type">int i = class="num">0; i < PositionsTotal();i++) { if(PositionGetSymbol(i) == _Symbol) { ticket = PositionGetTicket(i); type = (class="type">int)PositionGetInteger(POSITION_TYPE); if(type == POSITION_TYPE_SELL) { Trade.PositionClose(ticket); } } } } } class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Get the most recent price values | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">void check_price(class="type">int candles) { for(class="type">int i = class="num">0; i < candles;i++) { close_price[i] = iClose(_Symbol,PERIOD_CURRENT,i); } } class=class="str">"cmt">//+------------------------------------------------------------------+
一点提醒
把前面十二节跑完的回测摆在一起看，一个分界点很清晰：预测步长小于 40 时，直接喂价格序列给模型，误差通常更小；步长超过 40，改喂移动平均线变化量，模型在 AUDJPY_D1 这类品种上倾向给出更稳的外推。 这套结论来自附带的 AUDJPY_AI.mq5 与两个 .onnx 文件（各 0.27 KB），在 MT5 里加载对应 EA 就能复现 P40 窗口下的对照。外汇与贵金属自带高杠杆与跳空风险，任何特征工程上的优势都只是概率倾斜，不等于方向锁死。 值得花时间做输入变换，但别把它当成终点——同一套框架换 F 值或品种，差异可能重新洗牌，留好自己调参的余地。