G2+R3: 赛道映射模块 + akg_gate/akg_score 合成因子（07-30 三拍落地）
两锚三档 + 两段式；C 闸默认关待 yml 转正；requirements 加 PyYAML 需 rebuild。
This commit is contained in:
zlt
parent
ce87a1ea7d
commit
b813e42f52
|
|
@ -107,3 +107,10 @@ FROZEN_ROOT = os.environ.get("FROZEN_ROOT", "/app/data/frozen")
|
|||
# 07-28 语义更新：基座二批后 topic_context cap=1000、scan 大主题闸=600、quiet 全量落库。
|
||||
# 本值对齐基座"大主题闸"：members_total 超过它 = 基座制度回退，桥侧告警（factors.py）。
|
||||
UPSTREAM_MEMBER_CAP = int(os.environ.get("UPSTREAM_MEMBER_CAP", "600"))
|
||||
|
||||
# --- 赛道门槛 C（G2：config/frontier_tracks.yml + tracks.py）-----------------
|
||||
# yml 是唯一事实源（设计 §3.2）。C 闸默认关：映射还是 v0.1 草案，先跑
|
||||
# `python run.py tracks` 做覆盖体检、清单转正后再置 1——届时主榜(gate=2)
|
||||
# 再交赛道成员，观察档的产业链锚也并入赛道成员。
|
||||
TRACKS_YML = os.environ.get("TRACKS_YML", "config/frontier_tracks.yml")
|
||||
ENABLE_TRACK_GATE = os.environ.get("ENABLE_TRACK_GATE", "0") == "1"
|
||||
|
|
|
|||
|
|
@ -229,3 +229,35 @@ probe 三张读数提醒已钉 **07-30 周四 19:00**。
|
|||
快照完整性建议纳入周检/巡检（非二批）。
|
||||
④ 今晚 17:30 盯行数：≈2391=健康且池源同步；≈1675=sync 用滞后池源（新问题）；
|
||||
再挂=马上查日志。07-24 缺口的 sync 日志仍待查。
|
||||
|
||||
## 七、07-30 拍板与交付记录
|
||||
|
||||
1. **快照 T−1 结构性错位定案**：源侧每日 ~19:50 才发布当日行情，两拍 17:30/17:55
|
||||
永远只拿到前一日 → 传导天天用昨日 movers（07-28 扫描用 07-27 快照、07-29 用
|
||||
07-28，台账 mkt_trade_date 如实打标）。07-27 事故是这个常态的极端版；"缺日"是
|
||||
滞后不是丢失（某日行情在次日傍晚自动补上）。**拍点四案待拍**：A 整链后移 ~20:30 /
|
||||
B 加第三拍+当晚重扫 / C 正式接受 T−1 语义 / D 整链挪到次日盘前。倾向 D（盘前出
|
||||
选股计划，晚间披露的公告也能进当日事件）或 A；07-30 起记录源侧到点，与三批-5
|
||||
采信规则同题拍板。
|
||||
2. **9-4 主决策已拍：两锚三档**。业绩锚=券商覆盖（upside 可算）；产业链锚=图谱
|
||||
传导链（赛道清单转正后并入赛道成员）。主榜=有覆盖且 upside≥0（C 闸待转正）；
|
||||
观察档=无覆盖但有产业链锚，算分用传导+热度（暂拟 0.6z_T+0.4z_H），明确标低置信；
|
||||
两锚皆无=不采纳。升降档本身是信号（首次覆盖/新进传导链），进日报。
|
||||
3. **三批-3 已拍：主榜组内排序=两段式**。传导票按 log 传导分的组内中位数分强弱两档，
|
||||
档内按"还没热、还便宜"（0.6z(−热度)+0.4z(upside)）排。依据（07-29 截面）：传导
|
||||
权重 0.5 时 top50 有 48 只传导票=现行加权事实上就是传导优先；两种排法 top20 仅
|
||||
重合 14/20=组内排序就是榜单本身。三项相关 |ρ|≤0.063 正交复现；9-3 的 q 暂留 0
|
||||
（门槛②对有覆盖股只剔 33/994，S2 再议）。
|
||||
4. **probe 修复判收**：corr 节 scipy 崩点（pandas 的 Series.corr(spearman) 会
|
||||
import scipy，DataFrame.corr 不会）已改为复用整表矩阵，07-30 实机全节跑通。
|
||||
5. **07-30 交付（待实机验证）**：
|
||||
① 基座 transmission.py 传导主题黑名单（高新技术企业/小型微利企业/联营企业/无/
|
||||
报告分部——联营企业 298 名成员不触 600 闸，必须显式拦；源与目标两侧都挡）；
|
||||
② 桥 config/frontier_tracks.yml v0.1（6 confirmed + 4 candidate、61 主题映射、
|
||||
排除清单、TODO 决策点待圈）；③ 桥 tracks.py + `run.py tracks`（赛道覆盖体检 +
|
||||
成员表版本化快照落 data/，即本清单"赛道覆盖体检"的正式工具）；④ 桥 akg_gate /
|
||||
akg_score 两个新因子（三档+两段式，进 build all 日更；C 闸默认关
|
||||
=ENABLE_TRACK_GATE，yml 转正后开）；⑤ requirements.txt 加 PyYAML（**桥机需
|
||||
rebuild 镜像**；不 rebuild 只影响 tracks 命令，其余照跑）。
|
||||
**工作哲学（07-30 用户令）**：数据渐进式暴露是常态——不等全量抽完、不等干净
|
||||
截面、不等覆盖完备；缺锚降档不弃用，滞后交采信规则消化。
|
||||
141
factors.py
141
factors.py
|
|
@ -20,6 +20,9 @@ FACTORS = {
|
|||
"akg_heat": "t_factor_akg_heat",
|
||||
"akg_event": "t_factor_akg_event",
|
||||
"akg_transmission": "t_factor_akg_transmission",
|
||||
# 合成层（07-30 拍板：两锚三档 + 两段式），build all 一并日更
|
||||
"akg_gate": "t_factor_akg_gate",
|
||||
"akg_score": "t_factor_akg_score",
|
||||
}
|
||||
|
||||
# ---- 事件极性草案（设计 §5.3，待 §9-5 确认）----
|
||||
|
|
@ -304,7 +307,145 @@ def _warn_upstream_truncation(tr: pd.DataFrame) -> None:
|
|||
f"若构建窗口含 07-24 及更早的旧制度日，此告警对旧日属预期）")
|
||||
|
||||
|
||||
# ---------------------------------------------------------------- 合成层
|
||||
# 07-30 拍板：门槛=两锚三档（akg_gate），主榜组内排序=两段式（akg_score）。
|
||||
# 依据（07-29 截面权重体检）：传导权重 0.5 时 top50 有 48 只传导票——现行加权
|
||||
# 事实上就是传导优先；两种排法 top20 仅重合 14/20，组内排序即榜单本身。
|
||||
|
||||
def _robust_z(v: pd.Series) -> pd.Series:
|
||||
med = v.median()
|
||||
mad = (v - med).abs().median()
|
||||
if pd.isna(mad) or mad == 0:
|
||||
sd = v.std()
|
||||
return (v - v.mean()) / sd if sd and sd > 0 else v * 0.0
|
||||
return (v - med) / (1.4826 * mad)
|
||||
|
||||
|
||||
def _heat_day_for(ds: str):
|
||||
"""热度 T+1 到达 → 取 <= ds 的最新热度日（与 probe / 设计 §3.4 末口径一致）。"""
|
||||
df = db.read_mysql("heat", "SELECT MAX(trade_date) d FROM stock_fund_heat_scores "
|
||||
"WHERE trade_date <= %s", (ds,))
|
||||
v = None if df.empty else df.iloc[0, 0]
|
||||
return None if v is None or pd.isna(v) else pd.Timestamp(v).date().isoformat()
|
||||
|
||||
|
||||
def _col(df: pd.DataFrame, name: str) -> pd.Series:
|
||||
"""子因子 builder 输出 → 前缀码索引的一列（同股同日取最大）。"""
|
||||
if df is None or df.empty:
|
||||
return pd.Series(dtype=float, name=name)
|
||||
x = df.copy()
|
||||
x["k"] = x["stock_code"].map(common.to_prefix)
|
||||
return x.groupby("k")["factor_value"].max().astype(float).rename(name)
|
||||
|
||||
|
||||
def _panel(ds: str) -> pd.DataFrame:
|
||||
"""单日截面：upside / heat / transmission 三列，索引=前缀码。"""
|
||||
up = _col(build_upside(ds, ds), "upside")
|
||||
hd = _heat_day_for(ds)
|
||||
ht = _col(build_heat(hd, hd), "heat") if hd else pd.Series(dtype=float, name="heat")
|
||||
tr = _col(build_transmission(ds, ds), "transmission")
|
||||
return pd.concat([up, ht, tr], axis=1)
|
||||
|
||||
|
||||
def _track_set():
|
||||
"""已转正赛道的成员集合（前缀码）。C 闸未启用时返回 None。"""
|
||||
if not config.ENABLE_TRACK_GATE:
|
||||
return None
|
||||
try:
|
||||
import tracks
|
||||
df, _missing = tracks.resolve_members(only_confirmed=True)
|
||||
return {common.to_prefix(x) for x in df["ts_code"]}
|
||||
except Exception as e: # noqa: BLE001 —— 解析失败按闸未启用降级，只告警不中止
|
||||
print(f" ⚠️ 赛道成员表解析失败，C 闸按未启用处理: {e!r}")
|
||||
return None
|
||||
|
||||
|
||||
def _gate_of(p: pd.DataFrame, track_set) -> pd.Series:
|
||||
"""三档（07-30 拍板）：
|
||||
2 = 主榜：有业绩锚（券商覆盖且 upside>=0）；C 闸开启后再交赛道成员；
|
||||
1 = 观察档：无业绩锚但有产业链锚（当前=在传导链上；C 转正后并入赛道成员）；
|
||||
0 = 不采纳：两锚皆无，或有覆盖但 upside<0（贵了不买是绝对下限）。"""
|
||||
covered = p["upside"].notna()
|
||||
chain = p["transmission"].fillna(0.0) > 0
|
||||
if track_set:
|
||||
in_track = pd.Series(p.index.isin(track_set), index=p.index)
|
||||
chain = chain | in_track
|
||||
g = pd.Series(0.0, index=p.index)
|
||||
main = covered & (p["upside"] >= 0.0)
|
||||
if track_set:
|
||||
main = main & in_track
|
||||
g[main] = 2.0
|
||||
g[(~covered) & chain] = 1.0
|
||||
return g
|
||||
|
||||
|
||||
def _days_of(start, end):
|
||||
days = [pd.Timestamp(d).date().isoformat()
|
||||
for d in common.trading_days(start, end)]
|
||||
if not days and start == end:
|
||||
days = [start] # 日历查不到也照算单日：以调用方给的日期为准
|
||||
return days
|
||||
|
||||
|
||||
def build_gate(start, end):
|
||||
"""akg_gate ∈ {0,1,2}：全池出行——让平台看得见门槛本身，不只是池内排序。
|
||||
history 模式逐日重算三路子因子，区间大会慢；建议 daily / 短区间。"""
|
||||
ts = _track_set()
|
||||
rows = []
|
||||
for ds in _days_of(start, end):
|
||||
p = _panel(ds)
|
||||
if p.empty:
|
||||
continue
|
||||
g = _gate_of(p, ts)
|
||||
rows.append(pd.DataFrame({"trade_date": ds, "stock_code": g.index,
|
||||
"factor_value": g.values}))
|
||||
return pd.concat(rows, ignore_index=True) if rows else _EMPTY
|
||||
|
||||
|
||||
def build_score(start, end):
|
||||
"""akg_score：先档后分，仅 gate>0 出行，数值直接可排序。
|
||||
|
||||
主榜 = 200 + 传导档位×10 + 组内分：传导票按 log 传导分的组内中位数分成
|
||||
强(2)/弱(1)两档，无传导 0 档；组内分 = 0.6·z(−热度) + 0.4·z(upside)，
|
||||
即"还没热、还便宜"。观察档 = 100 + 0.6·z(传导) + 0.4·z(−热度)——无估值锚，
|
||||
低置信。组内分夹在 ±9.9 保证档位永不重叠；z 都在各自档内算。
|
||||
缺失处置同设计 §3.4：传导缺=0，热度缺=档内中位数。"""
|
||||
ts = _track_set()
|
||||
out = []
|
||||
for ds in _days_of(start, end):
|
||||
p = _panel(ds)
|
||||
if p.empty:
|
||||
continue
|
||||
g = _gate_of(p, ts)
|
||||
|
||||
P = p[g == 2.0].copy() # ---- 主榜
|
||||
if not P.empty:
|
||||
P["transmission"] = P["transmission"].fillna(0.0)
|
||||
P["heat"] = P["heat"].fillna(P["heat"].median())
|
||||
lt = np.log1p(P["transmission"])
|
||||
tier = pd.Series(0.0, index=P.index)
|
||||
hit = P["transmission"] > 0
|
||||
if hit.any():
|
||||
med = lt[hit].median()
|
||||
tier[hit] = np.where(lt[hit] >= med, 2.0, 1.0)
|
||||
inner = 0.6 * (-_robust_z(P["heat"])) + 0.4 * _robust_z(P["upside"])
|
||||
score = 200.0 + tier * 10.0 + inner.clip(-9.9, 9.9)
|
||||
out.append(pd.DataFrame({"trade_date": ds, "stock_code": P.index,
|
||||
"factor_value": score.values}))
|
||||
|
||||
O = p[g == 1.0].copy() # ---- 观察档
|
||||
if not O.empty:
|
||||
O["heat"] = O["heat"].fillna(O["heat"].median())
|
||||
zt = _robust_z(np.log1p(O["transmission"].fillna(0.0)))
|
||||
inner = 0.6 * zt + 0.4 * (-_robust_z(O["heat"]))
|
||||
score = 100.0 + inner.clip(-9.9, 9.9)
|
||||
out.append(pd.DataFrame({"trade_date": ds, "stock_code": O.index,
|
||||
"factor_value": score.values}))
|
||||
return pd.concat(out, ignore_index=True) if out else _EMPTY
|
||||
|
||||
|
||||
BUILDERS = {
|
||||
"akg_upside": build_upside, "akg_heat": build_heat,
|
||||
"akg_event": build_event, "akg_transmission": build_transmission,
|
||||
"akg_gate": build_gate, "akg_score": build_score,
|
||||
}
|
||||
|
|
|
|||
|
|
@ -3,3 +3,4 @@ numpy>=1.24
|
|||
psycopg[binary]>=3.1
|
||||
PyMySQL>=1.1
|
||||
python-dotenv>=1.0
|
||||
PyYAML>=6.0
|
||||
|
|
|
|||
22
run.py
22
run.py
|
|
@ -4,7 +4,8 @@
|
|||
python run.py apply-views [--dry-run] # 把插槽视图 DDL 应用到基座 PG
|
||||
python run.py probe # G1 体检（只读，详见 probe.py）
|
||||
python run.py freeze [--date D] # 输入冻结（G0.5，详见 freeze.py）
|
||||
python run.py register # 注册四子因子到 factor_metadata
|
||||
python run.py tracks # 赛道覆盖体检 + 成员表快照（G2）
|
||||
python run.py register # 注册全部因子到 factor_metadata
|
||||
python run.py build all --mode history --start 2024-01-01 --end 2025-12-31
|
||||
python run.py build akg_heat --mode daily --date 2026-07-24
|
||||
python run.py build akg_event --mode history --start 2025-01-01 --end 2026-07-24
|
||||
|
|
@ -88,11 +89,19 @@ _META = {
|
|||
"akg_heat": ("astock-kg 热度", "生态日频资金热度分(0~1)"),
|
||||
"akg_event": ("astock-kg 事件", "利好利空事件时间衰减加权分(仅公告来源,单文档封顶)"),
|
||||
"akg_transmission": ("astock-kg 传导", "板块传导未动成员传导强度(distinct源数×(1-已动比例))"),
|
||||
"akg_gate": ("astock-kg 门槛档位",
|
||||
"三档置信门槛(07-30拍板): 2=主榜(券商覆盖且upside>=0; 赛道C闸未启用, "
|
||||
"转正后再交赛道成员), 1=观察档(无券商覆盖但在图谱传导链上, 无估值锚, "
|
||||
"低置信), 0=不采纳(两锚皆无, 或upside<0)。全池出行, 让平台看得见门槛"),
|
||||
"akg_score": ("astock-kg 景气度漏斗",
|
||||
"先档后分: 主榜=200+传导档位x10+组内分(0.6z(-热度)+0.4z(upside), "
|
||||
"两段式07-30拍板); 观察档=100+0.6z(传导)+0.4z(-热度)。"
|
||||
"仅gate>0出行, 数值直接可排序"),
|
||||
}
|
||||
|
||||
|
||||
def cmd_register():
|
||||
print("注册四子因子：")
|
||||
print(f"注册因子（共 {len(_META)} 个）：")
|
||||
for code, (name, desc) in _META.items():
|
||||
common.register(code, name, factors.FACTORS[code], ["astock-kg", code.split("_", 1)[1]], desc)
|
||||
|
||||
|
|
@ -130,6 +139,7 @@ def main():
|
|||
sub = ap.add_subparsers(dest="cmd", required=True)
|
||||
sub.add_parser("views")
|
||||
sub.add_parser("register")
|
||||
sub.add_parser("tracks") # 赛道覆盖体检 + confirmed 成员表快照（G2）
|
||||
p = sub.add_parser("probe")
|
||||
p.add_argument("--section", choices=["all", "pools", "price", "upside", "corr"],
|
||||
default="all",
|
||||
|
|
@ -161,6 +171,14 @@ def main():
|
|||
freeze.snapshot(a.date)
|
||||
elif a.cmd == "register":
|
||||
cmd_register()
|
||||
elif a.cmd == "tracks":
|
||||
import tracks
|
||||
tracks.coverage_report()
|
||||
out, df, missing = tracks.snapshot(only_confirmed=True)
|
||||
print(f"\nconfirmed 成员表快照: {out}"
|
||||
f"（{df['ts_code'].nunique()} 只，{len(df)} 行 股票×赛道）")
|
||||
if missing:
|
||||
print(f"⚠️ {len(missing)} 个主题在 industry_pools 里没找到（见体检表）")
|
||||
elif a.cmd == "build":
|
||||
cmd_build(a.factor, a.mode, a.start, a.end, a.date, do_freeze=not a.no_freeze)
|
||||
|
||||
|
|
|
|||
|
|
@ -0,0 +1,106 @@
|
|||
"""赛道映射（硬门槛 C 的实现载体，设计 §3.2）。
|
||||
|
||||
config/frontier_tracks.yml 是唯一事实源：每个赛道列出 kg_themes
|
||||
（industry_pools 主题名）。本模块把它解析成 ts_code 级成员表，并落
|
||||
data/track_members_<日期>.csv 版本化快照——可 git diff、可审计：
|
||||
每只股票能追溯到因哪个赛道、哪个主题入选。
|
||||
|
||||
当前唯一的映射路径是主题名（kg_segments / kg_concepts 要等基座的环节
|
||||
投影表建成，即三批-4），source_rule 统一记 'pool_theme'。
|
||||
"""
|
||||
import datetime as dt
|
||||
import json
|
||||
import os
|
||||
|
||||
import pandas as pd
|
||||
|
||||
import config
|
||||
import db
|
||||
|
||||
try:
|
||||
import yaml
|
||||
except ImportError: # 镜像未装 PyYAML 时给出可执行的修复指令，而不是裸崩
|
||||
yaml = None
|
||||
|
||||
|
||||
def _need_yaml():
|
||||
if yaml is None:
|
||||
raise SystemExit(
|
||||
"缺 PyYAML：requirements.txt 已加，请在桥机重建镜像——\n"
|
||||
" docker compose build akg-factor-bridge && "
|
||||
"docker compose up -d akg-factor-bridge")
|
||||
|
||||
|
||||
def load_yml(path: str | None = None) -> dict:
|
||||
_need_yaml()
|
||||
with open(path or config.TRACKS_YML, "r", encoding="utf-8") as f:
|
||||
return yaml.safe_load(f)
|
||||
|
||||
|
||||
def resolve_members(only_confirmed: bool = True, path: str | None = None):
|
||||
"""yml + industry_pools 当前态 → (成员表, 未命中主题清单)。
|
||||
|
||||
成员表列：ts_code, name, track, status, theme。同股同赛道多主题只留一行，
|
||||
同股跨赛道保留多行。未命中 = yml 里写了、industry_pools 里查无此主题
|
||||
（通常是主题改名或池尚未涌现，体检时重点看）。
|
||||
"""
|
||||
d = load_yml(path)
|
||||
pools = db.read_pg("SELECT theme, members FROM industry_pools")
|
||||
by_theme = {}
|
||||
for _, r in pools.iterrows():
|
||||
ms = r["members"]
|
||||
if isinstance(ms, str):
|
||||
ms = json.loads(ms)
|
||||
by_theme[str(r["theme"]).strip()] = ms or []
|
||||
excl = {str(x).strip() for x in (d.get("exclude_themes") or [])}
|
||||
rows, missing = [], []
|
||||
for tr in d.get("tracks") or []:
|
||||
if only_confirmed and tr.get("status") != "confirmed":
|
||||
continue
|
||||
for theme in tr.get("kg_themes") or []:
|
||||
t = str(theme).strip()
|
||||
if t in excl:
|
||||
continue
|
||||
if t not in by_theme:
|
||||
missing.append((tr["name"], t))
|
||||
continue
|
||||
for m in by_theme[t]:
|
||||
ts = (m or {}).get("ts_code")
|
||||
if ts:
|
||||
rows.append((ts, m.get("name"), tr["name"],
|
||||
tr.get("status"), t))
|
||||
df = (pd.DataFrame(rows, columns=["ts_code", "name", "track",
|
||||
"status", "theme"])
|
||||
.drop_duplicates(["ts_code", "track"]))
|
||||
return df, missing
|
||||
|
||||
|
||||
def snapshot(only_confirmed: bool = True, path: str | None = None):
|
||||
"""成员表落 data/ 版本化快照（含 source_rule / updated_at，审计列）。"""
|
||||
df, missing = resolve_members(only_confirmed, path)
|
||||
os.makedirs("data", exist_ok=True)
|
||||
out = f"data/track_members_{dt.date.today().isoformat()}.csv"
|
||||
(df.assign(layer="", source_rule="pool_theme",
|
||||
updated_at=dt.datetime.now().isoformat(timespec="seconds"))
|
||||
.to_csv(out, index=False))
|
||||
return out, df, missing
|
||||
|
||||
|
||||
def coverage_report(path: str | None = None) -> pd.DataFrame:
|
||||
"""赛道覆盖体检（G1 清单的首件事）：逐赛道成员数与主题命中率，含 candidate。"""
|
||||
df, missing = resolve_members(only_confirmed=False, path=path)
|
||||
d = load_yml(path)
|
||||
print("赛道覆盖体检（industry_pools 当前态；成员数已去重）：")
|
||||
for tr in d.get("tracks") or []:
|
||||
sub = df[df["track"] == tr["name"]]
|
||||
n_theme = len(tr.get("kg_themes") or [])
|
||||
miss = [t for name, t in missing if name == tr["name"]]
|
||||
tag = "" if tr.get("status") == "confirmed" else "（candidate）"
|
||||
line = (f" {tr['name']}{tag}: 成员 {sub['ts_code'].nunique()} 只"
|
||||
f" | 主题命中 {sub['theme'].nunique()}/{n_theme}")
|
||||
if miss:
|
||||
line += f" | 未命中: {'、'.join(miss)}"
|
||||
print(line)
|
||||
conf = df[df["status"] == "confirmed"]
|
||||
print(f" —— confirmed 合计（去重）: {conf['ts_code'].nunique()} 只")
|
||||
return df
|
||||
Loading…
Reference in New Issue