claude_dev
2ff8ecaf52
compute_metrics rolling alpha/beta 原 for 循环每点 iloc[:i+1].cov/var = O(n²)。日线 111 天没事,分钟级 daily_df(万行)卡死(cta_c287f4b5 卡 5min CPU88%)。改 pandas expanding 向量化(cov/var/mean O(n)),数值与原循环一致(beta/alpha diff<1e-15,NaN 位置匹配)。pytest 6 绿。触发:b2c41c7 降级让 compute_metrics 首次在 benchmark 有数据时真跑(此前 benchmark 空被 skip),暴露 rolling O(n²) 性能 bug。
102 lines
4.9 KiB
Python
102 lines
4.9 KiB
Python
"""回测相对/绝对指标计算（empyrical，聚宽同源口径）。纯函数。"""
|
||
from dataclasses import dataclass, field
|
||
from typing import Dict, Literal
|
||
import math
|
||
import numpy as np
|
||
# empyrical 0.5.5 引用了 NumPy 2.0 已移除的别名（np.NINF / np.NaN / np.Inf / np.PINF），
|
||
# 不补回会在 sortino_ratio/downside_risk 等函数里抛 AttributeError，导致整块相对指标计算
|
||
# 失败、_metrics.json 不生成、结果页回退到 vnpy 原始字段（单位混乱）。
|
||
for _alias, _val in (("NINF", -np.inf), ("Inf", np.inf), ("PINF", np.inf),
|
||
("NaN", np.nan), ("NAN", np.nan), ("infty", np.inf)):
|
||
if not hasattr(np, _alias):
|
||
setattr(np, _alias, _val)
|
||
import pandas as pd
|
||
import empyrical
|
||
|
||
BenchmarkCode = Literal["hs300", "zz500"]
|
||
BENCHMARK_SYMBOL: Dict[str, str] = {"hs300": "sh000300", "zz500": "sz000905"}
|
||
|
||
|
||
@dataclass
|
||
class MetricsResult:
|
||
scalars: Dict[str, float] = field(default_factory=dict)
|
||
series: Dict[str, pd.Series] = field(default_factory=dict)
|
||
|
||
|
||
def compute_metrics(
|
||
daily_df: pd.DataFrame,
|
||
benchmark_returns: pd.Series,
|
||
period: int = 252,
|
||
) -> MetricsResult:
|
||
"""对 vnpy daily_df + 基准日收益计算聚宽级指标。
|
||
|
||
H4: 从 daily_df["balance"] 自算 simple return（vnpy df["return"] 是 log return，
|
||
empyrical 期望 simple return，直接用会失真）。
|
||
H5: period='daily' → empyrical 内部 252 交易日年化，口径统一。
|
||
benchmark 对齐：reindex 到策略交易日 + ffill，不丢策略日期（原 dropna 会丢停牌日）。
|
||
|
||
daily_df: vnpy calculate_result() 产出，须含 "balance" 列（账户余额），index 为日期。
|
||
benchmark_returns: 基准日收益率 Series，index 为日期（不必与 daily_df 对齐）。
|
||
period: 年化周期（整数，默认252交易日），empyrical 内部使用 'daily'
|
||
"""
|
||
# H4: simple return from balance（pct_change 首项 NaN → 0）
|
||
if "balance" not in daily_df.columns:
|
||
raise ValueError("daily_df 缺少 balance 列，无法计算 simple return")
|
||
strat = daily_df["balance"].astype(float).pct_change().fillna(0)
|
||
|
||
# benchmark ffill 对齐：reindex 到策略交易日，前向填充，不丢策略日期
|
||
b = benchmark_returns.reindex(daily_df.index).ffill().fillna(0)
|
||
|
||
aligned = pd.concat([strat.rename("s"), b.rename("b")], axis=1)
|
||
s, b = aligned["s"], aligned["b"]
|
||
|
||
scalars = {
|
||
"total_return": float(empyrical.cum_returns_final(s)),
|
||
"annual_return": float(empyrical.annual_return(s, period='daily')),
|
||
"alpha": float(empyrical.alpha(s, b, period='daily')),
|
||
"beta": float(empyrical.beta(s, b)),
|
||
"sharpe_ratio": float(empyrical.sharpe_ratio(s, period='daily')),
|
||
"sortino_ratio": float(empyrical.sortino_ratio(s, period='daily')),
|
||
"information_ratio": float(empyrical.excess_sharpe(s, b)),
|
||
"annual_volatility": float(empyrical.annual_volatility(s, period='daily')),
|
||
"max_drawdown": float(empyrical.max_drawdown(s)),
|
||
"benchmark_return": float(empyrical.cum_returns_final(b)),
|
||
"benchmark_volatility": float(empyrical.annual_volatility(b, period='daily')),
|
||
}
|
||
# Sanitize non-finite floats (NaN/Inf from degenerate inputs) → None for JSON safety
|
||
scalars = {k: (None if isinstance(v, float) and not math.isfinite(v) else v) for k, v in scalars.items()}
|
||
|
||
equity = empyrical.cum_returns(s)
|
||
bench_curve = empyrical.cum_returns(b)
|
||
# rolling alpha/beta (expanding window 用于画图，口径由 scalars 保证)
|
||
# 向量化 expanding (O(n)) 替代原 for 循环 (O(n²))——分钟级 daily_df(万行)原循环卡死
|
||
# (cta_c287f4b5 卡 5min CPU88% 根因)。expanding.cov/.var/.mean 数值与原循环一致
|
||
# (ddof=1)；var_b<=0(含单点 NaN)处 beta=NaN，与原 len>=2 且 var>0 守卫等价。
|
||
cov_sb = aligned["s"].expanding().cov(aligned["b"])
|
||
var_b = aligned["b"].expanding().var()
|
||
roll_beta = cov_sb / var_b.where(var_b > 0)
|
||
roll_alpha = (aligned["s"].expanding().mean() - roll_beta * aligned["b"].expanding().mean()) * period
|
||
# drawdown: 从峰值回落 (值 <= 0)
|
||
cummax = equity.cummax()
|
||
drawdown = (equity - cummax) / cummax
|
||
|
||
# Rolling annualized volatility (quarterly window) for the volatility chart
|
||
_vol_window = min(63, len(s))
|
||
if _vol_window >= 2:
|
||
vol_strategy = s.rolling(_vol_window, min_periods=2).std() * np.sqrt(period)
|
||
vol_benchmark = b.rolling(_vol_window, min_periods=2).std() * np.sqrt(period)
|
||
else:
|
||
vol_strategy = pd.Series([np.nan] * len(s), index=s.index)
|
||
vol_benchmark = pd.Series([np.nan] * len(s), index=s.index)
|
||
|
||
series = {
|
||
"equity_curve": equity,
|
||
"benchmark_curve": bench_curve,
|
||
"alpha": roll_alpha,
|
||
"beta": roll_beta,
|
||
"drawdown": drawdown,
|
||
"volatility_strategy": vol_strategy,
|
||
"volatility_benchmark": vol_benchmark,
|
||
}
|
||
return MetricsResult(scalars=scalars, series=series)
|