liaozhaorun/ProStock
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

refactor(experiment): 提取共用配置到 common 模块

Browse Source 
- 将因子定义、日期配置、股票池筛选等提取到 common.py
- 重构 learn_to_rank 和 regression 脚本，统一使用公共配置
- 简化代码结构，消除重复定义
This commit is contained in:
liaozhaorun
2026-03-15 05:46:19 +08:00
parent 6927d20de1
commit 0e9ea5d533
5 changed files with 1127 additions and 962 deletions
Download Patch File Download Diff File Expand all files Collapse all files

278
src/experiment/common.py Normal file
View File

@@ -0,0 +1,278 @@
"""实验脚本的共用配置和辅助函数。
此模块包含 regression.py 和 learn_to_rank.py 共用的代码，
避免重复维护两份相同的配置和函数。
"""
from datetime import datetime
from typing import List
import polars as pl
from src.factors import FactorEngine
# =============================================================================
# 日期范围配置（正确的 train/val/test 三分法）
# =============================================================================
TRAIN_START = "20200101"
TRAIN_END = "20231231"
VAL_START = "20240101"
VAL_END = "20241231"
TEST_START = "20250101"
TEST_END = "20261231"
# =============================================================================
# 因子配置
# =============================================================================
# 当前选择的因子列表（从 FACTOR_DEFINITIONS 中选择要使用的因子）
SELECTED_FACTORS = [
# ================= 1. 价格、趋势与路径依赖 =================
"ma_5",
"ma_20",
"ma_ratio_5_20",
"bias_10",
"high_low_ratio",
"bbi_ratio",
"return_5",
"return_20",
"kaufman_ER_20",
"mom_acceleration_10_20",
"drawdown_from_high_60",
"up_days_ratio_20",
# ================= 2. 波动率、风险调整与高阶矩 =================
"volatility_5",
"volatility_20",
"volatility_ratio",
"std_return_20",
"sharpe_ratio_20",
"min_ret_20",
"volatility_squeeze_5_60",
# ================= 3. 日内微观结构与异象 =================
"overnight_intraday_diff",
"upper_shadow_ratio",
"capital_retention_20",
"max_ret_20",
# ================= 4. 量能、流动性与量价背离 =================
"volume_ratio_5_20",
"turnover_rate_mean_5",
"turnover_deviation",
"amihud_illiq_20",
"turnover_cv_20",
"pv_corr_20",
"close_vwap_deviation",
# ================= 5. 基本面财务特征 =================
"roe",
"roa",
"profit_margin",
"debt_to_equity",
"current_ratio",
"net_profit_yoy",
"revenue_yoy",
"healthy_expansion_velocity",
# ================= 6. 基本面估值与截面动量共振 =================
"EP",
"BP",
"CP",
"market_cap_rank",
"turnover_rank",
"return_5_rank",
"EP_rank",
"pe_expansion_trend",
"value_price_divergence",
"active_market_cap",
"ebit_rank",
]
# 因子定义字典（完整因子库，用于存放尚未注册到metadata的因子）
FACTOR_DEFINITIONS = {}
def get_label_factor(label_name: str) -> dict:
"""获取Label因子定义字典。
Args:
label_name: label因子名称
Returns:
Label因子定义字典
"""
return {
label_name: "(ts_delay(close, -5) / ts_delay(open, -1)) - 1",
}
# =============================================================================
# 辅助函数
# =============================================================================
def register_factors(
engine: FactorEngine,
selected_factors: List[str],
factor_definitions: dict,
label_factor: dict,
) -> List[str]:
"""注册因子。
selected_factors 从 metadata 查询，factor_definitions 用 DSL 表达式注册。
Args:
engine: FactorEngine实例
selected_factors: 从metadata中选择的因子名称列表
factor_definitions: 通过表达式定义的因子字典
label_factor: label因子定义字典
Returns:
特征列名称列表
"""
print("=" * 80)
print("注册因子")
print("=" * 80)
# 注册 SELECTED_FACTORS 中的因子（已在 metadata 中）
print("\n注册特征因子（从 metadata）:")
for name in selected_factors:
engine.add_factor(name)
print(f" - {name}")
# 注册 FACTOR_DEFINITIONS 中的因子（通过表达式，尚未在 metadata 中）
print("\n注册特征因子（表达式）:")
for name, expr in factor_definitions.items():
engine.add_factor(name, expr)
print(f" - {name}: {expr}")
# 注册 label 因子（通过表达式）
print("\n注册 Label 因子（表达式）:")
for name, expr in label_factor.items():
engine.add_factor(name, expr)
print(f" - {name}: {expr}")
# 特征列 = SELECTED_FACTORS + FACTOR_DEFINITIONS 的 keys
feature_cols = selected_factors + list(factor_definitions.keys())
print(f"\n特征因子数: {len(feature_cols)}")
print(f" - 来自 metadata: {len(selected_factors)}")
print(f" - 来自表达式: {len(factor_definitions)}")
print(f"Label: {list(label_factor.keys())[0]}")
print(f"已注册因子总数: {len(engine.list_registered())}")
return feature_cols
def prepare_data(
engine: FactorEngine,
feature_cols: List[str],
start_date: str,
end_date: str,
label_name: str,
) -> pl.DataFrame:
"""准备数据。
计算因子并返回包含特征和label的数据框。
Args:
engine: FactorEngine实例
feature_cols: 特征列名称列表
start_date: 开始日期 (YYYYMMDD)
end_date: 结束日期 (YYYYMMDD)
label_name: label列名称
Returns:
包含因子计算结果的数据框
"""
print("\n" + "=" * 80)
print("准备数据")
print("=" * 80)
# 计算因子（全市场数据）
print(f"\n计算因子: {start_date} - {end_date}")
factor_names = feature_cols + [label_name] # 包含 label
data = engine.compute(
factor_names=factor_names,
start_date=start_date,
end_date=end_date,
)
print(f"数据形状: {data.shape}")
print(f"数据列: {data.columns}")
print(f"\n前5行预览:")
print(data.head())
return data
# =============================================================================
# 股票池筛选配置
# =============================================================================
def stock_pool_filter(df: pl.DataFrame) -> pl.Series:
"""股票池筛选函数（单日数据）。
筛选条件：
1. 排除创业板（代码以 300 开头）
2. 排除科创板（代码以 688 开头）
3. 排除北交所（代码以 8、9 或 4 开头）
4. 选取当日市值最小的500只股票
Args:
df: 单日数据框
Returns:
布尔Series，表示哪些股票被选中
"""
# 代码筛选（排除创业板、科创板、北交所）
code_filter = (
~df["ts_code"].str.starts_with("30") # 排除创业板
& ~df["ts_code"].str.starts_with("68") # 排除科创板
& ~df["ts_code"].str.starts_with("8") # 排除北交所
& ~df["ts_code"].str.starts_with("9") # 排除北交所
& ~df["ts_code"].str.starts_with("4") # 排除北交所
)
# 在已筛选的股票中，选取市值最小的500只
valid_df = df.filter(code_filter)
n = min(500, len(valid_df))
small_cap_codes = valid_df.sort("total_mv").head(n)["ts_code"]
# 返回布尔 Series：是否在被选中的股票中
return df["ts_code"].is_in(small_cap_codes)
# 定义筛选所需的基础列
STOCK_FILTER_REQUIRED_COLUMNS = ["total_mv"]
# =============================================================================
# 输出配置
# =============================================================================
OUTPUT_DIR = "output"
SAVE_PREDICTIONS = True
PERSIST_MODEL = False
# Top N 配置：每日推荐股票数量
TOP_N = 5 # 可调整为 10, 20 等
def get_output_path(model_type: str, test_start: str, test_end: str) -> str:
"""生成输出文件路径。
Args:
model_type: 模型类型（"regression" 或 "rank"）
test_start: 测试开始日期
test_end: 测试结束日期
Returns:
输出文件路径
"""
import os
# 确保输出目录存在
os.makedirs(OUTPUT_DIR, exist_ok=True)
# 生成文件名
start_dt = datetime.strptime(test_start, "%Y%m%d")
end_dt = datetime.strptime(test_end, "%Y%m%d")
date_str = f"{start_dt.strftime('%Y%m%d')}_{end_dt.strftime('%Y%m%d')}"
filename = f"{model_type}_output.csv"
return os.path.join(OUTPUT_DIR, filename)

901
src/experiment/learn_to_rank.ipynb
View File

File diff suppressed because one or more lines are too long

268
src/experiment/learn_to_rank.py
View File

@@ -1,4 +1,4 @@
#%% md
# %% md
# # Learn-to-Rank 排序学习训练流程
# #
# 本 Notebook 实现基于 LightGBM LambdaRank 的排序学习训练，用于股票排序任务。
@@ -9,9 +9,9 @@
# 2. **排序学习**: 使用 LambdaRank 目标函数，学习每日股票排序
# 3. **NDCG 评估**: 使用 NDCG@1/5/10/20 评估排序质量
# 4. **策略回测**: 基于排序分数构建 Top-k 选股策略
#%% md
# %% md
# ## 1. 导入依赖
#%%
# %%
import os
from datetime import datetime
from typing import List, Tuple, Optional
@@ -36,78 +36,32 @@ from src.training import (
from src.training.components.models import LightGBMLambdaRankModel
from src.training.config import TrainingConfig
#%% md
# ## 2. 辅助函数
#%%
def register_factors(
engine: FactorEngine,
selected_factors: List[str],
factor_definitions: dict,
label_factor: dict,
) -> List[str]:
"""注册因子（selected_factors 从 metadata 查询，factor_definitions 用 DSL 表达式注册）"""
print("=" * 80)
print("注册因子")
print("=" * 80)
# 注册 SELECTED_FACTORS 中的因子（已在 metadata 中）
print("\n注册特征因子（从 metadata）:")
for name in selected_factors:
engine.add_factor(name)
print(f" - {name}")
# 注册 FACTOR_DEFINITIONS 中的因子（通过表达式，尚未在 metadata 中）
print("\n注册特征因子（表达式）:")
for name, expr in factor_definitions.items():
engine.add_factor(name, expr)
print(f" - {name}: {expr}")
# 注册 label 因子（通过表达式）
print("\n注册 Label 因子（表达式）:")
for name, expr in label_factor.items():
engine.add_factor(name, expr)
print(f" - {name}: {expr}")
# 特征列 = SELECTED_FACTORS + FACTOR_DEFINITIONS 的 keys
feature_cols = selected_factors + list(factor_definitions.keys())
print(f"\n特征因子数: {len(feature_cols)}")
print(f" - 来自 metadata: {len(selected_factors)}")
print(f" - 来自表达式: {len(factor_definitions)}")
print(f"Label: {list(label_factor.keys())[0]}")
print(f"已注册因子总数: {len(engine.list_registered())}")
return feature_cols
# 从 common 模块导入共用配置和函数
from src.experiment.common import (
SELECTED_FACTORS,
FACTOR_DEFINITIONS,
get_label_factor,
register_factors,
prepare_data,
TRAIN_START,
TRAIN_END,
VAL_START,
VAL_END,
TEST_START,
TEST_END,
stock_pool_filter,
STOCK_FILTER_REQUIRED_COLUMNS,
OUTPUT_DIR,
SAVE_PREDICTIONS,
PERSIST_MODEL,
TOP_N,
)
def prepare_data(
engine: FactorEngine,
feature_cols: List[str],
start_date: str,
end_date: str,
) -> pl.DataFrame:
"""准备数据"""
print("\n" + "=" * 80)
print("准备数据")
print("=" * 80)
# 计算因子（全市场数据）
print(f"\n计算因子: {start_date} - {end_date}")
factor_names = feature_cols + [LABEL_NAME] # 包含 label
data = engine.compute(
factor_names=factor_names,
start_date=start_date,
end_date=end_date,
)
print(f"数据形状: {data.shape}")
print(f"数据列: {data.columns}")
print(f"\n前5行预览:")
print(data.head())
return data
# %% md
# ## 2. 本地辅助函数
# %%
# 注意：register_factors 和 prepare_data 已从 common 模块导入
def prepare_ranking_data(
@@ -240,92 +194,22 @@ def evaluate_ndcg_at_k(
return results
#%% md
# %% md
# ## 3. 配置参数
# #
# ### 3.1 因子定义
#%%
# 特征因子定义字典（复用 regression.ipynb 的因子定义）
LABEL_NAME = "future_return_5_rank"
# ### 3.1 因子与日期配置
# %%
# 注意：SELECTED_FACTORS, FACTOR_DEFINITIONS, 日期配置等已从 common 模块导入
# 本脚本特有的配置：
# 当前选择的因子列表（从 FACTOR_DEFINITIONS 中选择要使用的因子）
SELECTED_FACTORS = [
# ================= 1. 价格、趋势与路径依赖 =================
"ma_5",
"ma_20",
"ma_ratio_5_20",
"bias_10",
"high_low_ratio",
"bbi_ratio",
"return_5",
"return_20",
"kaufman_ER_20",
"mom_acceleration_10_20",
"drawdown_from_high_60",
"up_days_ratio_20",
# ================= 2. 波动率、风险调整与高阶矩 =================
"volatility_5",
"volatility_20",
"volatility_ratio",
"std_return_20",
"sharpe_ratio_20",
"min_ret_20",
"volatility_squeeze_5_60",
# ================= 3. 日内微观结构与异象 =================
"overnight_intraday_diff",
"upper_shadow_ratio",
"capital_retention_20",
"max_ret_20",
# ================= 4. 量能、流动性与量价背离 =================
"volume_ratio_5_20",
"turnover_rate_mean_5",
"turnover_deviation",
"amihud_illiq_20",
"turnover_cv_20",
"pv_corr_20",
"close_vwap_deviation",
# ================= 5. 基本面财务特征 =================
"roe",
"roa",
"profit_margin",
"debt_to_equity",
"current_ratio",
"net_profit_yoy",
"revenue_yoy",
"healthy_expansion_velocity",
"ebit_rank",
# ================= 6. 基本面估值与截面动量共振 =================
"EP",
"BP",
"CP",
"market_cap_rank",
"turnover_rank",
"return_5_rank",
"EP_rank",
"pe_expansion_trend",
"value_price_divergence",
"active_market_cap",
]
# Label 名称（排序学习使用原始收益率，会后续转换为分位数标签）
LABEL_NAME = "future_return_5"
# 因子定义字典（完整因子库）
FACTOR_DEFINITIONS = {
# "turnover_rate_volatility": "ts_std(log(turnover_rate), 20)"
}
# 获取 Label 因子定义
LABEL_FACTOR = get_label_factor(LABEL_NAME)
# Label 因子定义（不参与训练，用于计算目标）
LABEL_FACTOR = {
LABEL_NAME: "(ts_delay(close, -5) / ts_delay(open, -1)) - 1",
}
#%% md
# ### 3.2 训练参数配置
#%%
# 日期范围配置（正确的 train/val/test 三分法）
TRAIN_START = "20200101"
TRAIN_END = "20231231"
VAL_START = "20240101"
VAL_END = "20241231"
TEST_START = "20250101"
TEST_END = "20251231"
# 分位数配置
N_QUANTILES = 20 # 将 label 分为 20 组
# 分位数配置
@@ -352,44 +236,11 @@ MODEL_PARAMS = {
"label_gain": [i for i in range(1, N_QUANTILES + 1)],
}
# 股票池筛选函数
def stock_pool_filter(df: pl.DataFrame) -> pl.Series:
"""股票池筛选函数（单日数据）
筛选条件：
1. 排除创业板（代码以 300 开头）
2. 排除科创板（代码以 688 开头）
3. 排除北交所（代码以 8、9 或 4 开头）
4. 选取当日市值最小的500只股票
"""
code_filter = (
~df["ts_code"].str.starts_with("30")
& ~df["ts_code"].str.starts_with("68")
& ~df["ts_code"].str.starts_with("8")
& ~df["ts_code"].str.starts_with("9")
& ~df["ts_code"].str.starts_with("4")
)
valid_df = df.filter(code_filter)
n = min(500, len(valid_df))
small_cap_codes = valid_df.sort("total_mv").head(n)["ts_code"]
return df["ts_code"].is_in(small_cap_codes)
STOCK_FILTER_REQUIRED_COLUMNS = ["total_mv"]
# 输出配置
OUTPUT_DIR = "output"
SAVE_PREDICTIONS = True
PERSIST_MODEL = False
# Top N 配置：每日推荐股票数量
TOP_N = 5 # 可调整为 10, 20 等
#%% md
# 注意：stock_pool_filter, STOCK_FILTER_REQUIRED_COLUMNS, OUTPUT_DIR 等配置
# 已从 common 模块导入
# %% md
# ## 4. 训练流程
#%%
# %%
print("\n" + "=" * 80)
print("LightGBM LambdaRank 排序学习训练")
print("=" * 80)
@@ -411,6 +262,7 @@ data = prepare_data(
feature_cols=feature_cols,
start_date=TRAIN_START,
end_date=TEST_END,
label_name=LABEL_NAME,
)
# 4. 转换为排序学习格式（分位数标签）
@@ -469,9 +321,9 @@ trainer = Trainer(
feature_cols=feature_cols,
persist_model=PERSIST_MODEL,
)
#%% md
# %% md
# ### 4.1 股票池筛选
#%%
# %%
print("\n" + "=" * 80)
print("股票池筛选")
print("=" * 80)
@@ -493,9 +345,9 @@ if pool_manager:
else:
filtered_data = data
print(" 未配置股票池管理器，跳过筛选")
#%% md
# %% md
# ### 4.2 数据划分
#%%
# %%
print("\n" + "=" * 80)
print("数据划分")
print("=" * 80)
@@ -519,9 +371,9 @@ if splitter:
print(f"测试集日均样本数: {np.mean(test_group):.1f}")
else:
raise ValueError("必须配置数据划分器")
#%% md
# %% md
# ### 4.3 数据质量检查
#%%
# %%
print("\n" + "=" * 80)
print("数据质量检查（必须在预处理之前）")
print("=" * 80)
@@ -537,9 +389,9 @@ check_data_quality(test_data, feature_cols, raise_on_error=True)
print("[成功] 数据质量检查通过，未发现异常")
#%% md
# %% md
# ### 4.4 数据预处理
#%%
# %%
print("\n" + "=" * 80)
print("数据预处理")
print("=" * 80)
@@ -563,9 +415,9 @@ if processors:
print(f"\n处理后训练集形状: {train_data.shape}")
print(f"处理后验证集形状: {val_data.shape}")
print(f"处理后测试集形状: {test_data.shape}")
#%% md
# %% md
# ### 4.4 训练 LambdaRank 模型
#%%
# %%
print("\n" + "=" * 80)
print("训练 LambdaRank 模型")
print("=" * 80)
@@ -593,9 +445,9 @@ model.fit(
eval_set=(X_val, y_val, val_group),
)
print("训练完成！")
#%% md
# %% md
# ### 4.5 训练指标曲线
#%%
# %%
print("\n" + "=" * 80)
print("训练指标曲线")
print("=" * 80)
@@ -645,9 +497,9 @@ else:
best_val = max(val_metric_list)
print(f" {metric}: {best_val:.4f} (迭代 {best_iter_metric + 1})")
print(f"\n[重要提醒] 验证集仅用于早停/调参，测试集完全独立于训练过程！")
#%% md
# %% md
# ### 4.6 模型评估
#%%
# %%
print("\n" + "=" * 80)
print("模型评估")
print("=" * 80)
@@ -685,7 +537,7 @@ if importance is not None:
top_features = importance.sort_values(ascending=False).head(20)
for i, (feature, score) in enumerate(top_features.items(), 1):
print(f" {i:2d}. {feature:30s} {score:10.2f}")
#%%
# %%
# 确保输出目录存在
os.makedirs(OUTPUT_DIR, exist_ok=True)
@@ -731,7 +583,7 @@ print(f"\n 预览（前15行）:")
print(topn_to_save.head(15))
print("\n训练流程完成！")
#%% md
# %% md
# ## 5. 总结
# #
# 本 Notebook 实现了完整的 Learn-to-Rank 训练流程：
@@ -764,4 +616,4 @@ print("\n训练流程完成！")
# 2. **超参数调优**: 使用网格搜索优化 LambdaRank 参数
# 3. **模型集成**: 结合多个排序模型的预测
# 4. **更复杂的分组**: 考虑按行业分组排序
#
#

407
src/experiment/regression.ipynb
View File

@@ -15,7 +15,6 @@
"source": [
"import os\n",
"from datetime import datetime\n",
"from typing import List\n",
"\n",
"import polars as pl\n",
"\n",
@@ -25,7 +24,6 @@
" LightGBMModel,\n",
" STFilter,\n",
" StandardScaler,\n",
" # StockFilterConfig, # 已删除，使用 StockPoolManager + filter_func 替代\n",
" StockPoolManager,\n",
" Trainer,\n",
" Winsorizer,\n",
@@ -33,87 +31,27 @@
" check_data_quality,\n",
")\n",
"from src.training.config import TrainingConfig\n",
"\n"
]
},
{
"metadata": {},
"cell_type": "markdown",
"source": "## 2. 定义辅助函数"
},
{
"metadata": {},
"cell_type": "code",
"outputs": [],
"execution_count": null,
"source": [
"def register_factors(\n",
" engine: FactorEngine,\n",
" selected_factors: List[str],\n",
" factor_definitions: dict,\n",
" label_factor: dict,\n",
") -> List[str]:\n",
" \"\"\"注册因子（selected_factors 从 metadata 查询，factor_definitions 用 DSL 表达式注册）\"\"\"\n",
" print(\"=\" * 80)\n",
" print(\"注册因子\")\n",
" print(\"=\" * 80)\n",
"\n",
" # 注册 SELECTED_FACTORS 中的因子（已在 metadata 中）\n",
" print(\"\\n注册特征因子（从 metadata）:\")\n",
" for name in selected_factors:\n",
" engine.add_factor(name)\n",
" print(f\" - {name}\")\n",
"\n",
" # 注册 FACTOR_DEFINITIONS 中的因子（通过表达式，尚未在 metadata 中）\n",
" print(\"\\n注册特征因子（表达式）:\")\n",
" for name, expr in factor_definitions.items():\n",
" engine.add_factor(name, expr)\n",
" print(f\" - {name}: {expr}\")\n",
"\n",
" # 注册 label 因子（通过表达式）\n",
" print(\"\\n注册 Label 因子（表达式）:\")\n",
" for name, expr in label_factor.items():\n",
" engine.add_factor(name, expr)\n",
" print(f\" - {name}: {expr}\")\n",
"\n",
" # 特征列 = SELECTED_FACTORS + FACTOR_DEFINITIONS 的 keys\n",
" feature_cols = selected_factors + list(factor_definitions.keys())\n",
"\n",
" print(f\"\\n特征因子数: {len(feature_cols)}\")\n",
" print(f\" - 来自 metadata: {len(selected_factors)}\")\n",
" print(f\" - 来自表达式: {len(factor_definitions)}\")\n",
" print(f\"Label: {list(label_factor.keys())[0]}\")\n",
" print(f\"已注册因子总数: {len(engine.list_registered())}\")\n",
"\n",
" return feature_cols\n",
"\n",
"\n",
"def prepare_data(\n",
" engine: FactorEngine,\n",
" feature_cols: List[str],\n",
" start_date: str,\n",
" end_date: str,\n",
") -> pl.DataFrame:\n",
" print(\"\\n\" + \"=\" * 80)\n",
" print(\"准备数据\")\n",
" print(\"=\" * 80)\n",
"\n",
" # 计算因子（全市场数据）\n",
" print(f\"\\n计算因子: {start_date} - {end_date}\")\n",
" factor_names = feature_cols + [LABEL_NAME] # 包含 label\n",
"\n",
" data = engine.compute(\n",
" factor_names=factor_names,\n",
" start_date=start_date,\n",
" end_date=end_date,\n",
" )\n",
"\n",
" print(f\"数据形状: {data.shape}\")\n",
" print(f\"数据列: {data.columns}\")\n",
" print(f\"\\n前5行预览:\")\n",
" print(data.head())\n",
"\n",
" return data\n",
"# 从 common 模块导入共用配置和函数\n",
"from src.experiment.common import (\n",
" SELECTED_FACTORS,\n",
" FACTOR_DEFINITIONS,\n",
" get_label_factor,\n",
" register_factors,\n",
" prepare_data,\n",
" TRAIN_START,\n",
" TRAIN_END,\n",
" VAL_START,\n",
" VAL_END,\n",
" TEST_START,\n",
" TEST_END,\n",
" stock_pool_filter,\n",
" STOCK_FILTER_REQUIRED_COLUMNS,\n",
" OUTPUT_DIR,\n",
" SAVE_PREDICTIONS,\n",
" PERSIST_MODEL,\n",
" TOP_N,\n",
")\n",
"\n"
]
},
@@ -121,9 +59,9 @@
"metadata": {},
"cell_type": "markdown",
"source": [
"## 3. 配置参数\n",
"## 2. 配置参数\n",
"#\n",
"### 3.1 因子定义"
"### 2.1 标签定义"
]
},
{
@@ -132,177 +70,11 @@
"outputs": [],
"execution_count": null,
"source": [
"# 特征因子定义字典：新增因子只需在此处添加一行\n",
"# Label 名称（回归任务使用连续收益率）\n",
"LABEL_NAME = \"future_return_5\"\n",
"\n",
"# 当前选择的因子列表（从 FACTOR_DEFINITIONS 中选择要使用的因子）\n",
"SELECTED_FACTORS = [\n",
" # ================= 1. 价格、趋势与路径依赖 =================\n",
" \"ma_5\",\n",
" \"ma_20\",\n",
" \"ma_ratio_5_20\",\n",
" \"bias_10\",\n",
" \"high_low_ratio\",\n",
" \"bbi_ratio\",\n",
" \"return_5\",\n",
" \"return_20\",\n",
" \"kaufman_ER_20\",\n",
" \"mom_acceleration_10_20\",\n",
" \"drawdown_from_high_60\",\n",
" \"up_days_ratio_20\",\n",
" # ================= 2. 波动率、风险调整与高阶矩 =================\n",
" \"volatility_5\",\n",
" \"volatility_20\",\n",
" \"volatility_ratio\",\n",
" \"std_return_20\",\n",
" \"sharpe_ratio_20\",\n",
" \"min_ret_20\",\n",
" \"volatility_squeeze_5_60\",\n",
" # ================= 3. 日内微观结构与异象 =================\n",
" \"overnight_intraday_diff\",\n",
" \"upper_shadow_ratio\",\n",
" \"capital_retention_20\",\n",
" \"max_ret_20\",\n",
" # ================= 4. 量能、流动性与量价背离 =================\n",
" \"volume_ratio_5_20\",\n",
" \"turnover_rate_mean_5\",\n",
" \"turnover_deviation\",\n",
" \"amihud_illiq_20\",\n",
" \"turnover_cv_20\",\n",
" \"pv_corr_20\",\n",
" \"close_vwap_deviation\",\n",
" # ================= 5. 基本面财务特征 =================\n",
" \"roe\",\n",
" \"roa\",\n",
" \"profit_margin\",\n",
" \"debt_to_equity\",\n",
" \"current_ratio\",\n",
" \"net_profit_yoy\",\n",
" \"revenue_yoy\",\n",
" \"healthy_expansion_velocity\",\n",
" # ================= 6. 基本面估值与截面动量共振 =================\n",
" \"EP\",\n",
" \"BP\",\n",
" \"CP\",\n",
" \"market_cap_rank\",\n",
" \"turnover_rank\",\n",
" \"return_5_rank\",\n",
" \"EP_rank\",\n",
" \"pe_expansion_trend\",\n",
" \"value_price_divergence\",\n",
" \"active_market_cap\",\n",
" \"ebit_rank\",\n",
"]\n",
"\n",
"# 因子定义字典（完整因子库）\n",
"FACTOR_DEFINITIONS = {\n",
" # ================= 1. 价格、趋势与路径依赖 (Trend, Momentum & Path Dependency) =================\n",
" \"ma_5\": \"ts_mean(close, 5)\",\n",
" \"ma_20\": \"ts_mean(close, 20)\",\n",
" \"ma_ratio_5_20\": \"ts_mean(close, 5) / (ts_mean(close, 20) + 1e-8) - 1\", # 均线发散度\n",
" \"bias_10\": \"close / (ts_mean(close, 10) + 1e-8) - 1\", # 10日乖离率\n",
" \"high_low_ratio\": \"(close - ts_min(low, 20)) / (ts_max(high, 20) - ts_min(low, 20) + 1e-8)\", # 威廉指标变形\n",
" \"bbi_ratio\": \"(ts_mean(close, 3) + ts_mean(close, 6) + ts_mean(close, 12) + ts_mean(close, 24)) / (4 * close + 1e-8)\", # 多空指标比率\n",
" \"return_5\": \"(close / (ts_delay(close, 5) + 1e-8)) - 1\", # 5日动量\n",
" \"return_20\": \"(close / (ts_delay(close, 20) + 1e-8)) - 1\", # 20日动量\n",
" # [高阶] Kaufman 趋势效率 (极高价值) - 衡量趋势流畅度，剔除无序震荡\n",
" \"kaufman_ER_20\": \"abs(close - ts_delay(close, 20)) / (ts_sum(abs(close - ts_delay(close, 1)), 20) + 1e-8)\",\n",
" # [高阶] 动量加速度 - 寻找二阶导数大于0，正在加速爆发的股票\n",
" \"mom_acceleration_10_20\": \"(close / (ts_delay(close, 10) + 1e-8) - 1) - (ts_delay(close, 10) / (ts_delay(close, 20) + 1e-8) - 1)\",\n",
" # [高阶] 高点距离衰减 - 衡量套牢盘压力\n",
" \"drawdown_from_high_60\": \"close / (ts_max(high, 60) + 1e-8) - 1\",\n",
" # [高阶] 趋势一致性 - 过去20天内收红的天数比例\n",
" \"up_days_ratio_20\": \"ts_sum(close > ts_delay(close, 1), 20) / 20\",\n",
" # ================= 2. 波动率、风险调整与高阶矩 (Volatility & Risk-Adjusted Returns) =================\n",
" \"volatility_5\": \"ts_std(close, 5)\",\n",
" \"volatility_20\": \"ts_std(close, 20)\",\n",
" \"volatility_ratio\": \"ts_std(close, 5) / (ts_std(close, 20) + 1e-8)\", # 波动率期限结构\n",
" \"std_return_20\": \"ts_std((close / (ts_delay(close, 1) + 1e-8)) - 1, 20)\", # 真实收益率波动率\n",
" # [高阶] 夏普趋势比率 - 惩罚暴涨暴跌，奖励稳健爬坡\n",
" \"sharpe_ratio_20\": \"ts_mean(close / (ts_delay(close, 1) + 1e-8) - 1, 20) / (ts_std(close / (ts_delay(close, 1) + 1e-8) - 1, 20) + 1e-8)\",\n",
" # [高阶] 尾部崩盘风险 - 过去一个月最大单日跌幅\n",
" \"min_ret_20\": \"ts_min(close / (ts_delay(close, 1) + 1e-8) - 1, 20)\",\n",
" # [高阶] 波动率挤压比 - 寻找盘整到极致面临变盘的股票 (布林带收口)\n",
" \"volatility_squeeze_5_60\": \"ts_std(close, 5) / (ts_std(close, 60) + 1e-8)\",\n",
" # ================= 3. 日内微观结构与异象 (Intraday Microstructure & Anomalies) =================\n",
" # [高阶] 隔夜与日内背离 - 差值越小说明主力越喜欢在盘中吸筹\n",
" \"overnight_intraday_diff\": \"(open / (ts_delay(close, 1) + 1e-8) - 1) - (close / (open + 1e-8) - 1)\",\n",
" # [高阶] 上影线抛压极值 - 冲高回落被套牢的概率\n",
" \"upper_shadow_ratio\": \"(high - ((open + close + abs(open - close)) / 2)) / (high - low + 1e-8)\",\n",
" # [高阶] 资金沉淀率 - 衡量主力日内高抛低吸洗盘的剧烈程度\n",
" \"capital_retention_20\": \"ts_sum(abs(close - open), 20) / (ts_sum(high - low, 20) + 1e-8)\",\n",
" # [高阶] MAX 彩票效应 - 反转因子，剔除近期有过妖股连板特征的标的\n",
" \"max_ret_20\": \"ts_max(close / (ts_delay(close, 1) + 1e-8) - 1, 20)\",\n",
" # ================= 4. 量能、流动性与量价背离 (Volume, Liquidity & Divergence) =================\n",
" \"volume_ratio_5_20\": \"ts_mean(vol, 5) / (ts_mean(vol, 20) + 1e-8)\", # 相对放量比\n",
" \"turnover_rate_mean_5\": \"ts_mean(turnover_rate, 5)\", # 活跃度\n",
" \"turnover_deviation\": \"(turnover_rate - ts_mean(turnover_rate, 10)) / (ts_std(turnover_rate, 10) + 1e-8)\", # 换手率偏离度\n",
" # [高阶] Amihud 非流动性异象 (绝对核心) - 衡量砸盘/拉升的摩擦成本\n",
" \"amihud_illiq_20\": \"ts_mean(abs(close / (ts_delay(close, 1) + 1e-8) - 1) / (amount + 1e-8), 20)\",\n",
" # [高阶] 换手率惩罚因子 - 换手率忽高忽低说明游资接力，行情极不稳定\n",
" \"turnover_cv_20\": \"ts_std(turnover_rate, 20) / (ts_mean(turnover_rate, 20) + 1e-8)\",\n",
" # [高阶] 纯粹量价相关性 - 检验是否是\"放量上涨，缩量下跌\"的良性多头\n",
" \"pv_corr_20\": \"ts_corr(close / (ts_delay(close, 1) + 1e-8) - 1, vol, 20)\",\n",
" # [高阶] 收盘价与均价背离 - 专门抓尾盘突袭拉升骗线的股票\n",
" \"close_vwap_deviation\": \"close / (amount / (vol * 100 + 1e-8) + 1e-8) - 1\",\n",
" # ================= 5. 基本面财务特征 (Fundamental Quality & Structure) =================\n",
" \"roe\": \"n_income / (total_hldr_eqy_exc_min_int + 1e-8)\", # 净资产收益率\n",
" \"roa\": \"n_income / (total_assets + 1e-8)\", # 总资产收益率\n",
" \"profit_margin\": \"n_income / (revenue + 1e-8)\", # 销售净利率\n",
" \"debt_to_equity\": \"total_liab / (total_hldr_eqy_exc_min_int + 1e-8)\", # 杠杆率\n",
" \"current_ratio\": \"total_cur_assets / (total_cur_liab + 1e-8)\", # 短期偿债安全垫\n",
" # [高阶] 利润同比增速 (日频延后252天等于去年同期)\n",
" \"net_profit_yoy\": \"(n_income / (ts_delay(n_income, 252) + 1e-8)) - 1\",\n",
" # [高阶] 营收同比增速\n",
" \"revenue_yoy\": \"(revenue / (ts_delay(revenue, 252) + 1e-8)) - 1\",\n",
" # [高阶] 资产负债表扩张斜率 - 剔除单纯靠举债扩张的公司\n",
" \"healthy_expansion_velocity\": \"(total_assets / (ts_delay(total_assets, 252) + 1e-8) - 1) - (total_liab / (ts_delay(total_liab, 252) + 1e-8) - 1)\",\n",
" # ================= 6. 基本面估值与截面动量共振 (Valuation & Cross-Sectional Ranking) =================\n",
" # 估值水平绝对值 (Tushare 市值单位需要 * 10000 转换为元)\n",
" \"EP\": \"n_income / (total_mv * 10000 + 1e-8)\", # 盈利收益率 (1/PE)\n",
" \"BP\": \"total_hldr_eqy_exc_min_int / (total_mv * 10000 + 1e-8)\", # 账面市值比 (1/PB)\n",
" \"CP\": \"n_cashflow_act / (total_mv * 10000 + 1e-8)\", # 经营现金流收益率 (1/PCF)\n",
" # 全市场截面排名因子\n",
" \"market_cap_rank\": \"cs_rank(total_mv)\", # 规模因子 (Size)\n",
" \"turnover_rank\": \"cs_rank(turnover_rate)\",\n",
" \"return_5_rank\": \"cs_rank((close / (ts_delay(close, 5) + 1e-8)) - 1)\",\n",
" \"EP_rank\": \"cs_rank(n_income / (total_mv + 1e-8))\", # 谁最便宜\n",
" # [高阶] 戴维斯双击动量 - 估值相对上一年是否在扩张\n",
" \"pe_expansion_trend\": \"(total_mv / (n_income + 1e-8)) / (ts_delay(total_mv, 60) / (ts_delay(n_income, 60) + 1e-8) + 1e-8) - 1\",\n",
" # [高阶] 业绩与价格背离度 - 截面做差：利润排名全市场第一，但近20日价格排名倒数第一，捕捉被错杀的潜伏股\n",
" \"value_price_divergence\": \"cs_rank((n_income - ts_delay(n_income, 252)) / (abs(ts_delay(n_income, 252)) + 1e-8)) - cs_rank(close / (ts_delay(close, 20) + 1e-8))\",\n",
" # [高阶] 流动性溢价调整后市值 - 识别僵尸大盘股和极度活跃的小微盘\n",
" \"active_market_cap\": \"total_mv * ts_mean(turnover_rate, 20)\",\n",
" \"ebit_rank\": \"cs_rank(ebit)\",\n",
"}\n",
"\n",
"# Label 因子定义（不参与训练，用于计算目标）\n",
"LABEL_FACTOR = {\n",
" LABEL_NAME: \"(ts_delay(close, -5) / ts_delay(open, -1)) - 1\", # 未来5日收益率\n",
"}"
]
},
{
"metadata": {},
"cell_type": "markdown",
"source": "### 3.2 训练参数配置"
},
{
"metadata": {},
"cell_type": "code",
"outputs": [],
"execution_count": null,
"source": [
"# 日期范围配置（正确的 train/val/test 三分法）\n",
"# Train: 用于训练模型参数\n",
"# Val: 用于验证/早停/调参（位于 train 之后，test 之前）\n",
"# Test: 仅用于最终评估，完全独立于训练过程\n",
"TRAIN_START = \"20200101\"\n",
"TRAIN_END = \"20231231\"\n",
"VAL_START = \"20240101\"\n",
"VAL_END = \"20241231\"\n",
"TEST_START = \"20250101\"\n",
"TEST_END = \"20261231\"\n",
"# 获取 Label 因子定义\n",
"LABEL_FACTOR = get_label_factor(LABEL_NAME)\n",
"\n",
"# 模型参数配置\n",
"MODEL_PARAMS = {\n",
@@ -326,60 +98,7 @@
" # 数值稳定性\n",
" \"verbose\": -1,\n",
" \"random_state\": 42,\n",
"}\n",
"\n",
"\n",
"# 股票池筛选函数\n",
"# 使用新的 StockPoolManager API：传入自定义筛选函数和所需列/因子\n",
"# 筛选函数接收单日 DataFrame，返回布尔 Series\n",
"#\n",
"# 筛选逻辑（针对单日数据）：\n",
"# 1. 先排除创业板、科创板、北交所（ST过滤由STFilter组件处理）\n",
"# 2. 然后选取市值最小的500只股票\n",
"def stock_pool_filter(df: pl.DataFrame) -> pl.Series:\n",
" \"\"\"股票池筛选函数（单日数据）\n",
"\n",
" 筛选条件：\n",
" 1. 排除创业板（代码以 300 开头）\n",
" 2. 排除科创板（代码以 688 开头）\n",
" 3. 排除北交所（代码以 8、9 或 4 开头）\n",
" 4. 选取当日市值最小的500只股票\n",
" \"\"\"\n",
" # 代码筛选（排除创业板、科创板、北交所）\n",
" code_filter = (\n",
" ~df[\"ts_code\"].str.starts_with(\"30\") # 排除创业板\n",
" & ~df[\"ts_code\"].str.starts_with(\"68\") # 排除科创板\n",
" & ~df[\"ts_code\"].str.starts_with(\"8\") # 排除北交所\n",
" & ~df[\"ts_code\"].str.starts_with(\"9\") # 排除北交所\n",
" & ~df[\"ts_code\"].str.starts_with(\"4\") # 排除北交所\n",
" )\n",
"\n",
" # 在已筛选的股票中，选取市值最小的500只\n",
" # 按市值升序排序，取前500\n",
" valid_df = df.filter(code_filter)\n",
" n = min(1000, len(valid_df))\n",
" small_cap_codes = valid_df.sort(\"total_mv\").head(n)[\"ts_code\"]\n",
"\n",
" # 返回布尔 Series：是否在被选中的股票中\n",
" return df[\"ts_code\"].is_in(small_cap_codes)\n",
"\n",
"\n",
"# 定义筛选所需的基础列\n",
"STOCK_FILTER_REQUIRED_COLUMNS = [\"total_mv\"] # ST过滤由STFilter组件处理\n",
"\n",
"# 可选：定义筛选所需的因子（如果需要用因子进行筛选）\n",
"# STOCK_FILTER_REQUIRED_FACTORS = {\n",
"# \"market_cap_rank\": \"cs_rank(total_mv)\",\n",
"# }\n",
"\n",
"\n",
"# 输出配置（相对于本文件所在目录）\n",
"OUTPUT_DIR = \"output\"\n",
"SAVE_PREDICTIONS = True\n",
"PERSIST_MODEL = False\n",
"\n",
"# Top N 配置：每日推荐股票数量\n",
"TOP_N = 5 # 可调整为 10, 20 等"
"}"
]
},
{
@@ -420,6 +139,7 @@
" feature_cols=feature_cols,\n",
" start_date=TRAIN_START,\n",
" end_date=TEST_END,\n",
" label_name=LABEL_NAME,\n",
")\n",
"\n",
"# 4. 打印配置信息\n",
@@ -515,8 +235,6 @@
{
"metadata": {},
"cell_type": "code",
"outputs": [],
"execution_count": null,
"source": [
"# 步骤 2: 划分训练/验证/测试集（正确的三分法）\n",
"print(\"\\n[步骤 2/6] 划分训练集、验证集和测试集\")\n",
@@ -550,7 +268,9 @@
" train_data = filtered_data\n",
" test_data = filtered_data\n",
" print(\" 未配置划分器，全部作为训练集\")"
]
],
"outputs": [],
"execution_count": null
},
{
"metadata": {},
@@ -579,8 +299,6 @@
{
"metadata": {},
"cell_type": "code",
"outputs": [],
"execution_count": null,
"source": [
"# 步骤 4: 训练集数据处理\n",
"print(\"\\n[步骤 4/7] 训练集数据处理\")\n",
@@ -608,7 +326,9 @@
" null_count = train_data[col].null_count()\n",
" if null_count > 0:\n",
" print(f\" {col}: {null_count} ({null_count / len(train_data) * 100:.2f}%)\")"
]
],
"outputs": [],
"execution_count": null
},
{
"metadata": {},
@@ -828,8 +548,6 @@
{
"metadata": {},
"cell_type": "code",
"outputs": [],
"execution_count": null,
"source": [
"print(\"\\n\" + \"=\" * 80)\n",
"print(\"训练结果\")\n",
@@ -855,7 +573,9 @@
"sample_data = results.filter(results[\"trade_date\"] == sample_date).head(10)\n",
"print(f\"\\n示例日期 {sample_date} 的前10条预测:\")\n",
"print(sample_data.select([\"ts_code\", \"trade_date\", target_col, \"prediction\"]))"
]
],
"outputs": [],
"execution_count": null
},
{
"metadata": {},
@@ -978,6 +698,61 @@
"- 可以帮助理解哪些因子最有效"
]
},
{
"metadata": {},
"cell_type": "code",
"outputs": [],
"execution_count": null,
"source": [
"print(\"绘制特征重要性...\")\n",
"\n",
"fig, ax = plt.subplots(figsize=(10, 8))\n",
"lgb.plot_importance(\n",
" booster,\n",
" max_num_features=20,\n",
" importance_type=\"gain\",\n",
" title=\"Feature Importance (Gain)\",\n",
" ax=ax,\n",
")\n",
"ax.set_xlabel(\"Importance (Gain)\")\n",
"plt.tight_layout()\n",
"plt.show()\n",
"\n",
"# 打印重要性排名\n",
"importance_gain = pd.Series(\n",
" booster.feature_importance(importance_type=\"gain\"), index=feature_cols\n",
").sort_values(ascending=False)\n",
"\n",
"print(\"\\n[特征重要性排名 - Gain]\")\n",
"print(importance_gain)\n",
"\n",
"# 识别低重要性特征\n",
"zero_importance = importance_gain[importance_gain == 0].index.tolist()\n",
"if zero_importance:\n",
" print(f\"\\n[低重要性特征] 以下{len(zero_importance)}个特征重要性为0，可考虑删除:\")\n",
" for feat in zero_importance:\n",
" print(f\" - {feat}\")\n",
"else:\n",
" print(\"\\n所有特征都有一定重要性\")\n"
]
},
{
"metadata": {},
"cell_type": "code",
"outputs": [],
"execution_count": null,
"source": [
"# 导入可视化库\n",
"import matplotlib.pyplot as plt\n",
"import lightgbm as lgb\n",
"import pandas as pd\n",
"\n",
"# 从封装的model中取出底层Booster\n",
"booster = model.model\n",
"print(f\"模型类型: {type(booster)}\")\n",
"print(f\"特征数量: {len(feature_cols)}\")"
]
},
{
"metadata": {},
"cell_type": "code",

235
src/experiment/regression.py
View File

@@ -3,7 +3,6 @@
# %%
import os
from datetime import datetime
from typing import List
import polars as pl
@@ -13,7 +12,6 @@ from src.training import (
LightGBMModel,
STFilter,
StandardScaler,
# StockFilterConfig, # 已删除，使用 StockPoolManager + filter_func 替代
StockPoolManager,
Trainer,
Winsorizer,
@@ -22,167 +20,38 @@ from src.training import (
)
from src.training.config import TrainingConfig
# %% md
# ## 2. 定义辅助函数
# %%
def register_factors(
engine: FactorEngine,
selected_factors: List[str],
factor_definitions: dict,
label_factor: dict,
) -> List[str]:
"""注册因子（selected_factors 从 metadata 查询，factor_definitions 用 DSL 表达式注册）"""
print("=" * 80)
print("注册因子")
print("=" * 80)
# 注册 SELECTED_FACTORS 中的因子（已在 metadata 中）
print("\n注册特征因子（从 metadata）:")
for name in selected_factors:
engine.add_factor(name)
print(f" - {name}")
# 注册 FACTOR_DEFINITIONS 中的因子（通过表达式，尚未在 metadata 中）
print("\n注册特征因子（表达式）:")
for name, expr in factor_definitions.items():
engine.add_factor(name, expr)
print(f" - {name}: {expr}")
# 注册 label 因子（通过表达式）
print("\n注册 Label 因子（表达式）:")
for name, expr in label_factor.items():
engine.add_factor(name, expr)
print(f" - {name}: {expr}")
# 特征列 = SELECTED_FACTORS + FACTOR_DEFINITIONS 的 keys
feature_cols = selected_factors + list(factor_definitions.keys())
print(f"\n特征因子数: {len(feature_cols)}")
print(f" - 来自 metadata: {len(selected_factors)}")
print(f" - 来自表达式: {len(factor_definitions)}")
print(f"Label: {list(label_factor.keys())[0]}")
print(f"已注册因子总数: {len(engine.list_registered())}")
return feature_cols
def prepare_data(
engine: FactorEngine,
feature_cols: List[str],
start_date: str,
end_date: str,
) -> pl.DataFrame:
print("\n" + "=" * 80)
print("准备数据")
print("=" * 80)
# 计算因子（全市场数据）
print(f"\n计算因子: {start_date} - {end_date}")
factor_names = feature_cols + [LABEL_NAME] # 包含 label
data = engine.compute(
factor_names=factor_names,
start_date=start_date,
end_date=end_date,
)
print(f"数据形状: {data.shape}")
print(f"数据列: {data.columns}")
print(f"\n前5行预览:")
print(data.head())
return data
# 从 common 模块导入共用配置和函数
from src.experiment.common import (
SELECTED_FACTORS,
FACTOR_DEFINITIONS,
get_label_factor,
register_factors,
prepare_data,
TRAIN_START,
TRAIN_END,
VAL_START,
VAL_END,
TEST_START,
TEST_END,
stock_pool_filter,
STOCK_FILTER_REQUIRED_COLUMNS,
OUTPUT_DIR,
SAVE_PREDICTIONS,
PERSIST_MODEL,
TOP_N,
)
# %% md
# ## 3. 配置参数
# ## 2. 配置参数
#
# ### 3.1 因子定义
# ### 2.1 标签定义
# %%
# 特征因子定义字典：新增因子只需在此处添加一行
# Label 名称（回归任务使用连续收益率）
LABEL_NAME = "future_return_5"
# 当前选择的因子列表（从 FACTOR_DEFINITIONS 中选择要使用的因子）
SELECTED_FACTORS = [
# ================= 1. 价格、趋势与路径依赖 =================
"ma_5",
"ma_20",
"ma_ratio_5_20",
"bias_10",
"high_low_ratio",
"bbi_ratio",
"return_5",
"return_20",
"kaufman_ER_20",
"mom_acceleration_10_20",
"drawdown_from_high_60",
"up_days_ratio_20",
# ================= 2. 波动率、风险调整与高阶矩 =================
"volatility_5",
"volatility_20",
"volatility_ratio",
"std_return_20",
"sharpe_ratio_20",
"min_ret_20",
"volatility_squeeze_5_60",
# ================= 3. 日内微观结构与异象 =================
"overnight_intraday_diff",
"upper_shadow_ratio",
"capital_retention_20",
"max_ret_20",
# ================= 4. 量能、流动性与量价背离 =================
"volume_ratio_5_20",
"turnover_rate_mean_5",
"turnover_deviation",
"amihud_illiq_20",
"turnover_cv_20",
"pv_corr_20",
"close_vwap_deviation",
# ================= 5. 基本面财务特征 =================
"roe",
"roa",
"profit_margin",
"debt_to_equity",
"current_ratio",
"net_profit_yoy",
"revenue_yoy",
"healthy_expansion_velocity",
# ================= 6. 基本面估值与截面动量共振 =================
"EP",
"BP",
"CP",
"market_cap_rank",
"turnover_rank",
"return_5_rank",
"EP_rank",
"pe_expansion_trend",
"value_price_divergence",
"active_market_cap",
"ebit_rank",
]
# 因子定义字典（完整因子库）
FACTOR_DEFINITIONS = {
}
# Label 因子定义（不参与训练，用于计算目标）
LABEL_FACTOR = {
LABEL_NAME: "(ts_delay(close, -5) / ts_delay(open, -1)) - 1", # 未来5日收益率
}
# %% md
# ### 3.2 训练参数配置
# %%
# 日期范围配置（正确的 train/val/test 三分法）
# Train: 用于训练模型参数
# Val: 用于验证/早停/调参（位于 train 之后，test 之前）
# Test: 仅用于最终评估，完全独立于训练过程
TRAIN_START = "20200101"
TRAIN_END = "20231231"
VAL_START = "20240101"
VAL_END = "20241231"
TEST_START = "20250101"
TEST_END = "20261231"
# 获取 Label 因子定义
LABEL_FACTOR = get_label_factor(LABEL_NAME)
# 模型参数配置
MODEL_PARAMS = {
@@ -207,59 +76,6 @@ MODEL_PARAMS = {
"verbose": -1,
"random_state": 42,
}
# 股票池筛选函数
# 使用新的 StockPoolManager API：传入自定义筛选函数和所需列/因子
# 筛选函数接收单日 DataFrame，返回布尔 Series
#
# 筛选逻辑（针对单日数据）：
# 1. 先排除创业板、科创板、北交所（ST过滤由STFilter组件处理）
# 2. 然后选取市值最小的500只股票
def stock_pool_filter(df: pl.DataFrame) -> pl.Series:
"""股票池筛选函数（单日数据）
筛选条件：
1. 排除创业板（代码以 300 开头）
2. 排除科创板（代码以 688 开头）
3. 排除北交所（代码以 8、9 或 4 开头）
4. 选取当日市值最小的500只股票
"""
# 代码筛选（排除创业板、科创板、北交所）
code_filter = (
~df["ts_code"].str.starts_with("30") # 排除创业板
& ~df["ts_code"].str.starts_with("68") # 排除科创板
& ~df["ts_code"].str.starts_with("8") # 排除北交所
& ~df["ts_code"].str.starts_with("9") # 排除北交所
& ~df["ts_code"].str.starts_with("4") # 排除北交所
)
# 在已筛选的股票中，选取市值最小的500只
# 按市值升序排序，取前500
valid_df = df.filter(code_filter)
n = min(1000, len(valid_df))
small_cap_codes = valid_df.sort("total_mv").head(n)["ts_code"]
# 返回布尔 Series：是否在被选中的股票中
return df["ts_code"].is_in(small_cap_codes)
# 定义筛选所需的基础列
STOCK_FILTER_REQUIRED_COLUMNS = ["total_mv"] # ST过滤由STFilter组件处理
# 可选：定义筛选所需的因子（如果需要用因子进行筛选）
# STOCK_FILTER_REQUIRED_FACTORS = {
# "market_cap_rank": "cs_rank(total_mv)",
# }
# 输出配置（相对于本文件所在目录）
OUTPUT_DIR = "output"
SAVE_PREDICTIONS = True
PERSIST_MODEL = False
# Top N 配置：每日推荐股票数量
TOP_N = 5 # 可调整为 10, 20 等
# %% md
# ## 4. 训练流程
#
@@ -288,6 +104,7 @@ data = prepare_data(
feature_cols=feature_cols,
start_date=TRAIN_START,
end_date=TEST_END,
label_name=LABEL_NAME,
)
# 4. 打印配置信息