数据科学和机器学习（第 29 部分）：为 AI 训练目的而选择最佳外汇数据的基本技巧·综合运用
用 L1 惩罚把无关变量直接归零
线性回归靠最小化预测值和真实值的平方差来定系数，但特征一多就容易过拟合、方差炸开。Lasso 引入 L1 正则化，把系数按一个常量做绝对值惩罚，逼着不重要的系数精确落到 0——这不是简单缩小，而是顺手完成了特征选择，能直接把对应列从数据里删掉。 Ridge 用的是 L2（罚平方），只会把系数压小但不会归零；Elastic Net 则是两者混搭。Lasso 的惩罚强度由 alpha 控制，得靠超参搜索找合适值，不是拍脑袋定的。 跑通下面这段 Python 就能看见实效：在测试集上只保留 4 个非零系数特征（HIGH、LOW、CLOSE、30DAY_SMA），R² 达到 0.989，说明砍掉冗余列后模型依旧贴合。外汇与贵金属行情高度随机，此类特征筛选仅降低过拟合概率，不预示任何方向。 别把正态当圣经：当特征间高度相关时，Lasso 给哪列留系数倾向随机，估算可能不稳；若特征数多于样本数，它也可能跑歪。 代码逐行拆解： from sklearn.model_selection import train_test_split —— 导入切分训练/测试集工具 from sklearn.linear_model import Lasso —— 导入 Lasso 回归类 from sklearn.metrics import r2_score —— 导入 R² 评分函数 from sklearn.preprocessing import MinMaxScaler —— 导入归一化缩放器 y = df["TARGET_CLOSE"] —— 取目标收盘价列作标签 X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42) —— 按 8:2 切数据，固定随机种子保证可复现 scaler = MinMaxScaler() —— 建归一化器，把特征压到 0~1 区间 lasso = Lasso(alpha=0.001) # You need tune for the best penalty value —— 初始化 Lasso，alpha=0.001 是起点，实际要调 X_train = scaler.fit_transform(X_train) —— 用训练集拟合并转换 lasso.fit(X_train, y_train) —— 训练 Lasso print(f'Coefficients: {lasso.coef_}') —— 打印各特征系数 X_test = scaler.transform(X_test) —— 测试集用同一 scaler 转换 y_pred = lasso.predict(X_test) —— 预测测试集 mse = r2_score(y_test, y_pred) —— 算 R² print(f'Lasso regression test accuracy = {mse}') —— 输出准确率 selected_features = X.columns[lasso.coef_ != 0] —— 挑出系数非 0 的特征名 print(f'Selected Features: {selected_features}') —— 打印入选特征
from sklearn.model_selection class="kw">import train_test_split from sklearn.linear_model class="kw">import Lasso from sklearn.metrics class="kw">import r2_score from sklearn.preprocessing class="kw">import MinMaxScaler y = df["TARGET_CLOSE"] # Split the data into training and testing sets X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=class="num">0.2, random_state=class="num">42) # A scaling technique scaler = MinMaxScaler() # Initialize and fit the lasso model lasso = Lasso(alpha=class="num">0.001) # You need tune for the best penalty value # Train the scaler and transfrom data X_train = scaler.fit_transform(X_train) lasso.fit(X_train, y_train) print(f&class="macro">#x27;Coefficients: {lasso.coef_}&class="macro">#x27;) class="macro">#print coefficients # Predict on the test set X_test = scaler.transform(X_test) y_pred = lasso.predict(X_test) # Calculate mean squared error mse = r2_score(y_test, y_pred) print(f&class="macro">#x27;Lasso regression test accuracy = {mse}&class="macro">#x27;) # select all features with coefficents not equal to zero selected_features = X.columns[lasso.coef_ != class="num">0] print(f&class="macro">#x27;Selected Features: {selected_features}&class="macro">#x27;)
「用随机森林筛出真正驱动价格的字段」
决策树靠递归切分数据做预测，每个节点挑一个特征和一个阈值来拆，目标是把混杂度压到最低。某个特征的重要性，看它在整棵树里累计减少了多少杂质——在多个节点都拿来切，就把各处的减少量加总。随机森林是并行长出几十上百棵树，最终预测取均值或投票，特征重要性则是所有树的平均。 梯度提升机（如 XGBoost）按顺序建树，后一棵树专门补前一棵的错，重要性是所有树之和。不管哪种，跑完都能拿到一份特征重要性数值，用来挑出模型最倚重的输入。 别急着把选出的特征直接上实盘。先用测试集量一下随机森林分类器的准确率，确认这些字段在没见过的数据上也站得住脚。下面这段 Python 可直接在 MT5 导出的 CSV 上跑，阈值设 0.04 时筛出了 HIGH、CLOSE、OPEN_LAG3、30DAY_SMA 等 10 个字段，测试集准确率需你本地 print 出来看。 随机森林做特征选择，优势是树群融汇降低了单棵树的过拟合，重要性分数更稳；能扛高维数据不怎么掉性能，还能抓特征间的非线性互动。缺点是大数据集训练贵，相关特征会分到相似分数难区分，且可能偏爱连续或多级别特征，给类别少的字段打分偏低。外汇与贵金属波动剧烈、杠杆高风险极大，任何筛选结果都只是概率倾向，须经历史回测验证。
from sklearn.ensemble class="kw">import RandomForestClassifier y = [] # Loop through each row in the DataFrame to create the target variable for i in range(len(df)): if df.loc[i, &class="macro">#x27;TARGET_CLOSE&class="macro">#x27;] > df.loc[i, &class="macro">#x27;TARGET_OPEN&class="macro">#x27;]: y.append(class="num">1) else: y.append(class="num">0) # Split the data into training and testing sets X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=class="num">0.2, random_state=class="num">42) model = RandomForestClassifier(n_estimators=class="num">50, min_samples_split=class="num">10, max_depth=class="num">5, min_samples_leaf=class="num">5) model.fit(X_train, y_train) importances = model.feature_importances_ print(importances) selected_features = importances > class="num">0.04 selected_feature_names = X.columns[selected_features] print("selected features\n",selected_feature_names) [class="num">0.02691807 class="num">0.05334113 class="num">0.03780997 class="num">0.0563491 class="num">0.03162462 class="num">0.03486413 class="num">0.02652285 class="num">0.0237652 class="num">0.03398946 class="num">0.02822157 class="num">0.01794172 class="num">0.02818283 class="num">0.04052433 class="num">0.02821834 class="num">0.0386661 class="num">0.03921218 class="num">0.04406372 class="num">0.06162133 class="num">0.03103843 class="num">0.02206782 class="num">0.05104613 class="num">0.01700301 class="num">0.05191551 class="num">0.07251801 选择的特征 Index([&class="macro">#x27;HIGH&class="macro">#x27;, &class="macro">#x27;CLOSE&class="macro">#x27;, &class="macro">#x27;OPEN_LAG3&class="macro">#x27;, &class="macro">#x27;30DAY_SMA&class="macro">#x27;, &class="macro">#x27;7DAY_STDDEV&class="macro">#x27;, &class="macro">#x27;DAYOFYEAR&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_CLOSE&class="macro">#x27;], dtype=&class="macro">#x27;object&class="macro">#x27;) from sklearn.metrics class="kw">import accuracy_score test_pred = model.predict(X_test) print(f"Random forest test accuracy = ",accuracy_score(y_test, test_pred))
◍ 把 OHLC 四个价压成一个主成分
开盘、最高、最低、收盘这四列在相关矩阵里高度共线，硬塞进模型只是堆噪声。用 PCA 压到 1 维，新变量叫 OHLC，信息损耗比直觉里小得多。 拿线性回归跑对照：原始 4 特征数据的 R² 是 0.9938，降维后单特征 R² 是 0.9921，差 0.0017。对外汇或贵金属这类高噪品种，这种精度损失通常可忽略，但模型维度从 4 降到 1。 降维的好处是直接砍掉冗余、缓解过拟合，特征空间小了模型更好解释。代价也实在：PCA 要自己定保留几个分量，选错就丢信号；压出来的分量物理含义模糊，回测里容易错失价间微妙关系。 下面这段 Python 可直接复现上面的对照，把 df 换成你的 MT5 导出的 OHLCV 数据就能跑。
from sklearn.decomposition class="kw">import PCA from sklearn.linear_model class="kw">import LinearRegression pca = PCA(n_components=class="num">1) ohlc = pd.DataFrame({ "OPEN": df["OPEN"], "HIGH": df["HIGH"], "LOW": df["LOW"], "CLOSE": df["CLOSE"] }) y = df["TARGET_CLOSE"] # let us use the linear regression model model = LinearRegression() # for OHLC original data model.fit(ohlc, y) preds = model.predict(ohlc) print("ohlc_original data LR accuracy = ",r2_score(y, preds)) # For data reduced in dimension ohlc_reduced = pca.fit_transform(ohlc) print(ohlc_reduced[:class="num">10]) # print class="num">10 rows of the reduced data model.fit(ohlc_reduced, y) preds = model.predict(ohlc_reduced) print("ohlc_reduced data LR accuracy = ",r2_score(y, preds)) ohlc_original data LR accuracy = class="num">0.9937597843724363 [[-class="num">0.14447016] [-class="num">0.14997874] [-class="num">0.14129409] [-class="num">0.1293209 ] [-class="num">0.12659902] [-class="num">0.12895961] [-class="num">0.13831287] [-class="num">0.14061213] [-class="num">0.14719862] [-class="num">0.15752861]] ohlc_reduced data LR accuracy = class="num">0.9921387699876517 new_df = df.drop(columns=["OPEN", "HIGH", "LOW", "CLOSE"]) # new_df["OHLC"] = ohlc_reduced # Reorder the columns to make "ohlc" the first column cols = ["OHLC"] + [col for col in new_df.columns if col != "OHLC"] new_df = new_df[cols] new_df.head(class="num">10)
收束
把特征选择真正跑通，关键不在算法本身，而在你敢不敢动那几个 99% 相关的 OHLC 字段。前文给出的 feature_selection.ipynb 里，Open、High、Low 在样本集上相关系数超过 0.99，直接删变量能省训练时间，但也会把影线和突破强度这类非线性信息一起丢掉。 更稳的做法是把高/低转成相对价差再喂给模型，维度没少多少，几何结构却保留完整。MT5 接 Python 做策略测试时，这种预处理能让回测更贴近真实烛形逻辑。 外汇与贵金属杠杆高、价格跳跃频繁，任何特征工程都只是降低过拟合概率，不保证实盘优势。开 MT5 把 Timeseries_OHLC.csv 拖进笔记本跑一遍，比看结论更有用。