🎯 本周回顾
这一周我们干了一件事:把散落的单因子拼成一条能每周跑出名单的流水线。D1 用等权 / IC 加权 / 最大化 IR 把因子合成分数,实测年化差距 3.6 个点但 OOS 会反转,说明等权基准不能丢。D2 用 Gram-Schmidt 和对称正交处理共线性,条件是 IC 掉 20% 到 40% 但 IR 从 1.28 涨到 1.63。D3 把分数变成名单,Top1% 只比 Top5% 多 4 个点却放大 12 到 15 个点回撤,缓冲区把月换手从 45% 压到 22%。D4 用申万一级做行业 + 市值双重中性化,年化从 18.6% 降到 14.2%,但行业贡献从 41% 降到 8%。D5 算出净收益在季频见顶 10.0%,加缓冲后月频反超到 11.7%。D6 把每只票的入选理由拆成因子贡献,防止「不知道为什么赚钱」。
📋 本周方法回顾表
| 天 | 方法 | 关键数字 | 落地规则 |
|---|---|---|---|
| D1 | 等权 / IC 加权 / 最大化 IR | 13.2% vs 16.8%,OOS 反转 | 等权做永久基准 |
| D2 | 顺序正交 / 对称正交 | IR 1.28 → 1.51 → 1.63 | 看 IR 与回撤,不看 IC |
| D3 | 分位选股 + 缓冲区 | 换手 45% → 22% | 进 100 留 130 |
| D4 | 行业 + 市值中性化 | 18.6% → 14.2%,行业贡献 41% → 8% | 申万一级分组 z-score |
| D5 | 调仓频率与换手 | 季频净 10.0%,缓冲月频 11.7% | 按最短半衰期反推 |
| D6 | 因子贡献拆解 | 集中度 > 0.6 需复核 | 每期归档 holdings.csv |
💻 每周 Top20 完整流水线
import numpy as np
import pandas as pd
# ══════════════════════════════════════════════════════════
# 输入约定(由你的数据层提供,此脚本只做加工)
# panel: MultiIndex(date, code) 的长表,含:
# factor_value factor_quality factor_growth factor_momentum
# sw_l1 mktcap amount is_st suspended list_days
# forward_ret: 用于计算历史 IC(只用于权重估计,不参与选股)
# ══════════════════════════════════════════════════════════
FACTORS = ["factor_value", "factor_quality", "factor_growth", "factor_momentum"]
W = {"factor_value": 0.35, "factor_quality": 0.30,
"factor_growth": 0.20, "factor_momentum": 0.15}
COST_ONE_WAY = 0.0013
# ── 步骤 1:可投资性过滤 ────────────────────────────────
def filter_universe(df):
m = (~df["is_st"]) & (~df["suspended"]) & \
(df["list_days"] >= 60) & (df["amount"] >= 0.2)
return df[m].copy()
# ── 步骤 2:因子清洗(缩尾 + 行业/市值中性化)─────────
def winsorize(s, p=0.01):
return s.clip(s.quantile(p), s.quantile(1 - p))
def industry_neutralize(s, ind, min_n=10):
g = s.groupby(ind)
z = (s - g.transform("mean")) / g.transform("std").replace(0, np.nan)
gz = (s - s.mean()) / s.std(ddof=0)
return z.where(g.transform("count") >= min_n, gz).fillna(gz)
def size_neutralize(s, mktcap):
lm = np.log(mktcap.values.astype(float))
X = np.column_stack([np.ones(len(lm)), lm])
beta = np.linalg.lstsq(X, s.values, rcond=None)[0]
r = s.values - X @ beta
return pd.Series(r / r.std(ddof=0), index=s.index)
def neutralize_all(df):
out = df.copy()
for f in FACTORS:
z = winsorize(out[f].astype(float))
z = industry_neutralize(z, out["sw_l1"])
out[f + "_neu"] = size_neutralize(z, out["mktcap"])
return out
# ── 步骤 3:多因子合成 ──────────────────────────────────
def composite_score(df):
cols = [f + "_neu" for f in FACTORS]
w = pd.Series([W[f] for f in FACTORS], index=cols)
return df[cols].mul(w, axis=1).sum(axis=1)
# ── 步骤 4:缓冲区选股(进 100 / 留 130)────────────────
def select(ranked, prev_hold=None, n_new=20, n_keep=26):
ranked = ranked.sort_values("score", ascending=False).reset_index(drop=True)
ranked["rank"] = ranked.index + 1
if prev_hold is None:
return ranked.head(n_new)["code"].tolist()
keep = ranked[ranked["code"].isin(prev_hold) &
(ranked["rank"] <= n_keep)]["code"].tolist()
new = ranked[ranked["rank"] <= n_new - min(len(keep), n_new)]["code"].tolist()
picks = list(dict.fromkeys(keep + new))[:n_new]
# 数量不足时按排名补齐
for c in ranked["code"]:
if len(picks) >= n_new:
break
if c not in picks:
picks.append(c)
return picks
# ── 步骤 5:等权 + 5% 上限 ──────────────────────────────
def weights(picks, cap=0.05):
w = pd.Series(1.0 / len(picks), index=picks).clip(upper=cap)
return w / w.sum()
# ── 步骤 6:可解释性报告 ────────────────────────────────
def explain(df, picks):
cols = [f + "_neu" for f in FACTORS]
w = pd.Series([W[f] for f in FACTORS], index=cols)
sub = df.loc[df["code"].isin(picks)].set_index("code").loc[picks]
contrib = sub[cols].mul(w, axis=1)
contrib.columns = [c.replace("_neu", "") for c in cols]
rep = pd.DataFrame({
"score": sub["score"].round(3),
"行业": sub["sw_l1"],
"主导因子": contrib.idxmax(axis=1).str.replace("factor_", ""),
"集中度": (contrib.abs().max(axis=1) / contrib.abs().sum(axis=1)).round(2),
"市值分位": (sub["mktcap"].rank(pct=True)).round(2),
}, index=sub.index)
rep["风险"] = np.where(rep["集中度"] > 0.6, "单因子依赖", "")
cnt = rep["行业"].value_counts()
rep.loc[rep["行业"].map(cnt) == 1, "风险"] += " 行业内独苗"
return rep.sort_values("score", ascending=False)
# ══════════════════════════════════════════════════════════
# 主流程
# ══════════════════════════════════════════════════════════
def weekly_picks(panel_snapshot, prev_hold=None):
"""panel_snapshot: 某个交易日(T 日收盘)的截面 DataFrame"""
df = filter_universe(panel_snapshot)
if len(df) < 100:
raise ValueError(f"可投资池仅 {len(df)} 只,检查数据完整性")
df = neutralize_all(df)
df["score"] = composite_score(df)
picks = select(df, prev_hold=prev_hold, n_new=20, n_keep=26)
w = weights(picks)
rep = explain(df, picks)
rep["权重"] = w.round(4)
return picks, w, rep
# ── 示例:合成一份随机截面跑通全流程 ─────────────────────
if __name__ == "__main__":
np.random.seed(2026)
n = 800
snap = pd.DataFrame({
"code": [f"{600000 + i:06d}" for i in range(n)],
"sw_l1": np.random.choice(
["电子", "医药生物", "银行", "机械设备", "计算机", "食品饮料",
"电力设备", "有色金属", "煤炭", "基础化工"], n),
"mktcap": np.exp(np.random.randn(n) * 1.2 + 3),
"amount": np.exp(np.random.randn(n) + 1),
"is_st": np.random.rand(n) < 0.04,
"suspended": np.random.rand(n) < 0.01,
"list_days": np.random.randint(5, 3000, n),
"factor_value": np.random.randn(n),
"factor_quality": np.random.randn(n),
"factor_growth": np.random.randn(n),
"factor_momentum": np.random.randn(n),
})
picks, w, rep = weekly_picks(snap)
print(f"入选 {len(picks)} 只,权重和 {w.sum():.6f}")
print(rep.to_string())
print("\n行业分布:\n", rep["行业"].value_counts())
# 归档
# rep.to_csv("weekly_picks_20261123.csv", encoding="utf-8-sig")
✏️ 今日动手
把上面这段脚本接到你的数据层上,产出第一期真实名单。跑通之后每周固定一天(建议周五收盘后)执行,把 CSV 归档到 picks/YYYYMMDD.csv。连续做 8 周,你就有了一份自己的选股记录——它比任何回测都更能告诉你系统的真实边界。
🔖 下周预告
第5周进入组合构建与风险管理:名单有了,接下来是每只买多少。D1 从等权走向风险平价与波动率倒数加权,D2 讲协方差估计的收缩方法(Ledoit-Wolf),D3 做行业与风格的组合层约束,D4 讲最大回撤控制与动态减仓,D5 讲仓位管理与凯利公式在A股的实用修正,D6 讲组合归因(Brinson 与因子归因),D7 总结成一份可执行的组合管理手册。