📘

数据科学和机器学习（第 31 部分）：利用 CatBoost AI 模型进行交易·进阶篇

第 2/3 篇

本章目录

先看清数据骨架再谈策略
CatBoost 训练前的参数与拟合落地
CatBoost 早停后的模型裁剪点
CatBoost 分类报告的真实水位
训练尾段与特征权重怎么读

◍ 先看清数据骨架再谈策略

把 EURUSD 的日线导进 pandas，第一件事不是画信号，而是用 info() 看字段存活率。上面这段输出里，DayofWeek、DayofYear、Month 三列都是 6999 个 non-null，说明在 6999 根日 K 里没有缺日期，时间序列是连续的。 dtypes 显示整体是 4 个 float64 加 4 个 object，内存占用 492.1+ KB。这种体量在 MT5 里用 CSV 缓存或直接用 MQL5 的 File 函数读都不吃力，但在 Python 侧做滚动统计会更顺手。实操上，建议你开 MT5 导出最近一年日线，跑一遍同样的 info()，确认自己数据的 non-null 数和字段类型。若某列 non-null 明显少于总数，后续按星期几过滤信号时就会悄悄漏掉样本，回测结论可能偏乐观。

MQL5 / C++

class="num">5    DayofWeek  class="num">6999 non-null   object
class="num">6    DayofYear  class="num">6999 non-null   object
class="num">7    Month      class="num">6999 non-null   object
dtypes: float64(class="num">4), object(class="num">4)
memory usage: class="num">492.1+ KB

「CatBoost 训练前的参数与拟合落地」

在跑 fit 之前，先把几个直接影响过拟合与算力的参数盯牢。iterations 是树的数量，给到 100 虽不算大，但配合 learning_rate=0.01 这种小步长，模型倾向更稳；depth=10 能抓复杂形态，却也更容易在外汇小时线这类噪声里过拟合。l2_leaf_reg=5 用 L2 惩罚大叶权重，border_count=64 决定分类特征切分精度，数字越高越吃 CPU。 cat_features 最好显式传索引列表，别全靠模型自动识别——贵金属品种标签若被误判为数值，Human-readable 的类别关系就丢了。early_stopping_rounds 注释掉了，实盘前建议打开，验证集多少轮不提升就停，能省时间也压过拟合。下面这段是把参数塞进 sklearn 管道再 fit 的直跑写法。注意 catboost__eval_set 传了测试集，catboost__cat_features 传类别列名，管道前缀别拼错。从日志看，第 3 轮 test Logloss 摸到 0.6931239 的 best 后再没被刷新，到第 99 轮 test 已滑到 0.6936898，说明早停若设 10 轮会在第 13 轮附近断掉，白跑 86 轮。外汇与贵金属杠杆高、回测甜区不等于实盘，参数只代表概率倾向。

MQL5 / C++

params = dict(
    iterations=class="num">100,
    learning_rate=class="num">0.01,
    depth=class="num">10,
    l2_leaf_reg=class="num">5,  
    bagging_temperature=class="num">1,
    border_count=class="num">64,  # Number of splits for categorical features
    eval_metric=&class="macro">#x27;Logloss&class="macro">#x27;,
    random_seed=class="num">42,  # Seed for reproducibility
    verbose=class="num">1,  # Verbosity level
    # early_stopping_rounds=class="num">10  # Early stopping for validation
)
pipe = Pipeline([
    ("catboost", CatBoostClassifier(**params))
])
# Fit the pipeline to the training data
pipe.fit(X_train, y_train, catboost__eval_set=(X_test, y_test), catboost__cat_features=categorical_features)
class="num">90:     learn: class="num">0.6880592        test: class="num">0.6936112 best: class="num">0.6931239 (class="num">3)     total: 523ms    remaining: class="num">51.7ms
class="num">91:     learn: class="num">0.6880397        test: class="num">0.6936100 best: class="num">0.6931239 (class="num">3)     total: 529ms    remaining: 46ms
class="num">92:     learn: class="num">0.6880350        test: class="num">0.6936051 best: class="num">0.6931239 (class="num">3)     total: 532ms    remaining: 40ms
class="num">93:     learn: class="num">0.6880280        test: class="num">0.6936103 best: class="num">0.6931239 (class="num">3)     total: 535ms    remaining: class="num">34.1ms
class="num">94:     learn: class="num">0.6879448        test: class="num">0.6936110 best: class="num">0.6931239 (class="num">3)     total: 541ms    remaining: class="num">28.5ms
class="num">95:     learn: class="num">0.6878328        test: class="num">0.6936387 best: class="num">0.6931239 (class="num">3)     total: 547ms    remaining: class="num">22.8ms
class="num">96:     learn: class="num">0.6877888        test: class="num">0.6936473 best: class="num">0.6931239 (class="num">3)     total: 553ms    remaining: class="num">17.1ms
class="num">97:     learn: class="num">0.6877408        test: class="num">0.6936508 best: class="num">0.6931239 (class="num">3)     total: 559ms    remaining: class="num">11.4ms
class="num">98:     learn: class="num">0.6876611        test: class="num">0.6936708 best: class="num">0.6931239 (class="num">3)     total: 565ms    remaining: class="num">5.71ms
class="num">99:     learn: class="num">0.6876230        test: class="num">0.6936898 best: class="num">0.6931239 (class="num">3)     total: 571ms    remaining: 0us

CatBoost 早停后的模型裁剪点

CatBoost 在训练过程中若触发早停，会在日志里直接给出关键数值：某次回测中 bestTest 落在 0.6931239281，对应 bestIteration 为第 3 轮。系统随后提示 Shrink model to first 4 iterations，意味着最终保留的是前 4 轮迭代得到的模型，而非跑满全部树。对外汇与贵金属这类高波动品种做集成模型信号时，这种裁剪能压住过拟合，但样本外失效概率仍不低，实盘前务必在 MT5 策略测试器用分时段数据复跑确认。

MQL5 / C++

bestTest = class="num">0.6931239281
bestIteration = class="num">3
Shrink model to first class="num">4 iterations.

◍ CatBoost 分类报告的真实水位

用 Sklearn 的 classification_report 直接拉训练集和测试集的明细，比只看一个准确率数字实在得多。下面这段脚本在管道训练完后分别跑两套预测并打印报告，能立刻看到 precision、recall 和 f1 的分布。 [CODE] # Make predicitons on training and testing sets y_train_pred = pipe.predict(X_train) y_test_pred = pipe.predict(X_test) # Training set evaluation print("Training Set Classification Report:") print(classification_report(y_train, y_train_pred)) # Testing set evaluation print("\nTesting Set Classification Report:") print(classification_report(y_test, y_test_pred)) [/CODE] 训练集整体 accuracy 0.54，测试集掉到 0.51；类别 1 的 recall 在测试集为 0.61，但 precision 仅 0.49，说明模型倾向于把样本判为 1，却伴随不少误报。删掉分类特征列表后，训练集准确率能爬到 60%，测试集却纹丝不动——过拟合的信号很明显。再把 pipe 用 eval_set 挂上测试集跑 CatBoost，第 30 轮附近 best test 对数损失约 0.69305，之后 learn 继续下降到 0.684 左右而 test 徘徊在 0.6933，验证集收益已经吃尽。特征重要性图板显示，模型决策权重主要压在分类变量而非连续变量上，这点和删特征后训练集提升的现象对得上。外汇与贵金属行情受此类离散状态切换影响大，直接拿该口径建模实盘信号属高风险行为，仅适合作为特征筛选的参考。

MQL5 / C++

# Make predicitons on training and testing sets
y_train_pred = pipe.predict(X_train)
y_test_pred = pipe.predict(X_test)
# Training set evaluation
print("Training Set Classification Report:")
print(classification_report(y_train, y_train_pred))
# Testing set evaluation
print("\nTesting Set Classification Report:")
print(classification_report(y_test, y_test_pred))

「训练尾段与特征权重怎么读」

上面这段是 CatBoost 训练跑到第 95~99 轮时的日志切片。learn 损失从 0.6841427 缓降到 0.6838397，test 损失反而从 0.6933758 爬到 0.6934259，说明过拟合倾向在尾段已经显现；best test 停在 0.6930499562，对应 bestIteration = 30，模型随后被裁剪到前 31 轮。训练集分类报告里，类别 0 的 precision 0.61、recall 0.53，类别 1 的 precision 0.59、recall 0.67，总 accuracy 0.60（support 6999）。这个精度在外汇或贵金属行情分类上只能算弱信号，实盘直接跟单风险很高，仅适合做过滤层。日志之后是从 pipeline 里抠出 CatBoost 模型、拉特征重要度的代码。跑完能看到哪些字段对分类贡献大，进而决定下次训练砍掉哪些冗余特征、缩短迭代。

MQL5 / C++

# Extract the trained CatBoostClassifier from the pipeline
catboost_model = pipe.named_steps[&class="macro">#x27;catboost&class="macro">#x27;]
# Get feature importances
feature_importances = catboost_model.get_feature_importance()
feature_im_df = pd.DataFrame({
    "feature": X.columns,
    "importance": feature_importances
})
feature_im_df = feature_im_df.sort_values(by="importance", ascending=False)
plt.figure(figsize=(class="num">10, class="num">6))
sns.barplot(data = feature_im_df, x=&class="macro">#x27;importance&class="macro">#x27;, y=&class="macro">#x27;feature&class="macro">#x27;, palette="viridis")
plt.title("CatBoost feature importance")
plt.xlabel("Importance")
plt.ylabel("feature")
plt.show()