simon/myquant
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: 量化引擎加固 — 新增测试 + 数据/因子/回测层优化

Browse Source 
- 新增 finance/tests/ 6 个测试套件（agents/backtest/dao_upsert/factors/features/fundamental_lookahead）
- 数据层: data_manager / dao 优化，新增 upsert 逻辑
- 因子层: 基本面因子抽象定位 _mapping、ROE/PE/PB 重构
- 回测层: vectorbt/engine 大改动（251 行），report 增强
- ML 层: features/backtest_integration 特征工程与回测优化
- CLI: agent_cli 重构
- config/settings 扩充配置项
This commit is contained in:
Simon
2026-08-31 14:01:06 +08:00
parent 6acf938caf
commit 73d191b43a
28 changed files with 1418 additions and 373 deletions
Download Patch File Download Diff File Expand all files Collapse all files
finance/factors/fundamental/_mapping.py
+100
View File
@@ -0,0 +1,100 @@
"""
财务数据 → 日线映射（负责消除披露时点前视偏差）。
A 股季报的披露日远晚于报告期末：
- 一季报 / 年报 最迟约 4/30
- 中报 最迟约 8/31
- 三季报 最迟约 10/31
若直接把报告期 end_date 当日就"看到"本期财务结果，会引入前视偏差。
本模块统一在 end_date 上叠加一个保守的披露滞后：
1) 财务帧中若带 ann_date（实际披露日），则优先用 ann_date 作为可用日；
2) 否则按报告期月份推断法定披露时点，作为保守可用日。
"""
from __future__ import annotations
import pandas as pd
def _disclosure_available_date(end_date_ts: pd.Timestamp) -> pd.Timestamp:
"""根据报告期末推断法定披露可用日（无 ann_date 时的保守近似）。
一季度(0331)→4/30；中报(0630)→8/31；三季报(0930)→10/31；年报(1231)→次年4/30。
"""
if end_date_ts.month == 3 and end_date_ts.day == 31:
return pd.Timestamp(year=end_date_ts.year, month=4, day=30)
if end_date_ts.month == 6 and end_date_ts.day == 30:
return pd.Timestamp(year=end_date_ts.year, month=8, day=31)
if end_date_ts.month == 9 and end_date_ts.day == 30:
return pd.Timestamp(year=end_date_ts.year, month=10, day=31)
# 年报 12/31 → 次年 4/30
return pd.Timestamp(year=end_date_ts.year + 1, month=4, day=30)
def _eff_available_dates(fina_df: pd.DataFrame) -> pd.DataFrame:
"""计算每期财务的"可用日"（取最小滞后：有 ann_date 用它，否则法定截止）。"""
fina = fina_df.copy()
# 数值型 ann_date（YYYYMMDD）→ datetime；缺失用报告期末推断
if "ann_date" in fina.columns:
ann = pd.to_datetime(fina["ann_date"].astype(str), format="%Y%m%d", errors="coerce")
else:
ann = pd.Series(pd.NaT, index=fina.index)
end = pd.to_datetime(
fina["end_date"].astype(str), format="%Y%m%d", errors="coerce")
avail = ann.fillna(pd.Series(
[_disclosure_available_date(x) if not pd.isna(x) else pd.NaT for x in end],
index=end.index,
))
# 极少数 ann_date 早于报告期末（脏数据）时兜底用期末
avail = avail.where(avail >= end, end)
fina["_avail"] = avail
return fina
def effective_available_dates(fina_df: pd.DataFrame) -> pd.DataFrame:
"""
(公开) 返回带 _avail（披露可用日）的财务帧，供同比/环比等因子复用。
要求 fina_df 至少含 'end_date'；可选 'ann_date'。
"""
return _eff_available_dates(fina_df)
def map_fundamental_to_daily(
daily_df: pd.DataFrame,
fina_df: pd.DataFrame,
column: str,
) -> pd.Series:
"""
将季度财务数据按披露可用日映射到日线索引（前值填充）。
- 优先按 ann_date（实际披露日）对齐；
- 无 ann_date 时按报告期末的法定披露截止日保守对齐；
- 从而避免"财报在披露日之前就被回测看到"的前视偏差。
"""
if column not in fina_df.columns or "end_date" not in fina_df.columns:
return pd.Series(float("nan"), index=daily_df.index)
cols = ["end_date", column] + (["ann_date"] if "ann_date" in fina_df.columns else [])
fina = fina_df[cols].dropna(subset=["end_date", column]).copy()
if fina.empty:
return pd.Series(float("nan"), index=daily_df.index)
fina = _eff_available_dates(fina)
fina = fina.sort_values("_avail")
dates = pd.to_datetime(daily_df.index, format="%Y%m%d", errors="coerce")
result = pd.Series(float("nan"), index=daily_df.index)
avail_dates = fina["_avail"].values
for i, avail_dt in enumerate(avail_dates):
if pd.isna(avail_dt):
continue
mask = dates >= avail_dt
if i + 1 < len(avail_dates) and not pd.isna(avail_dates[i + 1]):
mask &= dates < avail_dates[i + 1]
result[mask] = fina[column].iloc[i]
return result.astype("float64")
finance/factors/fundamental/pe_pb.py
+4 -29
View File
@@ -7,6 +7,7 @@ PE / PB 估值因子。
import pandas as pd
from factors.base import BaseFactor
from factors.fundamental._mapping import map_fundamental_to_daily
class PEFactor(BaseFactor):
@@ -31,7 +32,7 @@ class PEFactor(BaseFactor):
if self._financial_df is None or self._financial_df.empty:
return pd.Series(float("nan"), index=df.index)
eps_series = _map_to_daily(df, self._financial_df, "eps")
eps_series = map_fundamental_to_daily(df, self._financial_df, "eps")
close = df["close"]
return close / eps_series.replace(0, float("nan"))
@@ -56,7 +57,7 @@ class PBFactor(BaseFactor):
if self._financial_df is None or self._financial_df.empty:
return pd.Series(float("nan"), index=df.index)
bvps_series = _map_to_daily(df, self._financial_df, "bvps")
bvps_series = map_fundamental_to_daily(df, self._financial_df, "bvps")
return df["close"] / bvps_series.replace(0, float("nan"))
def get_required_columns(self) -> list[str]:
@@ -80,34 +81,8 @@ class EPFactor(BaseFactor):
if self._financial_df is None or self._financial_df.empty:
return pd.Series(float("nan"), index=df.index)
eps_series = _map_to_daily(df, self._financial_df, "eps")
eps_series = map_fundamental_to_daily(df, self._financial_df, "eps")
return eps_series / df["close"].replace(0, float("nan")) * 100
def get_required_columns(self) -> list[str]:
return ["close"]
def _map_to_daily(
daily_df: pd.DataFrame,
fina_df: pd.DataFrame,
column: str,
) -> pd.Series:
"""将季度财务数据填充到日线索引（前值填充）。"""
fina = fina_df[["end_date", column]].dropna().copy()
fina["end_date"] = fina["end_date"].astype(str)
fina = fina.sort_values("end_date")
result = pd.Series(float("nan"), index=daily_df.index)
if fina.empty:
return result
dates = pd.to_datetime(daily_df.index, format="%Y%m%d", errors="coerce")
fina_dates = pd.to_datetime(fina["end_date"], format="%Y%m%d", errors="coerce")
for i, fina_date in enumerate(fina_dates):
mask = dates >= fina_date
if i + 1 < len(fina_dates):
mask &= dates < fina_dates.iloc[i + 1]
result[mask] = fina[column].iloc[i]
return result.astype("float64")
finance/factors/fundamental/roe.py
+27 -33
View File
@@ -5,6 +5,7 @@ ROE 因子。
import pandas as pd
from factors.base import BaseFactor
from factors.fundamental._mapping import map_fundamental_to_daily, effective_available_dates
class ROEFactor(BaseFactor):
@@ -45,28 +46,8 @@ class ROEFactor(BaseFactor):
fina_df: pd.DataFrame,
column: str,
) -> pd.Series:
"""将财务数据（季度）映射到日线索引。"""
fina = fina_df[["end_date", column]].dropna().copy()
fina["end_date"] = fina["end_date"].astype(str)
fina = fina.sort_values("end_date")
result = pd.Series(float("nan"), index=daily_df.index)
if fina.empty:
return result
dates = pd.to_datetime(daily_df.index, format="%Y%m%d", errors="coerce")
fina_dates = pd.to_datetime(fina["end_date"], format="%Y%m%d", errors="coerce")
for i, fina_date in enumerate(fina_dates):
mask = dates >= fina_date
if i + 1 < len(fina_dates):
mask &= dates < fina_dates.iloc[i + 1]
else:
pass # 最新一期覆盖所有后续日期
result[mask] = fina[column].iloc[i]
return result.astype("float64")
"""将财务数据（季度）映射到日线索引，按披露可用日消除前视。"""
return map_fundamental_to_daily(daily_df, fina_df, column)
class ROETTMDeltaFactor(BaseFactor):
@@ -82,23 +63,36 @@ class ROETTMDeltaFactor(BaseFactor):
if self._financial_df is None or self._financial_df.empty:
return pd.Series(float("nan"), index=df.index)
fina = self._financial_df[["end_date", "roe"]].dropna().copy()
fina["end_date"] = fina["end_date"].astype(str)
fina["year"] = fina["end_date"].str[:4].astype(int)
fina = fina.sort_values("end_date")
if "roe" not in self._financial_df.columns or "end_date" not in self._financial_df.columns:
return pd.Series(float("nan"), index=df.index)
# 基于财务帧计算披露可用日（含 ann_date 优先/法定滞后兜底）
fina = effective_available_dates(self._financial_df)
fina = fina[["_avail", "end_date", "roe"]].dropna(subset=["_avail", "roe"]).copy()
if fina.empty:
return pd.Series(float("nan"), index=df.index)
# 报告期月份（A 股季末：03/06/09/12）
end_dt = pd.to_datetime(fina["end_date"].astype(str), format="%Y%m%d", errors="coerce")
fina["report_month"] = end_dt.dt.month
fina["report_year"] = end_dt.dt.year
# 本年取数映射：{ (year, month): roe }
cur_map = dict(zip(zip(fina["report_year"], fina["report_month"]), fina["roe"]))
# 按披露可用日排序，逐期覆盖区间
fina = fina.sort_values("_avail")
# 按年分组计算 YoY 差值
roe_delta = pd.Series(float("nan"), index=df.index)
dates = pd.to_datetime(df.index, format="%Y%m%d", errors="coerce")
for _, row in fina.iterrows():
this_year = row["year"]
prev_row = fina[fina["year"] == this_year - 1]
if prev_row.empty:
avail_dt = row["_avail"]
if pd.isna(avail_dt):
continue
delta = row["roe"] - prev_row["roe"].iloc[-1]
f_date = pd.to_datetime(row["end_date"], format="%Y%m%d")
mask = dates >= f_date
prev_roe = cur_map.get((row["report_year"] - 1, row["report_month"]))
if prev_roe is None:
continue
delta = row["roe"] - prev_roe
mask = dates >= avail_dt
roe_delta[mask] = delta
return roe_delta.astype("float64")
finance/factors/registry.py
+7 -3
View File
@@ -101,6 +101,10 @@ def get_factor(name: str, **overrides) -> BaseFactor:
返回:
BaseFactor 实例
说明: 返回实例的 .name 恒等于注册键 name，即使构造器默认生成的
.name 与注册键不同（如 boll 的构造器默认 .name='boll_20'）——
注册键是列名的唯一事实源，避免 FactorEngine.compute 的列名漂移。
"""
if name not in _BUILTIN_FACTORIES:
raise KeyError(f"未知因子: '{name}'。可用: {list(_BUILTIN_FACTORIES)}")
@@ -109,9 +113,9 @@ def get_factor(name: str, **overrides) -> BaseFactor:
for k, v in overrides.items():
if hasattr(factor, k):
setattr(factor, k, v)
# 更新 factor.name
if hasattr(factor, "name"):
factor.name = name
# 无论是否覆盖参数，都强制 .name = 注册键，保证与分类/策略的引用一致
if hasattr(factor, "name"):
factor.name = name
return factor
finance/factors/sentiment/news_source.py
+16 -9
View File
@@ -10,6 +10,7 @@
"""
import json
import logging
import os
import time
from datetime import datetime, timedelta
@@ -21,6 +22,8 @@ import requests
# 确保 .env 已加载
import config.settings # noqa: F401
logger = logging.getLogger(__name__)
class NewsSource:
"""
@@ -87,7 +90,7 @@ class NewsSource:
frames.append(df)
time.sleep(self.request_delay)
except Exception as e:
print(" [WARN] AkShare 新闻获取失败 ({}): {}".format(ts_code, e))
logger.warning("[news_source] AkShare 新闻获取失败 (%s): %s", ts_code, e)
if self.use_xwlb:
try:
@@ -99,7 +102,7 @@ class NewsSource:
if not df.empty:
frames.append(df)
except Exception as e:
print(" [WARN] xwlb 新闻获取失败: {}".format(e))
logger.warning("[news_source] xwlb 新闻获取失败: %s", e)
if self.use_mcp:
try:
@@ -107,7 +110,7 @@ class NewsSource:
if not df.empty:
frames.append(df)
except Exception as e:
print(" [WARN] MCP 新闻获取失败: {}".format(e))
logger.warning("[news_source] MCP 新闻获取失败: %s", e)
if not frames:
return pd.DataFrame(columns=["date", "title", "content", "source", "url"])
@@ -133,7 +136,8 @@ class NewsSource:
symbol = ts_code.replace(".SZ", "").replace(".SH", "").replace(".BJ", "")
try:
df = ak.stock_news_em(symbol=symbol.zfill(6))
except Exception:
except Exception as e:
logger.warning("[news_source] AkShare 新闻接口失败 (%s): %s", ts_code, e)
return pd.DataFrame()
if df is None or df.empty:
@@ -204,7 +208,8 @@ class NewsSource:
df["url"] = ""
return df[["date", "title", "content", "source", "url"]]
except Exception:
except Exception as e:
logger.warning("[news_source] xwlb 新闻获取失败: %s", e)
return pd.DataFrame()
# ── MCP 数据源 ────────────────────────────────────────
@@ -232,7 +237,8 @@ class NewsSource:
df = pd.DataFrame(records)
df["source"] = "mcp_trendradar"
return df[["date", "title", "content", "source", "url"]]
except Exception:
except Exception as e:
logger.warning("[news_source] MCP 新闻获取失败: %s", e)
return pd.DataFrame()
def _mcp_initialize(self) -> str | None:
@@ -267,10 +273,10 @@ class NewsSource:
if session_id:
self._mcp_session_id = session_id
else:
print(" [WARN] MCP initialize 未返回 session-id")
logger.warning("[news_source] MCP initialize 未返回 session-id")
return session_id
except Exception as e:
print(" [WARN] MCP 连接失败: {}".format(e))
logger.warning("[news_source] MCP 连接失败: %s", e)
return None
def _mcp_call_tool(
@@ -299,7 +305,8 @@ class NewsSource:
data = json.loads(line[5:].strip())
return data.get("result", {})
return None
except Exception:
except Exception as e:
logger.warning("[news_source] MCP call_tool 失败: %s", e)
return None
@staticmethod
finance/factors/sentiment/sentiment_engine.py
+2 -1
View File
@@ -128,11 +128,12 @@ class SentimentEngine:
# 5. Qwen 情绪分析（有 API key 时才执行）
sentiment_df = self._analyze_news(news_df)
# 6. 计算因子
# 6. 计算因子（对齐注册表，产出全部 4 个注册情绪因子，含 news_sent_20）
factor_dfs = {}
if not sentiment_df.empty:
for factor_cls, kwargs in [
(NewsSentimentFactor, {"window": 5, "sentiment_df": sentiment_df}),
(NewsSentimentFactor, {"window": 20, "sentiment_df": sentiment_df}),
(SentimentConfidenceFactor, {"window": 5, "sentiment_df": sentiment_df}),
(SentimentMomentumFactor, {"period": 5, "sentiment_df": sentiment_df}),
]:
finance/factors/technical/rsi.py
+8 -2
View File
@@ -2,6 +2,7 @@
RSI 相对强弱因子。
"""
import numpy as np
import pandas as pd
from factors.base import BaseFactor
@@ -22,8 +23,13 @@ class RSIFactor(BaseFactor):
loss = (-delta).clip(lower=0)
avg_gain = gain.ewm(span=self.period, min_periods=self.period).mean()
avg_loss = loss.ewm(span=self.period, min_periods=self.period).mean()
rs = avg_gain / avg_loss.replace(0, float("nan"))
return 100 - 100 / (1 + rs)
# 标准 Wilder RSI：avg_loss==0 时 RSI 应 = 100，而非 NaN。
# 用 where 显式处理除零，避免 replace(0, nan) 把上涨趋势判为缺失。
rs = avg_gain / avg_loss.where(avg_loss != 0, np.nan)
rsi = 100 - 100 / (1 + rs)
# 上涨且无下跌的高位情形补 100（无 prior-loss 的窗口仍留 NaN 由上游填充）
rsi = rsi.where(avg_loss != 0, 100.0)
return rsi
def get_required_columns(self) -> list[str]:
return ["close"]