🧠
基于LSTM的趋势预测在趋势跟踪策略中的应用·进阶篇
（2/3）· 从MT5抽数到Python训模再到EA回测，趋势过滤的实战链路一次打通
直接用当前ADX判断趋势强度，往往在高点才确认趋势已近尾声，入场就吃震荡。本篇改用LSTM预测未来ADX，提前过滤弱趋势信号，让趋势跟踪少接劣质单。
本章目录
◍ LSTM训练时的三个坑和ONNX导出
用LSTM做行情序列预测，核心是把数据切成（样本数, 时间步长, 特征数）的三维张量。比如 time_step=5，就是用周一到周五前5根K线的ADX、RSI、Stationary三个特征，去推第6个时间点的 output 值。模型自己学每个时间步、每个特征的权重，最后由全连接层吐出一个数。 训练前最容易被忽略的是顺序。数据必须按时间排，train_test_split 里 shuffle 一定要设 False，否则把未来样本混进训练块，回测看着漂亮，实盘就是前瞻性偏差的典型翻车。示例里 X_train, X_eval = train_test_split(X, y, test_size=0.2, shuffle=False) 这行才是能用的写法。 模型别堆太深。外汇和贵金属这种样本量有限的高风险品种，LSTM(50) 加一个 Dense(1) 跑 50 个 epoch、batch_size=100 已经够用。叠太多层和单元，验证集损失容易和训练集脱节，过拟合后参数偏移会很大。 评估阶段看 MSE。代码跑完会打印 Evaluation Loss: 57.405677795410156，这就是验证集预测值和真实值的均方误差。判断标准只有一条：验证损失和训练损失接近、且相对 output 均值平方不算离谱，模型才没白训。 要让 MT5 能调用，得先转 ONNX。LSTM 原生不支持直接导出，需先把模型改成函数式并显式定义输入输出，输入维度第一维设 None，这样任意批量样本都能喂进去逐条出预测，不用固定每次推理的样本数。
MQL5 / C++
class="kw">import numpy as np class="kw">import matplotlib.pyplot as plt from sklearn.model_selection class="kw">import train_test_split from tensorflow.keras.models class="kw">import Sequential from tensorflow.keras.layers class="kw">import LSTM, Dense # Assume data is your DataFrame already loaded with the specified columns and a time-based index # data.columns should include [&class="macro">#x27;ADX&class="macro">#x27;, &class="macro">#x27;RSI&class="macro">#x27;, &class="macro">#x27;Stationary&class="macro">#x27;, &class="macro">#x27;output&class="macro">#x27;] # --- Step class="num">1: Data Preparation --- time_step = class="num">5 # Select features and target features = [&class="macro">#x27;ADX&class="macro">#x27;, &class="macro">#x27;RSI&class="macro">#x27;, &class="macro">#x27;Stationary&class="macro">#x27;] target = &class="macro">#x27;output&class="macro">#x27; # --- Step class="num">2: Create sequences for LSTM input --- def create_sequences(data, target_col, time_step): """ Create sequences of length time_step from the DataFrame. data: DataFrame of input features and target. target_col: Name of the target column. Returns: X, y arrays suitable for LSTM. """ X, y = [], [] feature_cols = data.columns.drop(target_col) for i in range(len(data) - time_step): seq_x = data.iloc[i:i+time_step][feature_cols].values # predict target at the next time step after the sequence seq_y = data.iloc[i+time_step][target_col] X.append(seq_x) y.append(seq_y) class="kw">return np.array(X), np.array(y) # Create sequences X, y = create_sequences(data, target_col=target, time_step=time_step) # --- Step class="num">3: Split into training and evaluation sets --- # Use a simple class="num">80/class="num">20 split for training and evaluation X_train, X_eval, y_train, y_eval = train_test_split(X, y, test_size=class="num">0.2, shuffle=False) # --- Step class="num">4: Build the LSTM model --- n_features = len(features) # number of features per time step model = Sequential() model.add(LSTM(class="num">50, input_shape=(time_step, n_features))) # LSTM layer with class="num">50 units model.add(Dense(class="num">1)) # output layer for regression model.compile(optimizer=&class="macro">#x27;adam&class="macro">#x27;, loss=&class="macro">#x27;mse&class="macro">#x27;) model.summary() # --- Step class="num">5: Train the model --- epochs = class="num">50 batch_size = class="num">100 history = model.fit( X_train, y_train, epochs=epochs, batch_size=batch_size, validation_data=(X_eval, y_eval) ) # --- Step class="num">6: Evaluate the model --- eval_loss = model.evaluate(X_eval, y_eval) print(f"Evaluation Loss: {eval_loss}") # --- Step class="num">7: Generate Predictions and Plot --- # Generate predictions on the evaluation set predictions = model.predict(X_eval).flatten() # Create a plot for predictions vs actual values plt.figure(figsize=(class="num">12, class="num">6)) plt.plot(predictions, label=&class="macro">#x27;Predicted Output&class="macro">#x27;, class="type">color=&class="macro">#x27;red&class="macro">#x27;) plt.plot(y_eval, label=&class="macro">#x27;Actual Output&class="macro">#x27;, class="type">color=&class="macro">#x27;blue&class="macro">#x27;) plt.title(&class="macro">#x27;LSTM Predictions vs Actual Output&class="macro">#x27;) plt.xlabel(&class="macro">#x27;Sample Index&class="macro">#x27;) plt.ylabel(&class="macro">#x27;Output Value&class="macro">#x27;) plt.legend() plt.show() 评估亏损值：class="num">57.405677795410156 class="kw">import tensorflow as tf class="kw">import tf2onnx # Define the input shape based on your LSTM requirements: (time_step, n_features) time_step = class="num">5 n_features = class="num">3
把 Keras 时序模型导成 ONNX 给 MT5 用
在 MT5 里跑深度学习信号，第一步是把训练好的 Keras 时序模型转成 ONNX，这样绕开 Python 依赖、直接进推理引擎。下面这段是转换的核心代码，输入维度必须和训练时一致，否则转换会直接报错。 代码里 input_signature 的 TensorSpec 用了 (None, time_step, n_features)，None 对应批量维，time_step 和 n_features 得是你特征工程里定的真实值，比如 30 根 K 线、每根 5 个特征就是 (None,30,5)。opset=15 是 ONNX 算子集版本，低于 13 对动态形状支持会出问题。 转完落地的文件叫 regression2024.onnx，MT5 的 ONNX 模块加载它时路径要写绝对路径。外汇和贵金属波动受消息面扰动大，模型历史拟合不等于未来概率，实盘前务必用 2023 以后的样本外数据回测。
MQL5 / C++
inputs = tf.keras.Input(shape=(time_step, n_features), name="input") outputs = model(inputs) functional_model = tf.keras.Model(inputs=inputs, outputs=outputs) input_signature = ( tf.TensorSpec((None, time_step, n_features), dtype=tf.float32, name="input"), ) output_path = "regression2024.onnx" onnx_model, _ = tf2onnx.convert.from_keras( functional_model, input_signature=input_signature, # matching the input signature opset=class="num">15, output_path=output_path ) print(f"Model successfully converted to ONNX at {output_path}")
「把LSTM塞进金叉EA的骨架里」
把训练好的ONNX模型丢进 /MQL5/Files 之后，回到MetaEditor，我们直接在经典金叉趋势跟踪框架上改。快慢均线交叉跟趋势，价格穿过慢线离场给跟踪止损留空间——这套逻辑本身不新，但接下来要让LSTM在交叉信号出现时才跑一次预测，过滤掉弱趋势。 先声明全局变量：模型输入/输出维度参数、存数据的双维数组，以及一个模型句柄管加载和取结果。OnInit() 里除了初始化RSI、ADX，还要核对MQL5侧声明的维度跟Python训练时一致，否则32位精度对不上会直接初始化失败。 每根新K线触发更新函数，循环5个时间步把数据转成浮点塞进全局数组，顺序必须是历史在前、新数据追加在后。OnTick() 只在新K线跑检查，避免同根K线重复计算；当magic编号持仓全平，buypos和sellpos重置为0，防止已有仓位还开新单。 模型推理不在每根K线跑，只在MA交叉且无持仓时调用，把预测ADX值写进out数组。低于阈值视为趋势弱不开仓，高于才入场——回测里这能省掉大量无意义评估，XAUUSD 1h上用15/25周期快慢线配0.01手，外汇贵金属杠杆高，实盘前务必自己跑一遍验证阈值。
MQL5 / C++
class="macro">#include <Trade/Trade.mqh> class=class="str">"cmt">//XAU - 1h. CTrade trade; input class="type">int MaPeriodsFast = class="num">15; input class="type">int MaPeriodsSlow = class="num">25; input class="type">int MaPeriods = class="num">200; input class="type">class="kw">double lott = class="num">0.01; class="type">ulong buypos = class="num">0, sellpos = class="num">0; input class="type">int Magic = class="num">0; class="type">int barsTotal = class="num">0; class="type">int handleMaFast; class="type">int handleMaSlow; class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Expert initialization function | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">int OnInit() { trade.SetExpertMagicNumber(Magic); handleMaFast =iMA(_Symbol,PERIOD_CURRENT,MaPeriodsFast,class="num">0,MODE_SMA,PRICE_CLOSE); handleMaSlow =iMA(_Symbol,PERIOD_CURRENT,MaPeriodsSlow,class="num">0,MODE_SMA,PRICE_CLOSE); 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">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Expert tick function | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">void OnTick() { class="type">int bars = iBars(_Symbol,PERIOD_CURRENT); class=class="str">"cmt">//Beware, the last element of the buffer list is the most recent data, not [class="num">0] if (barsTotal!= bars){ barsTotal = bars; class="type">class="kw">double maFast[]; class="type">class="kw">double maSlow[]; CopyBuffer(handleMaFast,BASE_LINE,class="num">1,class="num">2,maFast); CopyBuffer(handleMaSlow,BASE_LINE,class="num">1,class="num">2,maSlow); class="type">class="kw">double bid = SymbolInfoDouble(_Symbol, SYMBOL_BID); class="type">class="kw">double ask = SymbolInfoDouble(_Symbol, SYMBOL_ASK);
◍ 持仓清理与交叉信号的执行顺序
这段逻辑里最容易被忽视的是平仓判断必须排在开仓之前。代码先用 iClose 取当前周期前一根收盘价 lastClose，只有当买持仓存在且收盘价跌破慢线 maSlow[1] 时才平多；卖持仓则是收盘价站上慢线才平空。顺序错了就可能在同根 K 线上既平又开，造成重复下单。 快慢线死叉与金叉的判定只看相邻两根：maFast[1]>maSlow[1] 且 maFast[0]<maSlow[0] 为死叉倾向平多开空，反之金叉倾向平空开多；同时要求 buypos==sellpos（都为 0 或同号）才调用 executeBuy / executeSell，避免已有同向仓时追单。 持仓票据有效性靠 OnTradeTransaction 回写。订单新增时若 Magic 匹配，BUY 单ticket 写入 buypos，SELL 单写入 sellpos；主循环里再用 PositionSelectByTicket 配合 POSITION_MAGIC 校验，票号失效或非本 EA 魔法则把 buypos/sellpos 归零，防止残票干扰下一轮信号。 executeSell / executeBuy 里用 SymbolInfoDouble 取实时买价并 NormalizeDouble 到品种小数位，再走 trade.Sell / trade.Buy，开仓后立刻用 trade.ResultOrder 拿订单号回填。外汇与贵金属杠杆高，这类自动平仓逻辑若魔数设错可能误平手动仓，上 MT5 前先开模拟盘跑一轮验证。
MQL5 / C++
class="type">class="kw">double lastClose = iClose(_Symbol, PERIOD_CURRENT, class="num">1); class=class="str">"cmt">//The order below matters if(buypos>class="num">0&& lastClose<maSlow[class="num">1]) trade.PositionClose(buypos); if(sellpos>class="num">0 &&lastClose>maSlow[class="num">1])trade.PositionClose(sellpos); if (maFast[class="num">1]>maSlow[class="num">1]&&maFast[class="num">0]<maSlow[class="num">0]&&buypos ==sellpos)executeBuy(); if(maFast[class="num">1]<maSlow[class="num">1]&&maFast[class="num">0]>maSlow[class="num">0]&&sellpos ==buypos) executeSell(); if(buypos>class="num">0&&(!PositionSelectByTicket(buypos)|| PositionGetInteger(POSITION_MAGIC) != Magic)){ buypos = class="num">0; } if(sellpos>class="num">0&&(!PositionSelectByTicket(sellpos)|| PositionGetInteger(POSITION_MAGIC) != Magic)){ sellpos = class="num">0; } } class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Expert trade transaction handling function | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">void OnTradeTransaction(const MqlTradeTransaction& trans, const class="type">MqlTradeRequest& request, const class="type">MqlTradeResult& result) { if (trans.type == TRADE_TRANSACTION_ORDER_ADD) { COrderInfo order; if (order.Select(trans.order)) { if (order.Magic() == Magic) { if (order.OrderType() == ORDER_TYPE_BUY) { buypos = order.Ticket(); } else if (order.OrderType() == ORDER_TYPE_SELL) { sellpos = order.Ticket(); } } } } } class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Execute sell trade function | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">void executeSell() { class="type">class="kw">double bid = SymbolInfoDouble(_Symbol, SYMBOL_BID); bid = NormalizeDouble(bid,_Digits); trade.Sell(lott,_Symbol,bid); sellpos = trade.ResultOrder(); } class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Execute buy trade function | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">void executeBuy() {
把 LSTM 模型塞进 EA 的初始化与取数
EA 启动时先把交易魔术码设好，再挂四条指标句柄：快慢 SMA 各一条，ADX 与 RSI 周期都写死 14。注意 ADX 默认取主线（缓冲区 0），和原文注释里的『-3』无关，那是别的写法残留，照抄容易误导。 ONNX 模型从编译进资源的 regression2024.onnx 读入，输入张量形状锁成 {1,5,3}、输出 {1,1}，也就是用最近 5 根 K 线的 3 个特征跑出 1 个预测值。若 OnnxSetInputShape 或 OnnxSetOutputShape 任一返回 false，直接打印错误码并释放句柄返回 -1，EA 起不来。 getData 里用 CopyBuffer 从 handleAdx、handleRsi 各拽 5 个值，偏移从 1 开始——意味着跳过当前未闭合的这根 K 线，只用已闭合的 5 根。外汇与贵金属杠杆高，模型信号只是概率倾向，真要挂单前建议先开 MT5 用策略测试器跑一遍这句取数逻辑看是否越界。
MQL5 / C++
class="type">class="kw">double ask = SymbolInfoDouble(_Symbol, SYMBOL_ASK); ask = NormalizeDouble(ask,_Digits); trade.Buy(lott,_Symbol,ask); buypos = trade.ResultOrder(); } class="macro">#resource "\\Files\\regression2024.onnx" as class="type">uchar lstm_onnx[] class="type">float data[class="num">1][class="num">5][class="num">3]; class="type">float out[class="num">1][class="num">1]; class="type">long lstmHandle = INVALID_HANDLE; const class="type">long input_shape[] = {class="num">1,class="num">5,class="num">3}; const class="type">long output_shape[]={class="num">1,class="num">1}; class="type">int handleMaFast; class="type">int handleMaSlow; class="type">int handleAdx; class=class="str">"cmt">// Average Directional Movement Index - class="num">3 class="type">int handleRsi; class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Expert initialization function | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">int OnInit() {class=class="str">"cmt">//Initialize model trade.SetExpertMagicNumber(Magic); handleMaFast =iMA(_Symbol,PERIOD_CURRENT,MaPeriodsFast,class="num">0,MODE_SMA,PRICE_CLOSE); handleMaSlow =iMA(_Symbol,PERIOD_CURRENT,MaPeriodsSlow,class="num">0,MODE_SMA,PRICE_CLOSE); handleAdx=iADX(_Symbol,PERIOD_CURRENT,class="num">14);class=class="str">"cmt">//Average Directional Movement Index - class="num">3 handleRsi = iRSI(_Symbol,PERIOD_CURRENT,class="num">14,PRICE_CLOSE); class=class="str">"cmt">// Load the ONNX model lstmHandle = OnnxCreateFromBuffer(lstm_onnx, ONNX_DEFAULT); class=class="str">"cmt">//--- specify the shape of the input data if(!OnnxSetInputShape(lstmHandle,class="num">0,input_shape)) { Print("OnnxSetInputShape failed, error ",GetLastError()); OnnxRelease(lstmHandle); class="kw">return(-class="num">1); } class=class="str">"cmt">//--- specify the shape of the output data if(!OnnxSetOutputShape(lstmHandle,class="num">0,output_shape)) { Print("OnnxSetOutputShape failed, error ",GetLastError()); OnnxRelease(lstmHandle); class="kw">return(-class="num">1); } if (lstmHandle == INVALID_HANDLE) { Print("Error creating model OnnxCreateFromBuffer ", GetLastError()); class="kw">return(INIT_FAILED); } class="kw">return(INIT_SUCCEEDED); } class="type">void getData(){ class="type">class="kw">double rsi[]; class="type">class="kw">double adx[]; CopyBuffer(handleAdx,class="num">0,class="num">1,class="num">5,adx); CopyBuffer(handleRsi,class="num">0,class="num">1,class="num">5,rsi);
让小布替你跑这套回测诊断
这些LSTM信号过滤与样本外回测的诊断维度，小布盯盘的AIGC已内置，打开对应品种页即可看到胜率和盈亏比的变化曲线，你只需核对参数边界。
常见问题
当前ADX高常意味着趋势发挥完毕，追入易接末尾震荡；预测未来值可提前识别强度上升段，过滤掉趋势不足时的信号。
目前小布内置的是通用趋势强度诊断与回测视图，自定义模型需先导出信号到MT5，再由小布做盘口层面的聚合呈现。
对小时级而言一年外样本能暴露过拟合，但是否充分取决于品种波动机制，建议交叉多个年份分段验证。
RSI本身有界不必强差分，但需检查其与原收益率的协整关系，避免引入伪平稳特征拖垮LSTM记忆单元。