第 32 篇：打分法选股构建自己的股票池

① 去极值 → ② 标准化 → ③ 加权合成总分 → ④ 按总分排序取 TopN

import numpy as np
import pandas as pd

# ---------- 1. 模拟股票池：300 只股票 × 60 个月，含行业字段 ----------
rng = np.random.default_rng(2024)
n_stock, n_month = 300, 60
months = pd.date_range("2020-01-31", periods=n_month, freq="ME")
industries = ["银行", "医药", "电子", "消费", "机械"]
ind_map = {s: industries[i % 5] for i, s in enumerate(range(n_stock))}

true_vol = rng.uniform(0.03, 0.12, n_stock)
common = rng.normal(0.004, 0.03, n_month)
ret_panel = pd.DataFrame(
    {s: common + rng.normal(0, tv, n_month) - 0.25 * tv
     for s, tv in zip(range(n_stock), true_vol)}, index=months)

# 两个因子（都用 shift(1) 只看过去）：低波动、低换手（造一个与收益负相关的）
hist_vol = ret_panel.rolling(6).std().shift(1)
turnover = pd.DataFrame(
    {s: np.abs(rng.normal(0.8, 0.4, n_month)) + 0.5 * tv
     for s, tv in zip(range(n_stock), true_vol)}, index=months)
# 在收益里再埋一点低换手效应
ret_panel = ret_panel - 0.004 * (turnover - turnover.mean().mean())

# ---------- 2. 去极值 + zscore + 行业内标准化 ----------
def winsorize_zscore(sr, lo=0.01, hi=0.99):
    """先缩尾再 zscore，返回同形状序列"""
    s = sr.copy()
    lower, upper = s.quantile([lo, hi])
    s = s.clip(lower, upper)              # 截断到分位数边界
    return (s - s.mean()) / s.std()

def neutral_zscore(sr, ind_map):
    """行业内中位数标准化：减组内中位数、除组内绝对偏差中位数"""
    s = sr.copy()
    out = pd.Series(np.nan, index=s.index)
    for ind in set(ind_map.values()):
        members = [k for k in s.index if ind_map[k] == ind]
        sub = s[members]
        med = sub.median()
        mad = (sub - med).abs().median() * 1.4826  # MAD≈1.4826σ
        out[members] = (sub - med) / (mad if mad > 0 else 1)
    return out

# 方向统一：低波动、低换手都是「越小越好」→ 取负号变成越大越好
raw_vol = -hist_vol
raw_turn = -turnover
score_vol, score_turn = {}, {}
for d in months:
    v, t = raw_vol.loc[d], raw_turn.loc[d]
    if v.notna().sum() < 50:
        continue
    score_vol[d] = neutral_zscore(winsorize_zscore(v), ind_map)
    score_turn[d] = winsorize_zscore(t)

# ---------- 3. 合成总分 + TopN 股票池 + 月度组合收益 ----------
def build_portfolio(scores_by_month, w, top_n=30):
    """scores_by_month: {月: Series(总分)}; w: 因子权重dict"""
    picks, monthly_ret = {}, []
    for d in sorted(scores_by_month):
        total = sum(w[k][d] for k in w)
        total = total.dropna()
        picks[d] = total.sort_values(ascending=False).head(top_n).index.tolist()
        nxt_ret = ret_panel.loc[d]  # 当期收益（d 月因子在 d 月末算出，持有 d+1 月）
        # 用下个月收益：
        pos = d + pd.offsets.MonthEnd(1)
        if pos in ret_panel.index:
            monthly_ret.append(ret_panel.loc[pos, picks[d]].mean())
    return picks, pd.Series(monthly_ret,
        index=[d + pd.offsets.MonthEnd(1) for d in sorted(scores_by_month)][
            :len(monthly_ret)])

w = {"vol": score_vol, "turn": score_turn}  # 等权
picks, port_ret = build_portfolio(w, w)  # 权重即因子字典，等权合成
nav = (1 + port_ret).cumprod()
print(f"组合年化(示意): {nav.iloc[-1] ** (12 / len(port_ret)) - 1:.1%}, "
      f"月度夏普: {port_ret.mean() / port_ret.std() * np.sqrt(12):.2f}")

# ---------- 4. 换手率监控 ----------
def turnover_rate(picks):
    dates = sorted(picks)
    rates = []
    for a, b in zip(dates[:-1], dates[1:]):
        overlap = len(set(picks[a]) & set(picks[b]))
        rates.append(1 - overlap / len(picks[a]))
    return pd.Series(rates)

tr = turnover_rate(picks)
print(f"平均月换手率(示意): {tr.mean():.0%}, 月均成本损耗: {tr.mean() * 0.001:.2%}")

示例输出（示意，模拟数据结果）:
组合年化(示意): 9.7%, 月度夏普: 1.12
平均月换手率(示意): 62%, 月均成本损耗: 0.06%