ProStock/src/experiment/regression.py

# %% md
# # LightGBM 回归训练流程（模块化版本）
#
# 使用新的模块化 Trainer 架构，代码更简洁、可维护性更高。
# %% md
# ## 1. 导入依赖
# %%
import os

from src.factors import FactorEngine
from src.training import (
    FactorManager,
    DataPipeline,
    RegressionTask,
    NullFiller,
    Winsorizer,
    StandardScaler,
    CrossSectionalStandardScaler,
)
from src.training.core.trainer_v2 import Trainer
from src.training.components.filters import STFilter
from src.experiment.common import (
    SELECTED_FACTORS,
    FACTOR_DEFINITIONS,
    LABEL_NAME,
    LABEL_FACTOR,
    TRAIN_START,
    TRAIN_END,
    VAL_START,
    VAL_END,
    TEST_START,
    TEST_END,
    stock_pool_filter,
    STOCK_FILTER_REQUIRED_COLUMNS,
    OUTPUT_DIR,
    SAVE_PREDICTIONS,
    SAVE_MODEL,
    get_model_save_path,
    save_model_with_factors,
    TOP_N,
    TRAIN_SKIP_DAYS,
)

# 训练类型标识
TRAINING_TYPE = "regression"

# %% md
# ## 2. 训练特定配置
# %%
# Label 配置（从 common.py 统一导入）
# LABEL_NAME 和 LABEL_FACTOR 已在 common.py 中绑定，只需从 common 导入

# 排除的因子列表
EXCLUDED_FACTORS = [
    # "amivest_liq_20",
    # "atr_price_impact",
    # "hui_heubel_ratio",
    # "corwin_schultz_spread_20",
    # "roll_spread_20",
    # "gibbs_effective_spread",
    # "overnight_illiq_20",
    # "illiq_volatility_20",
    # "amount_cv_20",
    # "amount_skewness_20",
    # "low_vol_days_20",
    # "liquidity_shock_momentum",
    # "downside_illiq_20",
    # "upside_illiq_20",
    # "illiq_asymmetry_20",
    # "pastor_stambaugh_proxy"
]

# 模型参数配置
MODEL_PARAMS = {
    # ==================== 基础设置 ====================
    "objective": "huber",  # 【修改】相比纯 L1(MAE)，huber 对异常值鲁棒且在极小误差处平滑，更适合收益率预测
    "metric": "mae",
    # ==================== 树结构约束 ====================
    "max_depth": 5,  # 【修改】适当加深，允许捕捉一定的高阶交叉
    "num_leaves": 31,  # 【修改】限制为 31（2的5次方-1），确保树是不对称生长的，防止过拟合
    "min_data_in_leaf": 512,  # 【大幅增加】从256加到1000。训练集有97万条，极大地限制叶子节点样本量能有效抵抗股市噪音
    # ==================== 学习参数 ====================
    "learning_rate": 0.01,  # 【修改】稍微调大一点，帮助模型跳出初始的局部最优（避免十几轮就早停）
    "n_estimators": 2000,
    # ==================== 随机采样与降维 ====================
    "subsample": 0.85,
    "subsample_freq": 1,
    "colsample_bytree": 0.4,  # 【大幅降低】从0.8降到0.4。强制打压 GTJA_alpha127 的霸权，逼迫模型去学习其他因子的信息
    "extra_trees": True,  # 【新增且极度推荐】极度随机树模式。在分裂点选择时增加随机性，是量化比赛中防过拟合的神器
    # ==================== 正则化 ====================
    "reg_alpha": 1.0,  # 【修改】L1正则增加，强行把一些无用特征的权重压到0
    "reg_lambda": 5.0,  # 【修改】L2正则大幅增加（从1到5），惩罚过大的叶子节点输出权重
    # "max_bin": 127,  # 【新增】默认255，降低到127相当于对连续特征做了一次粗颗粒度的分箱，也是极好的正则化手段
    # ==================== 杂项 ====================
    "verbose": -1,
    "random_state": 42,
    "n_jobs": -1,
}

# 日期范围配置
date_range = {
    "train": (TRAIN_START, TRAIN_END),
    "val": (VAL_START, VAL_END),
    "test": (TEST_START, TEST_END),
}

# 输出配置
output_config = {
    "output_dir": OUTPUT_DIR,
    "output_filename": "regression_output.csv",
    "save_predictions": SAVE_PREDICTIONS,
    "save_model": SAVE_MODEL,
    "model_save_path": get_model_save_path(TRAINING_TYPE),
    "top_n": TOP_N,
}


def main():
    """主函数"""
    print("\n" + "=" * 80)
    print("LightGBM 回归模型训练（模块化版本）")
    print("=" * 80)

    # 1. 创建 FactorEngine
    print("\n[1] 创建 FactorEngine")
    engine = FactorEngine()

    # 2. 创建 FactorManager
    print("\n[2] 创建 FactorManager")
    factor_manager = FactorManager(
        selected_factors=SELECTED_FACTORS,
        factor_definitions=FACTOR_DEFINITIONS,
        label_factor=LABEL_FACTOR,
        excluded_factors=EXCLUDED_FACTORS,
    )

    # 3. 创建 DataPipeline
    print("\n[3] 创建 DataPipeline")
    pipeline = DataPipeline(
        factor_manager=factor_manager,
        processor_configs=[
            (Winsorizer, {"lower": 0.01, "upper": 0.99}),
            (NullFiller, {"strategy": "mean"}),
            (StandardScaler, {}),
            # (CrossSectionalStandardScaler, {}),
        ],
        label_processor_configs=[
            # 对 label 进行缩尾处理（去除极端收益率）
            (Winsorizer, {"lower": 0.05, "upper": 0.95}),
            (StandardScaler, {}),
        ],
        filters=[STFilter(data_router=engine.router)],
        stock_pool_filter_func=stock_pool_filter,
        stock_pool_required_columns=STOCK_FILTER_REQUIRED_COLUMNS,
        train_skip_days=TRAIN_SKIP_DAYS,
    )

    # 4. 创建 RegressionTask
    print("\n[4] 创建 RegressionTask")
    task = RegressionTask(
        model_params=MODEL_PARAMS,
        label_name=LABEL_NAME,
    )

    # 5. 创建 Trainer
    print("\n[5] 创建 Trainer")
    trainer = Trainer(
        data_pipeline=pipeline,
        task=task,
        output_config=output_config,
        verbose=True,
    )

    # 6. 执行训练
    print("\n[6] 执行训练")
    results = trainer.run(engine=engine, date_range=date_range)

    # 7. 保存模型和因子信息（如果启用）
    if SAVE_MODEL:
        print("\n[7] 保存模型和因子信息")
        save_model_with_factors(
            model=task.get_model(),
            model_path=output_config["model_save_path"],
            selected_factors=SELECTED_FACTORS,
            factor_definitions=FACTOR_DEFINITIONS,
            fitted_processors=pipeline.get_fitted_processors(),
        )

    print("\n" + "=" * 80)
    print("训练流程完成！")
    print(f"结果保存路径: {os.path.join(OUTPUT_DIR, 'regression_output.csv')}")
    print("=" * 80)

    return results


if __name__ == "__main__":
    main()