数据科学和机器学习（第 30 部分）：预测股票市场的幂对、卷积神经网络（CNN）、和递归神经网络（RNN）·进阶篇
「把 CNN 特征流改造成 RNN 时态输入」
上一步抽出的特征不能直接喂给普通全连接层，要先过 RNN 这一关。RNN 在扫这些特征时会顺着时间轴走，把前后步的依赖关系带进权重里，这和本系列之前用 CNN 那套只吃空间局部性的逻辑不一样。 关键改动在 Flatten 层：原 CNN 架构里它把 3D 张量压成 2D，而 RNN、LSTM、GRU 全都要（批次大小、时间步长、特征）这种 3D 输入。所以动手搭网络时，Flatten 必须删掉，否则维度对不上直接报 shape 错。 下面这段 Keras 式堆叠就是替换后的主干：先 1D 池化降采样，接 50 单元的 SimpleRNN，两次 Dropout 0.5 防过拟合，末端 softmax 出信号分类。外汇与贵金属行情高波动、高杠杆，此类模型仅给出概率倾向，实盘前务必在 MT5 历史数据上回测。
model.add(MaxPooling1D(pool_size=class="num">2)) model.add(SimpleRNN(class="num">50, activation=&class="macro">#x27;relu&class="macro">#x27;)) model.add(Dropout(class="num">0.5)) model.add(Dense(class="num">50, activation=&class="macro">#x27;relu&class="macro">#x27;)) model.add(Dropout(class="num">0.5)) model.add(Dense(units=len(np.unique(y)), activation=&class="macro">#x27;softmax&class="macro">#x27;)) # Softmax for binary classification(class="num">1 buy, class="num">0 sell signal)
把 RNN+CNN 模型塞进 MT5 实盘前要做的事
模型在 Python 端跑完 14 个 epoch 后，测试集准确率停在 54%，分类报告里类 0（卖出）precision 0.70 / recall 0.40，类 1（买入）precision 0.45 / recall 0.74，样本量 195。这个水平不算高，但作为简单 RNN 叠加 CNN 的基线已经能区分多空倾向。 训练好后必须存成 ONNX，同时把 StandardScaler 的 mean 和 scale 写成 bin 文件。MQL5 里用 #resource 把 onnx 和 scaler 参数挂成资源，在 OnInit 中初始化 CConvNet 与 StandardizationScaler，预测前先拿 scaler 归一化输入，再丢进 CNN 模型取信号和概率。 特征矩阵 x_data 按 3 列（Open、High、Low）和训练时相同的时间步长排布，类标识须升序（0 卖、1 买）。回测区间取 2020.01.01–2024.09.01，EA 挂在 4 小时图而非日线，因为日线新烛在收市才生成，会漏单；数据仍用 CopyRates 从日线取，不影响决策。该设置下 EA 盈利交易占比约 90%，但外汇/贵金属品种切换时此结果不必然复现，属高风险验证。 下面这段是 Python 训练与导出核心，以及 MQL5 资源挂载片段，逐行看便知衔接点。
model.summary() # Compile the model optimizer = Adam(learning_rate=class="num">0.0001) model.compile(optimizer=optimizer, loss=&class="macro">#x27;binary_crossentropy&class="macro">#x27;, metrics=[&class="macro">#x27;accuracy&class="macro">#x27;]) # Train the model early_stopping = EarlyStopping(monitor=&class="macro">#x27;val_loss&class="macro">#x27;, patience=class="num">10, restore_best_weights=True) history = model.fit(X_train, y_train_encoded, epochs=class="num">1000, batch_size=class="num">16, validation_split=class="num">0.2, callbacks=[early_stopping]) plt.figure(figsize=(class="num">7.5, class="num">6)) plt.plot(history.history[&class="macro">#x27;loss&class="macro">#x27;], label=&class="macro">#x27;Training Loss&class="macro">#x27;) plt.plot(history.history[&class="macro">#x27;val_loss&class="macro">#x27;], label=&class="macro">#x27;Validation Loss&class="macro">#x27;) plt.xlabel(&class="macro">#x27;Epochs&class="macro">#x27;) plt.ylabel(&class="macro">#x27;Loss&class="macro">#x27;) plt.title(&class="macro">#x27;Training Loss Curve&class="macro">#x27;) plt.legend() plt.savefig("training loss curve-rnn-cnn-clf.png") plt.show() # Evaluating the Trained Model y_pred = model.predict(X_test) classes_in_y = np.unique(y) y_pred_binary = classes_in_y[np.argmax(y_pred, axis=class="num">1)] # Confusion Matrix cm = confusion_matrix(y_test, y_pred_binary) sns.heatmap(cm, annot=True, fmt=&class="macro">#x27;d&class="macro">#x27;, cmap=&class="macro">#x27;Blues&class="macro">#x27;) plt.xlabel("Predicted Label") plt.ylabel("True Label") plt.title("Confusion Matrix") plt.savefig("confusion-matrix RNN + CNN.png") print("Classification Report\n", classification_report(y_test, y_pred_binary)) onnx_file_name = "rnn+cnn.TSLA.D1.onnx" spec = (tf.TensorSpec((None, time_step, X_train.shape[class="num">2]), tf.float16, name="class="kw">input"),) model.output_names = [&class="macro">#x27;outputs&class="macro">#x27;] onnx_model, _ = tf2onnx.convert.from_keras(model, input_signature=spec, opset=class="num">13) with open(onnx_file_name, "wb") as f: f.write(onnx_model.SerializeToString()) scaler.mean_.tofile(f"{onnx_file_name.replace(&class="macro">#x27;.onnx&class="macro">#x27;,&class="macro">#x27;&class="macro">#x27;)}.standard_scaler_mean.bin") scaler.scale_.tofile(f"{onnx_file_name.replace(&class="macro">#x27;.onnx&class="macro">#x27;,&class="macro">#x27;&class="macro">#x27;)}.standard_scaler_scale.bin") class="macro">#include <MALE5\Convolutional Neural Networks(CNNs)\ConvNet.mqh> class="macro">#include <MALE5\preprocessing.mqh> CConvNet cnn; StandardizationScaler *scaler; class="macro">#resource "\Files\rnn+cnn.TSLA.D1.onnx" as class="type">uchar onnx_model[] class="macro">#resource "\Files\rnn+cnn.TSLA.D1.standard_scaler_mean.bin" as class="type">class="kw">double standardization_mean[]
◍ 把CNN信号接进D1实盘前要先过标准化
在 MT5 里加载训练好的 CNN 模型，第一道坎不是预测，而是把盘口数据先过一遍与训练时一致的标准化。代码里用 #resource 把 rnn+cnn.TSLA.D1.standard_scaler_scale.bin 读成 standardization_std[]，再 new StandardizationScaler(standardization_mean, standardization_std) 把均值和标准差灌进缩放器，这一步漏掉，模型输入分布偏移，输出信号会直接失真。
新 K 线开盘才触发推理：NewBar() 为真时 CopyRates(Symbol(), PERIOD_D1, 1, time_step, rates) 取最近 time_step=5 根日线，把每根 open/high/low 填进 x_data 这个 5×3 矩阵。注意这里只用了三个价格字段，训练时若含 close 或 vol 就得改列数，否则维度对不上会初始化失败。
填完矩阵后 x_data = scaler.transform(x_data) 做归一，随后 cnn.predict_bin() 出 0/1 信号、cnn.predict_proba() 出各类概率，用 Comment 打到图表上方便肉眼核对。实盘里 signal==1 时先 ClosePos(POSITION_TYPE_SELL) 平空，再检查无多单后 m_trade.Buy(min_lot, ...) 开最小手数多单——外汇与贵金属杠杆高，这类基于历史分布的模型信号仅代表概率倾向，实盘前务必用策略测试器跑样本外验证。
class="macro">#resource "\\Files\\rnn+cnn.TSLA.D1.standard_scaler_scale.bin" as class="type">class="kw">double standardization_std[] class="type">int OnInit() { class=class="str">"cmt">//--- if (!cnn.Init(onnx_model)) class=class="str">"cmt">//Initialize the Convolutional neural network class="kw">return INIT_FAILED; scaler = new StandardizationScaler(standardization_mean, standardization_std); class=class="str">"cmt">//Initialize the saved scaler by populating it with values ... ... class="kw">return (INIT_SUCCEEDED); } if (NewBar()) class=class="str">"cmt">//Trade at the opening of a new candle { CopyRates(Symbol(), PERIOD_D1, class="num">1, time_step, rates); for (class="type">ulong i=class="num">0; i<x_data.Rows(); i++) { x_data[i][class="num">0] = rates[i].open; x_data[i][class="num">1] = rates[i].high; x_data[i][class="num">2] = rates[i].low; } class=class="str">"cmt">//--- x_data = scaler.transform(x_data); class=class="str">"cmt">//Normalize the data class="type">int signal = cnn.predict_bin(x_data, classes_in_data_); class=class="str">"cmt">//getting a trading signal from the RNN model vector probabilities = cnn.predict_proba(x_data); class=class="str">"cmt">//probability for each class Comment("Probability = ",probabilities,"\nSignal = ",signal); class="kw">input class="type">int time_step = class="num">5; class="kw">input class="type">int magic_number = class="num">24092024; class="kw">input class="type">int slippage = class="num">100; class="type">MqlRates rates[]; matrix x_data(time_step, class="num">3); class=class="str">"cmt">//class="num">3 columns for open, high and low vector classes_in_data_ = {class="num">0, class="num">1}; class=class="str">"cmt">//unique target variables as they are in the target variable in your training data class="type">int OldNumBars = class="num">0; class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Expert initialization function | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">int OnInit() { class=class="str">"cmt">//--- class="type">class="kw">double min_lot = SymbolInfoDouble(Symbol(), SYMBOL_VOLUME_MIN); class="type">MqlTick ticks; SymbolInfoTick(Symbol(), ticks); if (signal==class="num">1) class=class="str">"cmt">//if the signal is bullish { ClosePos(POSITION_TYPE_SELL); class=class="str">"cmt">//close sell trades when the signal is buy if (!PosExists(POSITION_TYPE_BUY)) class=class="str">"cmt">//There are no buy positions { if (!m_trade.Buy(min_lot, Symbol(), ticks.ask, class="num">0 , class="num">0)) class=class="str">"cmt">//Open a buy trade
「空头信号下的平仓与反向开仓逻辑」
当指标返回 signal==0 时，这段逻辑判定为空头信号，先平掉所有多单再决定是否开空。ClosePos(POSITION_TYPE_BUY) 会扫掉账户里全部买入持仓，避免多空对锁或逆势扛单。 紧接着用 PosExists(POSITION_TYPE_SELL) 检查是否已有卖单；若没有，才以 m_trade.Sell(min_lot, Symbol(), ticks.bid, 0, 0) 按当前 bid 价开最小手数空单。两个止损止盈参数传 0，意味着本段不挂硬性止损，实盘中外汇与贵金属波动剧烈，这种写法可能让单笔回撤失控。 若 Sell 返回失败，printf 会把 GetLastError() 的错误码打进日志，方便在 MT5 终端直接排查 4109（交易被禁）或 4756（报价过期）之类问题。建议把这段粘进 EA 的 OnTick 末尾，用欧美 5 分钟图跑一周模拟盘验证信号翻转时的成交表现。
printf("Failed to open a buy position err=%d",GetLastError()); } } else if (signal==class="num">0) class=class="str">"cmt">//Bearish signal { ClosePos(POSITION_TYPE_BUY); class=class="str">"cmt">//close all buy trades when the signal is sell if (!PosExists(POSITION_TYPE_SELL)) class=class="str">"cmt">//There are no Sell positions { if (!m_trade.Sell(min_lot, Symbol(), ticks.bid, class="num">0 , class="num">0)) class=class="str">"cmt">//open a sell trade printf("Failed to open a sell position err=%d",GetLastError()); } } else class=class="str">"cmt">//There was an error class="kw">return;