simonandClaude Opus 4.7
271a9343a5
7 Sprints 全部完成: Sprint 0: 基础设施 (DataManager + MariaDB) Sprint 1: 因子引擎 (34因子/12分类) Sprint 2: VectorBT 回测 (5策略+截面) Sprint 3: Optuna 优化 (+Walk-Forward) Sprint 4: ML 模型 (LightGBM+CatBoost) Sprint 5: Qwen 情绪因子 (三源新闻+日期对齐) Sprint 6: Agent 系统 (4Agent+日报.md/.html) 生产加固 (15项): Tushare双源fallback, SSH自动恢复, pool_pre_ping, save_daily先删后插, load_dotenv绝对路径, 日报5d/20d修复, RiskAgent改上证指数, 昨日对比+数据截止, mac_report utf8mb4, CLAUDE-*.md 9条已知Bug, demo全参数化, djapi数据源归一化, indexDatas API修正 Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
46 lines
1.7 KiB
Markdown
46 lines
1.7 KiB
Markdown
# CLAUDE-ml.md — ML 模型层
|
||
|
||
## FeatureEngine (`finance/models/features.py`)
|
||
|
||
因子 → 特征矩阵 + 标签。防前视偏差。
|
||
|
||
```python
|
||
from models.features import FeatureEngine
|
||
fe = FeatureEngine(lookahead=5, label_type="regression")
|
||
X, y = fe.build(factor_df, price_df, fit=True)
|
||
# fit=True: Winsorize(1%/99%) → ffill → median fill → RobustScaler.fit → 标签计算
|
||
# fit=False: 复用训练时的 scaler + 有效特征
|
||
```
|
||
|
||
## LightGBM (`finance/models/lightgbm/model.py`)
|
||
|
||
```python
|
||
from models.lightgbm.model import LightGBMModel
|
||
model = LightGBMModel(params={"n_estimators": 200, "learning_rate": 0.03}, eval_ratio=0.2)
|
||
model.fit(X_train, y_train) # val>=50 行才启用早停
|
||
pred = model.predict(X_test)
|
||
imp = model.get_feature_importance() # → DataFrame
|
||
cv = model.cv_evaluate(X, y, n_folds=5) # TimeSeriesSplit
|
||
```
|
||
|
||
## CatBoost (`finance/models/catboost/model.py`)
|
||
|
||
同接口。`get_feature_importance()` / `cv_evaluate()`。
|
||
|
||
## ML 策略 (`finance/models/backtest_integration.py`)
|
||
|
||
```python
|
||
from models.backtest_integration import MLStrategy, MLBenchmark
|
||
strategy = MLStrategy(model, fe, buy_quantile=0.7, sell_quantile=0.3, rebalance_freq=5)
|
||
# 预测值分位 → 动态阈值 → 交易信号
|
||
benchmark = MLBenchmark([lgb, cb], fe, price_df, factor_df)
|
||
result = benchmark.run() # → DataFrame: model × (IC, return, sharpe, win_rate, trades)
|
||
```
|
||
|
||
## 重要约束
|
||
|
||
- lookahead 固定，不输入模型（防目标泄露）
|
||
- 单股票 IC≈0 是正常现象（噪声主导），多股票截面才是 ML 发挥价值的地方
|
||
- 特征工程严禁使用未来数据（RobustScaler fit 在训练集，transform 在测试集）
|
||
- 交叉验证用 TimeSeriesSplit（不 shuffle）
|