《数据科学与机器学习（第25部分）：使用循环神经网络（RNN）进行外汇时间序列预测》·进阶篇
（2/3）· 跳过基础概念，直接拆解简单RNN在MT5中的训练、特征筛选与模型导出细节
不少交易者把RNN当成黑箱直接套用现成脚本，却不知道序列数据怎么切、特征权重怎么看，导致回测漂亮实盘崩盘。本篇接上篇概念铺垫，继续深挖简单RNN在外汇预测里的具体工程实现。
「把滞后列清掉再切时序序列」
原始数据集有 28 列，其中 16 个带 'lag' 字样的滞后列（如 OPEN_LAG1~CLOSE_LAG3、DIFF_LAG1_OPEN 等）是为非时序模型准备的。RNN 本身能捕捉时间步内的依赖，这些滞后值反而冗余，必须整列 drop 掉，清完剩 12 列。 切分训练集和测试集时，比例取 70% / 30%。用 train_test_split 必须 shuffle=False，否则时序顺序被打乱，RNN 学到的就不是真实的价格演化路径。外汇与贵金属属高风险品种，顺序错乱会直接放大过拟合概率。 删掉 TARGET_CLOSE 和 TARGET_OPEN 两个目标列后，自变量剩 10 个特征。下面这段代码先抓 lag 列名再删除，然后不洗牌切分，最后用 create_sequences 按时间步长 7 转序列： [CODE] lagged_columns = [col for col in data.columns if "lag" in col.lower()] #let us obtain all the columns with the name lag print("lagged columns: ",lagged_columns) data = data.drop(columns=lagged_columns) #drop them lagged columns: ['OPEN_LAG1', 'HIGH_LAG1', 'LOW_LAG1', 'CLOSE_LAG1', 'OPEN_LAG2', 'HIGH_LAG2', 'LOW_LAG2', 'CLOSE_LAG2', 'OPEN_LAG3', 'HIGH_LAG3', 'LOW_LAG3', 'CLOSE_LAG3', 'DIFF_LAG1_OPEN', 'DIFF_LAG1_HIGH', 'DIFF_LAG1_LOW', 'DIFF_LAG1_CL # Split the data X = data.drop(columns=["TARGET_CLOSE","TARGET_OPEN"]) #dropping the target variables Y = data["TARGET_CLOSE"] test_size = 0.3 #70% of the data should be used for training purpose while the rest 30% should be used for testing x_train, x_test, y_train, y_test = train_test_split(X, Y, shuffle=False, test_size = test_size) # this is timeseries data so we don't shuffle print(f"x_train {x_train.shape}\nx_test {x_test.shape}\ny_train{y_train.shape}\ny_test{y_test.shape}") def create_sequences(X, Y, time_step): if len(X) != len(Y): raise ValueError("X and y must have the same length") X = np.array(X) Y = np.array(Y) Xs, Ys = [], [] for i in range(X.shape[0] - time_step): Xs.append(X[i:(i + time_step), :]) # Include all features with slicing Ys.append(Y[i + time_step]) return np.array(Xs), np.array(Ys) time_step = 7 #we consider the past 7 days from sklearn.preprocessing import StandardScaler scaler = StandardScaler() x_train = scaler.fit_transform(x_train) x_test = scaler.transform(x_test) x_train_seq, y_train_seq = create_sequences(x_train, y_train, time_step) x_test_seq, y_test_seq = create_sequences(x_test, y_test, time_step) print(f"Sequential data\n\nx_train {x_train_seq.shape}\nx_test {x_test_seq.shape}\ny_train{y_train_seq.shape}\ny_test{y_test_seq.shape}") Sequential data x_train (693, 7, 10) x_test (293, 7, 10) y_train(693,) y_test(293,) [/CODE] 逐行看：第1行用列表推导抓出所有含 'lag' 的列名；第3行 drop 掉它们。X 去掉两个目标列，Y 只留 TARGET_CLOSE。train_test_split 带 shuffle=False 是关键，不洗牌才能保住日线顺序。 create_sequences 逻辑：X、Y 先转 numpy 数组，for 循环从 0 跑到 len-时间步，每次取 X[i:i+7, :] 作输入窗口，对应的标签是 Y[i+7]。时间步设 7，等价于用过去 7 天日线信息预测第 8 天。 跑完得到 x_train 形状 (693, 7, 10)、x_test (293, 7, 10)——693 和 293 是序列样本数，7 是回溯天数，10 是特征数。开 MT5 导出日线 CSV 后，照这个 shape 核对自己的数据就能验证管道通不通。
lagged_columns = [col for col in data.columns if "lag" in col.lower()] class="macro">#let us obtain all the columns with the name lag print("lagged columns: ",lagged_columns) data = data.drop(columns=lagged_columns) class="macro">#drop them lagged columns: [&class="macro">#x27;OPEN_LAG1&class="macro">#x27;, &class="macro">#x27;HIGH_LAG1&class="macro">#x27;, &class="macro">#x27;LOW_LAG1&class="macro">#x27;, &class="macro">#x27;CLOSE_LAG1&class="macro">#x27;, &class="macro">#x27;OPEN_LAG2&class="macro">#x27;, &class="macro">#x27;HIGH_LAG2&class="macro">#x27;, &class="macro">#x27;LOW_LAG2&class="macro">#x27;, &class="macro">#x27;CLOSE_LAG2&class="macro">#x27;, &class="macro">#x27;OPEN_LAG3&class="macro">#x27;, &class="macro">#x27;HIGH_LAG3&class="macro">#x27;, &class="macro">#x27;LOW_LAG3&class="macro">#x27;, &class="macro">#x27;CLOSE_LAG3&class="macro">#x27;, &class="macro">#x27;DIFF_LAG1_OPEN&class="macro">#x27;, &class="macro">#x27;DIFF_LAG1_HIGH&class="macro">#x27;, &class="macro">#x27;DIFF_LAG1_LOW&class="macro">#x27;, &class="macro">#x27;DIFF_LAG1_CL # Split the data X = data.drop(columns=["TARGET_CLOSE","TARGET_OPEN"]) class="macro">#dropping the target variables Y = data["TARGET_CLOSE"] test_size = class="num">0.3 #class="num">70% of the data should be used for training purpose while the rest class="num">30% should be used for testing x_train, x_test, y_train, y_test = train_test_split(X, Y, shuffle=False, test_size = test_size) # this is timeseries data so we don&class="macro">#x27;t shuffle print(f"x_train {x_train.shape}\nx_test {x_test.shape}\ny_train{y_train.shape}\ny_test{y_test.shape}") def create_sequences(X, Y, time_step): if len(X) != len(Y): raise ValueError("X and y must have the same length") X = np.array(X) Y = np.array(Y) Xs, Ys = [], [] for i in range(X.shape[class="num">0] - time_step): Xs.append(X[i:(i + time_step), :]) # Include all features with slicing Ys.append(Y[i + time_step]) class="kw">return np.array(Xs), np.array(Ys) time_step = class="num">7 class="macro">#we consider the past class="num">7 days from sklearn.preprocessing class="kw">import StandardScaler scaler = StandardScaler() x_train = scaler.fit_transform(x_train) x_test = scaler.transform(x_test) x_train_seq, y_train_seq = create_sequences(x_train, y_train, time_step) x_test_seq, y_test_seq = create_sequences(x_test, y_test, time_step) print(f"Sequential data\n\nx_train {x_train_seq.shape}\nx_test {x_test_seq.shape}\ny_train{y_train_seq.shape}\ny_test{y_test_seq.shape}") Sequential data x_train(class="num">693, class="num">7, class="num">10) x_test(class="num">293, class="num">7, class="num">10) y_train(class="num">693,) y_test(class="num">293,)
RNN 回归跑分落后树模型
在测试集上评估完表现后，这套简单循环网络的 R² 准确率落在 78%。回看前一篇的同标的回归任务，LightGBM 这类非时序模型拿过 86.76% 的准确率，说明在纯数值回归上，未堆深度的 RNN 暂时跑不赢梯度提升树。 训练时用了提前停止：监控验证集损失，连续 5 轮不改善就回滚到最优权重。日志里第 95–100 轮 val_loss 从 4.4433e-05 缓降到 4.4329e-05，模型仍在缓慢收敛但边际极小。 别把 RNN 当万能时序锤。若你的 MT5 导出序列用 LightGBM 已到 86%+，先别急着换循环层，先查序列窗口和特征滞后阶数是否喂对了。外汇与贵金属波动受事件驱动，模型回测高准确率不代表实盘概率占优，杠杆品种高风险。
early_stopping = EarlyStopping(monitor=&class="macro">#x27;val_loss&class="macro">#x27;, patience=class="num">5, restore_best_weights=True) history = reg_model.fit(x_train_seq, y_train_seq, epochs=class="num">100, batch_size=class="num">64, verbose=class="num">1, validation_data=(x_test_seq, y_test_seq), callbacks=[early_stopping]) Epoch class="num">95/class="num">100 class="num">11/class="num">11 ━━━━━━━━━━━━━━━━━━━━ 0s 8ms/step - loss: class="num">6.4504e-05 - val_loss: class="num">4.4433e-05 Epoch class="num">96/class="num">100 class="num">11/class="num">11 ━━━━━━━━━━━━━━━━━━━━ 0s 8ms/step - loss: class="num">6.4380e-05 - val_loss: class="num">4.4408e-05 Epoch class="num">97/class="num">100 class="num">11/class="num">11 ━━━━━━━━━━━━━━━━━━━━ 0s 8ms/step - loss: class="num">6.4259e-05 - val_loss: class="num">4.4386e-05 Epoch class="num">98/class="num">100 class="num">11/class="num">11 ━━━━━━━━━━━━━━━━━━━━ 0s 8ms/step - loss: class="num">6.4140e-05 - val_loss: class="num">4.4365e-05 Epoch class="num">99/class="num">100 class="num">11/class="num">11 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - loss: class="num">6.4024e-05 - val_loss: class="num">4.4346e-05 Epoch class="num">100/class="num">100 class="num">11/class="num">11 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step - loss: class="num">6.3910e-05 - val_loss: class="num">4.4329e-05 from sklearn.metrics class="kw">import r2_score y_pred = reg_model.predict(x_test_seq) # Make predictions on the test set # Plot the actual vs predicted values plt.figure(figsize=(class="num">12, class="num">6)) plt.plot(y_test_seq, label=&class="macro">#x27;Actual Values&class="macro">#x27;) plt.plot(y_pred, label=&class="macro">#x27;Predicted Values&class="macro">#x27;) plt.xlabel(&class="macro">#x27;Samples&class="macro">#x27;) plt.ylabel(&class="macro">#x27;TARGET_CLOSE&class="macro">#x27;) plt.title(&class="macro">#x27;Actual vs Predicted Values&class="macro">#x27;) plt.legend() plt.show() print("RNN accuracy =",r2_score(y_test_seq, y_pred))
◍ SHAP 拆解 RNN 到底看了哪些行情片段
用 SHAP 跑了一遍 RNN 回归模型的决策归因，抽样 100 条训练序列、nsamples=100 做 KernelExplainer 近似计算。输出的 summary_plot 里，带最新时间步标记的特征权重明显压过最旧时间步——模型更像在盯「句子最后几个词」做判断，而不是均匀消化整段历史。 这种偏向近期信息的模式，对人类交易员未必符合直觉，但在机器学习里是常见现象：越靠近预测点的输入，梯度贡献往往越大。外汇与贵金属波动受突发流动性影响，近期价量信号权重高，也意味着模型对隔夜跳空类风险可能反应偏慢，实盘需警惕。 要注意的是，这里用的是 KernelExplainer 而非 DeepExplainer，后者在 RNN 上试过但报错频繁，所以归因结果只能作参考、不能当唯一依据。另外回归模型只预测下一根收盘/开盘价，不如分类器直接给方向信号实用，下一步该转向搭一个 RNN 分类器。
class="kw">import shap # Wrap the model prediction for KernelExplainer def rnn_predict(data): data = data.reshape((data.shape[class="num">0], time_step, x_train.shape[class="num">1])) class="kw">return reg_model.predict(data).flatten() # Use SHAP to explain the model sampled_idx = np.random.choice(len(x_train_seq), size=class="num">100, replace=False) explainer = shap.KernelExplainer(rnn_predict, x_train_seq[sampled_idx].reshape(class="num">100, -class="num">1)) shap_values = explainer.shap_values(x_test_seq[:class="num">100].reshape(class="num">100, -class="num">1), nsamples=class="num">100) # Update feature names for SHAP feature_names = [f&class="macro">#x27;{original_feat}_t{t}&class="macro">#x27; for t in range(time_step) for original_feat in X.columns] # Plot the SHAP values shap.summary_plot(shap_values, x_test_seq[:class="num">100].reshape(class="num">100, -class="num">1), feature_names=feature_names, max_display=len(feature_names), show=False) # Adjust layout and set figure size plt.subplots_adjust(left=class="num">0.12, bottom=class="num">0.1, right=class="num">0.9, top=class="num">0.9) plt.gcf().set_size_inches(class="num">7.5, class="num">14) plt.tight_layout() plt.savefig("regressor-rnn feature-importance.png") plt.show()
「用RNN给K线方向做二分类」
把回归思路搬来分类，第一件事是改目标变量：不再预测数值，而是标出每根K线是收高于开（偏多）还是否则偏空。下面这段就是按收盘价与开盘价比较打 0/1 标签，再转成独热编码喂给序列。 序列切好之后要立刻做 one-hot，否则后面建模会维度对不上。代码里 y_train_encoded 形状是 (693, 2)、y_test_encoded 是 (293, 2)，说明训练集 693 个样本、测试集 293 个，二分类两列。 网络尾巴和回归器不同：最后一层用 sigmoid，单元数必须等于类别数，这里写死 2 个。训练时设了 EarlyStopping(patience=5)，但实际跑下来 6 个 epoch 左右就收敛了，val_loss 在 0.687 附近横住。 在准确率上，这个 RNN 分类器比 LightGBM 分类器高出约 3 个百分点；不过外汇与贵金属波动随机性强，实盘信号失效概率不低，任何模型输出只作概率参考。
Y = [] target_open = data["TARGET_OPEN"] target_close = data["TARGET_CLOSE"] for i in range(len(target_open)): if target_close[i] > target_open[i]: # if the candle closed above where it opened thats a buy signal Y.append(class="num">1) else: class="macro">#otherwise it is a sell signal Y.append(class="num">0) Y = np.array(Y) class="macro">#converting this array to NumPy classes_in_y = np.unique(Y) # obtaining classes present in the target variable for the sake of setting the number of outputs in the RNN from tensorflow.keras.utils class="kw">import to_categorical y_train_encoded = to_categorical(y_train_seq) y_test_encoded = to_categorical(y_test_seq) print(f"One hot encoded\n\ny_train {y_train_encoded.shape}\ny_test {y_test_encoded.shape}") One hot encoded y_train(class="num">693, class="num">2) y_test(class="num">293, class="num">2) cls_model = Sequential() cls_model.add(Input(shape=(time_step, x_train.shape[class="num">1]))) # input layer cls_model.add(SimpleRNN(class="num">50, activation=&class="macro">#x27;relu&class="macro">#x27;)) cls_model.add(Dense(class="num">50, activation=&class="macro">#x27;relu&class="macro">#x27;)) cls_model.add(Dense(units=len(classes_in_y), activation=&class="macro">#x27;sigmoid&class="macro">#x27;, name=&class="macro">#x27;outputs&class="macro">#x27;)) adam_optimizer = Adam(learning_rate = class="num">0.001) cls_model.compile(optimizer=adam_optimizer, loss=&class="macro">#x27;binary_crossentropy&class="macro">#x27;) # Compile the model cls_model.summary() early_stopping = EarlyStopping(monitor=&class="macro">#x27;val_loss&class="macro">#x27;, patience=class="num">5, restore_best_weights=True) history = cls_model.fit(x_train_seq, y_train_encoded, epochs=class="num">100, batch_size=class="num">64, verbose=class="num">1, validation_data=(x_test_seq, y_test_encoded), callbacks=[early_stopping]) Model: "sequential_1" ┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓ ┃ Layer(type) ┃ Output Shape ┃ Param # ┃ ┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩ │ simple_rnn_1(SimpleRNN) │ (None, class="num">50) │ class="num">3,class="num">050 │ ├─────────────────────────────────┼────────────────────────┼───────────────┤ │ dense_2(Dense) │ (None, class="num">50) │ class="num">2,class="num">550 │ ├─────────────────────────────────┼────────────────────────┼───────────────┤ │ outputs(Dense) │ (None, class="num">2) │ class="num">102 │ └─────────────────────────────────┴────────────────────────┴───────────────┘ Epoch class="num">1/class="num">100 class="num">11/class="num">11 ━━━━━━━━━━━━━━━━━━━━ 2s 36ms/step - loss: class="num">0.7242 - val_loss: class="num">0.6872 Epoch class="num">2/class="num">100 class="num">11/class="num">11 ━━━━━━━━━━━━━━━━━━━━ 0s 9ms/step - loss: class="num">0.6883 - val_loss: class="num">0.6891 Epoch class="num">3/class="num">100 class="num">11/class="num">11 ━━━━━━━━━━━━━━━━━━━━ 0s 8ms/step - loss: class="num">0.6817 - val_loss: class="num">0.6909 Epoch class="num">4/class="num">100 class="num">11/class="num">11 ━━━━━━━━━━━━━━━━━━━━ 0s 8ms/step - loss: class="num">0.6780 - val_loss: class="num">0.6940 Epoch class="num">5/class="num">100 class="num">11/class="num">11 ━━━━━━━━━━━━━━━━━━━━ 0s 8ms/step - loss: class="num">0.6743 - val_loss: class="num">0.6974 Epoch class="num">6/class="num">100
分类报告里的真实命中率
上面这段输出是模型跑完 11 轮训练后的终端日志，最后落在一份分类报告上。训练损失停在 0.6707，验证损失 0.6998，两者差距不大，说明没明显过拟合，但绝对水平都不低。 报告按 0、1 两类给出精度：类别 0 的 precision 0.53、recall 0.27，f1 只有 0.36，样本 137 个；类别 1 的 precision 0.55、recall 0.79，f1 0.65，样本 156 个。整体 accuracy 0.55，macro avg f1 0.50，weighted avg f1 0.51。 类别 0 的召回率掉到 0.27，意味着真实为 0 的样本里七成以上被错判成 1。如果你拿这套标签做欧元兑美元或黄金的波段过滤，类别 0 对应的那侧信号大概率会漏掉大半，实战里可能倾向错过起涨段。外汇与贵金属波动受事件驱动强，这种概率级模型只能做辅助，高风险始终在。
class="num">11/class="num">11 ━━━━━━━━━━━━━━━━━━━━ 0s 8ms/step - loss: class="num">0.6707 - val_loss: class="num">0.6998 class="num">10/class="num">10 ━━━━━━━━━━━━━━━━━━━━ 0s 19ms/step Classification Report precision recall f1-score support class="num">0 class="num">0.53 class="num">0.27 class="num">0.36 class="num">137 class="num">1 class="num">0.55 class="num">0.79 class="num">0.65 class="num">156 accuracy class="num">0.55 class="num">293 macro avg class="num">0.54 class="num">0.53 class="num">0.50 class="num">293 weighted avg class="num">0.54 class="num">0.55 class="num">0.51 class="num">293
◍ 把RNN分类器落盘成MT5能跑的ONNX
Keras的循环神经网络不像Scikit-learn模型那样能直接走Pipeline导出，想进MetaTrader 5必须转成ONNX。缩放环节有两种路：要么在MQL5里收完数据现场缩放，要么在Python侧把标准化器的参数存好、用MT5预处理库加载，两条都能避开管道不兼容的坑。 标准化缩放器的核心就是均值和标准差，只要把这两组数组原样写出成二进制，缩放逻辑就完整保留了，后续MT5侧读入即可对齐训练时的分布。外汇与贵金属行情高波动，模型推理仅作概率参考，不代表方向必现。 下面这段把时序输入定死为time_step长度、特征数为x_train列数，用opset 13导出rnn.EURUSD.D1.onnx，顺手把mean_和scale_存成bin。开MT5把这三个文件丢进MQL5/Files，就能在EA里调ONNX推理了。
class="kw">import tf2onnx # Convert the Keras model to ONNX spec = (tf.TensorSpec((None, time_step, x_train.shape[class="num">1]), tf.float16, name="input"),) cls_model.output_names=[&class="macro">#x27;output&class="macro">#x27;] onnx_model, _ = tf2onnx.convert.from_keras(cls_model, input_signature=spec, opset=class="num">13) # Save the ONNX model to a file with open("rnn.EURUSD.D1.onnx", "wb") as f: f.write(onnx_model.SerializeToString()) # Save the mean and scale parameters to binary files scaler.mean_.tofile("standard_scaler_mean.bin") scaler.scale_.tofile("standard_scaler_scale.bin")