289 lines
11 KiB
Python
289 lines
11 KiB
Python
|
"""
|
|||
|
|
验证问题2.3因子差异根本原因的测试
|
|||
|
|
|
|||
|
|
理论分析：
|
|||
|
|
1. GTJA_alpha032: cs_rank(high)和cs_rank(vol)是截面排名
|
|||
|
|
- 不同LOOKBACK_DAYS下载面股票数量可能不同（3Y包含更多历史股票）
|
|||
|
|
- cs_rank = rank/count，count不同导致排名分母不同
|
|||
|
|
|
|||
|
|
2. GTJA_alpha077: ts_decay_linear + cs_rank
|
|||
|
|
- ts_decay_linear使用np.convolve，边界效应
|
|||
|
|
- cs_rank嵌套导致排名基准变化
|
|||
|
|
|
|||
|
|
3. GTJA_alpha121: ts_rank嵌套ts_corr
|
|||
|
|
- ts_rank使用滑动窗口，边界敏感
|
|||
|
|
- ts_corr的边界效应叠加
|
|||
|
|
|
|||
|
|
验证方法：直接计算2Y和3Y数据下的截面股票数量差异
|
|||
|
|
"""
|
|||
|
|
|
|||
|
|
import polars as pl
|
|||
|
|
from datetime import datetime, timedelta
|
|||
|
|
from src.factors import FactorEngine
|
|||
|
|
import numpy as np
|
|||
|
|
|
|||
|
|
|
|||
|
|
# =============================================================================
|
|||
|
|
# 配置
|
|||
|
|
# =============================================================================
|
|||
|
|
PREDICT_START = "20250101"
|
|||
|
|
PREDICT_END = "20250131"
|
|||
|
|
LOOKBACK_2Y = 365 * 3
|
|||
|
|
LOOKBACK_3Y = 365 * 4
|
|||
|
|
|
|||
|
|
|
|||
|
|
def get_lookback_start_date(start_date: str, lookback_days: int) -> str:
|
|||
|
|
start_dt = datetime.strptime(start_date, "%Y%m%d")
|
|||
|
|
lookback_dt = start_dt - timedelta(days=lookback_days)
|
|||
|
|
return lookback_dt.strftime("%Y%m%d")
|
|||
|
|
|
|||
|
|
|
|||
|
|
def verify_cross_section_stock_count():
|
|||
|
|
"""
|
|||
|
|
验证：不同LOOKBACK_DAYS下载面股票数量是否不同
|
|||
|
|
|
|||
|
|
关键问题：如果3Y数据包含更多历史股票，那么在相同日期D，
|
|||
|
|
2Y和3Y的截面股票数量可能不同，导致cs_rank分母不同
|
|||
|
|
"""
|
|||
|
|
print("=" * 80)
|
|||
|
|
print("验证截面股票数量差异")
|
|||
|
|
print("=" * 80)
|
|||
|
|
|
|||
|
|
# 加载2Y数据
|
|||
|
|
actual_start_2y = get_lookback_start_date(PREDICT_START, LOOKBACK_2Y)
|
|||
|
|
engine_2y = FactorEngine()
|
|||
|
|
data_2y = engine_2y.compute(
|
|||
|
|
factor_names=["close"],
|
|||
|
|
start_date=actual_start_2y,
|
|||
|
|
end_date=PREDICT_END,
|
|||
|
|
)
|
|||
|
|
data_2y = data_2y.filter(data_2y["trade_date"] >= PREDICT_START)
|
|||
|
|
|
|||
|
|
# 加载3Y数据
|
|||
|
|
actual_start_3y = get_lookback_start_date(PREDICT_START, LOOKBACK_3Y)
|
|||
|
|
engine_3y = FactorEngine()
|
|||
|
|
data_3y = engine_3y.compute(
|
|||
|
|
factor_names=["close"],
|
|||
|
|
start_date=actual_start_3y,
|
|||
|
|
end_date=PREDICT_END,
|
|||
|
|
)
|
|||
|
|
data_3y = data_3y.filter(data_3y["trade_date"] >= PREDICT_START)
|
|||
|
|
|
|||
|
|
# 统计截面股票数量
|
|||
|
|
stocks_per_date_2y = (
|
|||
|
|
data_2y.group_by("trade_date")
|
|||
|
|
.agg(pl.col("ts_code").count().alias("stock_count"))
|
|||
|
|
.sort("trade_date")
|
|||
|
|
)
|
|||
|
|
|
|||
|
|
stocks_per_date_3y = (
|
|||
|
|
data_3y.group_by("trade_date")
|
|||
|
|
.agg(pl.col("ts_code").count().alias("stock_count"))
|
|||
|
|
.sort("trade_date")
|
|||
|
|
)
|
|||
|
|
|
|||
|
|
print("\n2Y数据 - 每天截面股票数量:")
|
|||
|
|
print(stocks_per_date_2y)
|
|||
|
|
|
|||
|
|
print("\n3Y数据 - 每天截面股票数量:")
|
|||
|
|
print(stocks_per_date_3y)
|
|||
|
|
|
|||
|
|
# 比较差异
|
|||
|
|
comparison = stocks_per_date_2y.join(
|
|||
|
|
stocks_per_date_3y, on="trade_date", suffix="_3y"
|
|||
|
|
).with_columns(
|
|||
|
|
[(pl.col("stock_count_3y") - pl.col("stock_count")).alias("count_diff")]
|
|||
|
|
)
|
|||
|
|
|
|||
|
|
print("\n股票数量差异 (3Y - 2Y):")
|
|||
|
|
print(comparison)
|
|||
|
|
|
|||
|
|
diff_count = comparison.filter(pl.col("count_diff") != 0).height
|
|||
|
|
print(f"\n有差异的日期数: {diff_count}")
|
|||
|
|
if diff_count > 0:
|
|||
|
|
print("结论：2Y和3Y的截面股票数量确实不同！")
|
|||
|
|
else:
|
|||
|
|
print("结论：2Y和3Y的截面股票数量相同，cs_rank分母应该一致")
|
|||
|
|
|
|||
|
|
|
|||
|
|
def verify_cs_rank_formula_behavior():
|
|||
|
|
"""
|
|||
|
|
验证cs_rank在不同count下的行为
|
|||
|
|
|
|||
|
|
场景：如果2Y截面有3000只股票，3Y截面有3100只股票
|
|||
|
|
假设股票X的rank都是1500（中间位置）
|
|||
|
|
- 2Y: cs_rank = 1500/3000 = 0.5
|
|||
|
|
- 3Y: cs_rank = 1500/3100 ≈ 0.484
|
|||
|
|
"""
|
|||
|
|
print("\n" + "=" * 80)
|
|||
|
|
print("验证 cs_rank 在不同 count 下的行为")
|
|||
|
|
print("=" * 80)
|
|||
|
|
|
|||
|
|
# 模拟：相同rank但不同count的情况
|
|||
|
|
df = pl.DataFrame(
|
|||
|
|
{
|
|||
|
|
"trade_date": ["20250101"] * 2,
|
|||
|
|
"ts_code": ["001", "002"],
|
|||
|
|
"rank": [1500.0, 1500.0], # 相同rank
|
|||
|
|
"count_2y": [3000, 3000], # 2Y count
|
|||
|
|
"count_3y": [3100, 3100], # 3Y count
|
|||
|
|
}
|
|||
|
|
)
|
|||
|
|
|
|||
|
|
result = df.with_columns(
|
|||
|
|
[
|
|||
|
|
(pl.col("rank") / pl.col("count_2y")).alias("cs_rank_2y"),
|
|||
|
|
(pl.col("rank") / pl.col("count_3y")).alias("cs_rank_3y"),
|
|||
|
|
]
|
|||
|
|
)
|
|||
|
|
|
|||
|
|
print("模拟结果（rank=1500的情况）:")
|
|||
|
|
print(
|
|||
|
|
result.select(
|
|||
|
|
["ts_code", "rank", "count_2y", "cs_rank_2y", "count_3y", "cs_rank_3y"]
|
|||
|
|
)
|
|||
|
|
)
|
|||
|
|
print(f"\n差异: cs_rank_2y - cs_rank_3y = {1500 / 3000 - 1500 / 3100:.6f}")
|
|||
|
|
|
|||
|
|
|
|||
|
|
def analyze_alpha032_deep():
|
|||
|
|
"""
|
|||
|
|
深入分析GTJA_alpha032 = -1 * ts_sum(cs_rank(ts_corr(cs_rank(high), cs_rank(vol), 3)), 3)
|
|||
|
|
|
|||
|
|
差异来源层级：
|
|||
|
|
1. cs_rank(high) - 截面排名，count不同导致差异
|
|||
|
|
2. cs_rank(vol) - 截面排名，count不同导致差异
|
|||
|
|
3. ts_corr(..., 3) - 滚动相关，使用上述排名结果作为输入
|
|||
|
|
4. cs_rank(ts_corr(...)) - 嵌套：对ts_corr结果再做截面排名
|
|||
|
|
5. ts_sum(..., 3) - 滚动求和
|
|||
|
|
"""
|
|||
|
|
print("\n" + "=" * 80)
|
|||
|
|
print("GTJA_alpha032 深入分析")
|
|||
|
|
print("=" * 80)
|
|||
|
|
|
|||
|
|
print("""
|
|||
|
|
公式: (-1 * ts_sum(cs_rank(ts_corr(cs_rank(high), cs_rank(vol), 3)), 3))
|
|||
|
|
|
|||
|
|
差异传递链:
|
|||
|
|
┌─────────────────────────────────────────────────────────────┐
|
|||
|
|
│ L1: high, vol │
|
|||
|
|
│ ↓ │
|
|||
|
|
│ L2: cs_rank(high), cs_rank(vol) │
|
|||
|
|
│ 问题: 2Y和3Y的截面股票数量可能不同 │
|
|||
|
|
│ 例如: 股票X在2Y下cs_rank=1500/3000=0.5 │
|
|||
|
|
│ 股票X在3Y下cs_rank=1500/3100=0.484 │
|
|||
|
|
│ ↓ │
|
|||
|
|
│ L3: ts_corr(cs_rank(high), cs_rank(vol), 3) │
|
|||
|
|
│ 问题: 输入的cs_rank值已经不同，ts_corr结果自然不同 │
|
|||
|
|
│ ↓ │
|
|||
|
|
│ L4: cs_rank(ts_corr(...)) │
|
|||
|
|
│ 问题: 每天对ts_corr结果做截面排名，count可能不同 │
|
|||
|
|
│ ↓ │
|
|||
|
|
│ L5: ts_sum(..., 3) │
|
|||
|
|
│ 问题: ts_sum对L4的排名结果求和，边界效应 │
|
|||
|
|
│ ↓ │
|
|||
|
|
│ L6: -1 * ... │
|
|||
|
|
└─────────────────────────────────────────────────────────────┘
|
|||
|
|
|
|||
|
|
结论：GTJA_alpha032的差异主要来自L2的cs_rank嵌套问题
|
|||
|
|
""")
|
|||
|
|
|
|||
|
|
|
|||
|
|
def analyze_alpha077_deep():
|
|||
|
|
"""
|
|||
|
|
深入分析GTJA_alpha077 = min_(cs_rank(ts_decay_linear(...)), cs_rank(ts_decay_linear(...)))
|
|||
|
|
|
|||
|
|
涉及：ts_decay_linear（np.convolve边界效应）+ cs_rank嵌套
|
|||
|
|
"""
|
|||
|
|
print("\n" + "=" * 80)
|
|||
|
|
print("GTJA_alpha077 深入分析")
|
|||
|
|
print("=" * 80)
|
|||
|
|
|
|||
|
|
print("""
|
|||
|
|
公式: min_(cs_rank(ts_decay_linear(DECAY1)), cs_rank(ts_decay_linear(DECAY2)))
|
|||
|
|
|
|||
|
|
其中:
|
|||
|
|
DECAY1 = ((((high + low) / 2) + high) - ((amount / vol) + high))
|
|||
|
|
DECAY2 = ts_corr(((high + low) / 2), ts_mean(vol, 40), 3)
|
|||
|
|
|
|||
|
|
差异来源:
|
|||
|
|
1. ts_decay_linear使用np.convolve
|
|||
|
|
- mode='valid' 只返回完全重叠的结果
|
|||
|
|
- 边界处(前window-1个)数据会是NaN
|
|||
|
|
- 但2Y和3Y的边界位置相同，结果应该一样
|
|||
|
|
|
|||
|
|
2. 真正的问题：cs_rank(ts_decay_linear(...))
|
|||
|
|
- 每天对ts_decay_linear结果做截面排名
|
|||
|
|
- 2Y和3Y截面股票数量不同 → cs_rank分母不同 → 结果不同
|
|||
|
|
|
|||
|
|
结论：差异主要来自cs_rank嵌套问题
|
|||
|
|
""")
|
|||
|
|
|
|||
|
|
|
|||
|
|
def analyze_alpha121_deep():
|
|||
|
|
"""
|
|||
|
|
深入分析GTJA_alpha121 = (cs_rank(...) ** ts_rank(ts_corr(...), 3)) * -1
|
|||
|
|
|
|||
|
|
涉及：ts_rank（滑动窗口边界敏感）+ ts_corr + cs_rank嵌套
|
|||
|
|
"""
|
|||
|
|
print("\n" + "=" * 80)
|
|||
|
|
print("GTJA_alpha121 深入分析")
|
|||
|
|
print("=" * 80)
|
|||
|
|
|
|||
|
|
print("""
|
|||
|
|
公式: ((cs_rank(((amount / vol) - min_((amount / vol), 12))) ** ts_rank(ts_corr(...), 3)) * -1)
|
|||
|
|
|
|||
|
|
差异来源:
|
|||
|
|
1. ts_rank(ts_corr(...), 3)
|
|||
|
|
- ts_rank使用sliding_window_view
|
|||
|
|
- 对边界敏感：不同起始点导致边界处有效窗口数量不同
|
|||
|
|
- ts_corr本身也有边界效应
|
|||
|
|
|
|||
|
|
2. cs_rank((amount / vol) - min_(..., 12))
|
|||
|
|
- 截面排名问题
|
|||
|
|
- 2Y和3Y截面股票数量不同
|
|||
|
|
|
|||
|
|
3. cs_rank(...) ** ts_rank(...)
|
|||
|
|
- 嵌套：截面排名结果作为指数
|
|||
|
|
- 两个差异来源叠加
|
|||
|
|
|
|||
|
|
结论：差异来自ts_rank/ts_corr边界效应 + cs_rank嵌套问题
|
|||
|
|
""")
|
|||
|
|
|
|||
|
|
|
|||
|
|
def summarize_root_causes():
|
|||
|
|
"""
|
|||
|
|
总结问题2.7因子的根本原因
|
|||
|
|
"""
|
|||
|
|
print("\n" + "=" * 80)
|
|||
|
|
print("问题2.3因子差异根本原因总结")
|
|||
|
|
print("=" * 80)
|
|||
|
|
|
|||
|
|
print("""
|
|||
|
|
┌─────────────────┬──────────────────────────────────────────────────────┐
|
|||
|
|
│ 因子 │ 根本原因 │
|
|||
|
|
├─────────────────┼──────────────────────────────────────────────────────┤
|
|||
|
|
│ GTJA_alpha016 │ cs_rank嵌套：ts_corr结果再做截面排名，count不同 │
|
|||
|
|
│ GTJA_alpha032 │ cs_rank嵌套：cs_rank(high)和cs_rank(vol)截面count不同 │
|
|||
|
|
│ GTJA_alpha077 │ ts_decay_linear边界效应 + cs_rank嵌套 │
|
|||
|
|
│ GTJA_alpha091 │ cs_rank嵌套：max_(close,5)结果再做截面排名 │
|
|||
|
|
│ GTJA_alpha121 │ ts_rank/ts_corr边界效应 + cs_rank嵌套 │
|
|||
|
|
│ GTJA_alpha130 │ ts_decay_linear边界效应 + cs_rank嵌套 │
|
|||
|
|
│ GTJA_alpha141 │ cs_rank嵌套：ts_corr结果再做截面排名 │
|
|||
|
|
└─────────────────┴──────────────────────────────────────────────────────┘
|
|||
|
|
|
|||
|
|
核心问题：cs_rank是截面排名函数，当不同LOOKBACK_DAYS下截面股票数量
|
|||
|
|
不同时，cs_rank的分母(count)不同，导致归一化排名结果不同。
|
|||
|
|
|
|||
|
|
这是一个结构性问题，与具体实现无关。
|
|||
|
|
""")
|
|||
|
|
|
|||
|
|
|
|||
|
|
if __name__ == "__main__":
|
|||
|
|
verify_cross_section_stock_count()
|
|||
|
|
verify_cs_rank_formula_behavior()
|
|||
|
|
analyze_alpha032_deep()
|
|||
|
|
analyze_alpha077_deep()
|
|||
|
|
analyze_alpha121_deep()
|
|||
|
|
summarize_root_causes()
|