Sources: extracted from two upstream archives (skill-repo, skills-main),
merged with the following policy:
- 15 broken symlinks (pointing to /Users/jameslee/.cc-switch/skills or
../../.agents/skills on a foreign machine) discarded
- 3 real name collisions with identical content (ai-pair, ifind-http-api,
zhipu-websearch) kept as one copy
- Functional overlaps deduped keeping the strongest variant:
- docx family: kept docx (official, full toolchain) + docx-cn
(GB/T 9704 Chinese official-document constants),
dropped docx_writer (no scripts, name collided with docx)
- humanizer family: kept humanizer-zh (6 zh reference docs),
dropped humanizer (en, redundant for CN workflow)
- Skills that only ran in a foreign environment removed:
ablemind-ops, app-publish, hlb-design-system, openclaw-adj-skill,
claude-driver
- alphapai excluded from this public repo because its SKILL.md hard-coded
live credentials
Result: 25 skills, 572 files, ~7.5 MB.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
8.7 KiB
8.7 KiB
风险指标公式参考
指标速查表
| 指标 | 含义 | 越大越好? | 典型范围 |
|---|---|---|---|
| 年化收益率 | 复合年化回报 | 是 | -20% ~ 50% |
| 年化波动率 | 收益率标准差年化 | 否 | 5% ~ 40% |
| 最大回撤 | 峰谷最大跌幅 | 否 | 5% ~ 60% |
| 夏普比率 | 单位风险超额收益 | 是 | -1 ~ 3 |
| 信息比率 | 单位跟踪误差超额收益 | 是 | -1 ~ 2 |
| 卡尔玛比率 | 年化收益 / 最大回撤 | 是 | 0 ~ 5 |
| 索提诺比率 | 单位下行风险超额收益 | 是 | -1 ~ 5 |
| 胜率 | 正收益天数占比 | 是 | 40% ~ 65% |
1. 年化收益率 (Annualized Return)
公式:
R_{ann} = \left(\frac{NAV_{end}}{NAV_{start}}\right)^{\frac{252}{n}} - 1
其中 n 为交易日天数,252 为 A 股年交易日。
Python:
def annualized_return(nav: pd.Series) -> float:
"""nav: 日频净值序列,index 为日期"""
total_return = nav.iloc[-1] / nav.iloc[0]
n_days = len(nav) - 1
if n_days <= 0:
return 0.0
return total_return ** (252 / n_days) - 1
Excel 公式:
= (末日净值/首日净值) ^ (252 / (交易日数-1)) - 1
2. 年化波动率 (Annualized Volatility)
公式:
\sigma_{ann} = \text{std}(r_t) \times \sqrt{252}
其中 r_t = \ln(NAV_t / NAV_{t-1}) 为日对数收益率。
Python:
def annualized_volatility(nav: pd.Series) -> float:
"""nav: 日频净值序列"""
log_returns = np.log(nav / nav.shift(1)).dropna()
return log_returns.std() * np.sqrt(252)
Excel 公式:
= STDEV(日收益率区域) * SQRT(252)
注意: 使用对数收益率
LN(B3/B2)而非简单收益率,因对数收益率可加性更好。Excel 中若已有简单收益率列,可直接用STDEV近似。
3. 最大回撤 (Maximum Drawdown)
公式:
MDD = \max_{t} \left(\frac{\text{RunningMax}_t - NAV_t}{\text{RunningMax}_t}\right)
Python:
def max_drawdown(nav: pd.Series) -> float:
"""返回正数(如 0.15 表示回撤 15%)"""
running_max = nav.cummax()
drawdown = (running_max - nav) / running_max
return drawdown.max()
Excel 辅助列方法:
辅助列 RunningMax: C2 = MAX($B$2:B2) (向下填充)
辅助列 Drawdown: D2 = (C2 - B2) / C2
最大回撤: = MAX(D:D)
4. 夏普比率 (Sharpe Ratio)
公式:
Sharpe = \frac{R_{ann} - R_f}{\sigma_{ann}}
其中 R_f 为无风险利率,A 股通常取一年期定存利率(当前约 1.5%)或 0。
Python:
def sharpe_ratio(nav: pd.Series, rf: float = 0.015) -> float:
ann_ret = annualized_return(nav)
ann_vol = annualized_volatility(nav)
if ann_vol == 0:
return 0.0
return (ann_ret - rf) / ann_vol
Excel 公式:
= (年化收益率 - 无风险利率) / 年化波动率
5. 信息比率 (Information Ratio)
公式:
IR = \frac{R_{p,ann} - R_{b,ann}}{TE}
其中 TE = \text{std}(r_p - r_b) \times \sqrt{252} 为年化跟踪误差。
Python:
def information_ratio(nav_p: pd.Series, nav_b: pd.Series) -> float:
"""nav_p: 组合净值, nav_b: 基准净值(需对齐日期)"""
ret_p = np.log(nav_p / nav_p.shift(1)).dropna()
ret_b = np.log(nav_b / nav_b.shift(1)).dropna()
# 对齐
common = ret_p.index.intersection(ret_b.index)
excess = ret_p.loc[common] - ret_b.loc[common]
te = excess.std() * np.sqrt(252)
if te == 0:
return 0.0
ann_excess = annualized_return(nav_p) - annualized_return(nav_b)
return ann_excess / te
Excel 公式:
跟踪误差 = STDEV(组合日收益 - 基准日收益) * SQRT(252)
信息比率 = (组合年化 - 基准年化) / 跟踪误差
6. 卡尔玛比率 (Calmar Ratio)
公式:
Calmar = \frac{R_{ann}}{MDD}
Python:
def calmar_ratio(nav: pd.Series) -> float:
mdd = max_drawdown(nav)
if mdd == 0:
return 0.0
return annualized_return(nav) / mdd
Excel 公式:
= 年化收益率 / 最大回撤
7. 索提诺比率 (Sortino Ratio)
公式:
Sortino = \frac{R_{ann} - R_f}{\sigma_{down}}
其中 \sigma_{down} = \text{std}(\min(r_t - r_f/252, 0)) \times \sqrt{252} 为下行波动率。
Python:
def sortino_ratio(nav: pd.Series, rf: float = 0.015) -> float:
log_returns = np.log(nav / nav.shift(1)).dropna()
daily_rf = rf / 252
downside = log_returns[log_returns < daily_rf] - daily_rf
if len(downside) == 0:
return 0.0
downside_vol = downside.std() * np.sqrt(252)
if downside_vol == 0:
return 0.0
return (annualized_return(nav) - rf) / downside_vol
Excel 公式:
下行波动率辅助列: = IF(日收益<无风险日利率, 日收益-无风险日利率, 0)
下行标准差: = STDEV(IF(辅助列<>0, 辅助列)) (Ctrl+Shift+Enter 数组公式)
索提诺比率: = (年化收益 - 无风险利率) / (下行标准差 * SQRT(252))
8. 胜率 (Win Rate)
公式:
WinRate = \frac{\text{count}(r_t > 0)}{n}
Python:
def win_rate(nav: pd.Series) -> float:
daily_returns = nav.pct_change().dropna()
if len(daily_returns) == 0:
return 0.0
return (daily_returns > 0).sum() / len(daily_returns)
Excel 公式:
= COUNTIF(日收益率区域, ">0") / COUNT(日收益率区域)
Brinson 三因素归因
公式:
| 因子 | 公式 | 含义 |
|---|---|---|
| 配置效应 | (w_p^i - w_b^i) \times (r_b^i - r_b) |
超配高收益行业的贡献 |
| 选股效应 | w_b^i \times (r_p^i - r_b^i) |
行业内个股选择的贡献 |
| 交互效应 | (w_p^i - w_b^i) \times (r_p^i - r_b^i) |
配置与选股的交叉项 |
| 总归因 | 配置 + 选股 + 交互 | 各因子之和 |
其中:
w_p^i: 组合中行业i的权重w_b^i: 基准中行业i的权重r_p^i: 组合中行业i的收益率r_b^i: 基准中行业i的收益率r_b: 基准整体收益率
Python:
def compute_brinson(
port_weights: pd.Series,
bench_weights: pd.Series,
port_returns: pd.Series,
bench_returns: pd.Series,
) -> pd.DataFrame:
"""
port_weights: 组合行业权重 (index=行业名)
bench_weights: 基准行业权重
port_returns: 组合行业收益率
bench_returns: 基准行业收益率
"""
bench_total = (bench_weights * bench_returns).sum()
allocation = (port_weights - bench_weights) * (bench_returns - bench_total)
selection = bench_weights * (port_returns - bench_returns)
interaction = (port_weights - bench_weights) * (port_returns - bench_returns)
total = allocation + selection + interaction
return pd.DataFrame({
"配置效应": allocation,
"选股效应": selection,
"交互效应": interaction,
"总归因": total,
})
Excel 公式(假设 A=行业, B=组合权重, C=基准权重, D=组合收益, E=基准收益):
基准整体收益 G1: = SUMPRODUCT(C:C, E:E)
配置效应 F2: = (B2-C2) * (E2-$G$1)
选股效应 G2: = C2 * (D2-E2)
交互效应 H2: = (B2-C2) * (D2-E2)
总归因 I2: = F2 + G2 + H2
辅助函数
对齐两个净值序列
def align_nav(nav1: pd.Series, nav2: pd.Series) -> tuple[pd.Series, pd.Series]:
"""按日期取交集并对齐"""
common = nav1.index.intersection(nav2.index)
return nav1.loc[common], nav2.loc[common]
滚动波动率
def rolling_volatility(nav: pd.Series, window: int = 20) -> pd.Series:
"""window 日滚动年化波动率"""
log_ret = np.log(nav / nav.shift(1))
return log_ret.rolling(window).std() * np.sqrt(252)
回撤序列
def drawdown_series(nav: pd.Series) -> pd.Series:
"""返回每日回撤幅度(正数表示回撤)"""
running_max = nav.cummax()
return (running_max - nav) / running_max
Top-N 回撤
def top_drawdowns(nav: pd.Series, n: int = 5) -> pd.DataFrame:
"""提取前 N 次最大回撤的起止日期和幅度"""
dd = drawdown_series(nav)
results = []
remaining = dd.copy()
for _ in range(n):
if remaining.max() == 0:
break
end_idx = remaining.idxmax()
# 向前找到回撤起点(上一个0回撤点)
before = remaining.loc[:end_idx]
start_candidates = before[before == 0]
start_idx = start_candidates.index[-1] if len(start_candidates) > 0 else before.index[0]
results.append({
"起始日期": start_idx,
"最低点日期": end_idx,
"回撤幅度": remaining[end_idx],
})
# 清除这段区间
remaining.loc[start_idx:end_idx] = 0
return pd.DataFrame(results)