183 lines
8.9 KiB
Python
183 lines
8.9 KiB
Python
"""逻辑状态四态的覆盖率读数：只读，不写库，不改任何判决。
|
||
|
||
回答一个问题：四态这套东西今天在真实数据上能覆盖多少票、各落到哪个状态。
|
||
它是判断"这一路该接多大面"的依据，也是接进计划装配之前必须先看的读数。
|
||
|
||
三路各自的取数在这里，归一与合成都调 logic_state 里的纯函数——读数脚本绝不能自己
|
||
另写一套判据，否则读到的就不是系统真会给出的状态。
|
||
|
||
甲路 研报论断 基座 PG v_factor_logic（经 sources.logic_claims）
|
||
乙路 产业研判 平台 MySQL t_akg_judgement_snapshot（经 judgement.load_previous）
|
||
丙路 券商行动 平台 MySQL gp_report_rc：两个等长窗口的每股收益预测中位数与机构数
|
||
丁路 公司事件 无数据源，恒定缺失
|
||
|
||
丙路的取数口径（三条都要照做，否则读数是错的）：
|
||
一，两个窗口必须等长。不等长会让八成的票假显示覆盖收缩——实测前 135 天对近 45 天时
|
||
有 907 只票误报。
|
||
二，同一财年同一预测期才可比，按 quarter 精确匹配，跨财年比较没有意义。
|
||
三，同一家机构在窗口里可能发多篇，先按机构取最近一篇再算中位数，否则发得勤的
|
||
机构会被重复计入。
|
||
|
||
跑法（155 上）：docker exec akg_factor_bridge python3 logic_state_coverage.py [数据日]
|
||
"""
|
||
from __future__ import annotations
|
||
|
||
import datetime as dt
|
||
import statistics as st
|
||
import sys
|
||
from collections import Counter, defaultdict
|
||
|
||
import common
|
||
import config
|
||
import db
|
||
import judgement
|
||
import logic_state as ls
|
||
import sources
|
||
|
||
# 丙路两个窗口各自的长度（自然日）。等长是硬要求，见模块说明第一条。
|
||
BROKER_WINDOW_DAYS = 45
|
||
|
||
|
||
def broker_paths(codes, ds: str) -> dict:
|
||
"""按票算丙路信号。返回前缀码到 logic_state.signal 的字典（算不出的票不进字典）。"""
|
||
end = dt.date.fromisoformat(ds)
|
||
mid = end - dt.timedelta(days=BROKER_WINDOW_DAYS)
|
||
start = end - dt.timedelta(days=BROKER_WINDOW_DAYS * 2)
|
||
# noqa: SLF001 —— _to_dot 是同仓自用的代码格式转换
|
||
dotted = sorted({sources._to_dot(c) for c in codes if c}) # noqa: SLF001
|
||
if not dotted:
|
||
return {}
|
||
out: dict[str, dict] = {}
|
||
# 一次拉两个窗口的全部行，按票在内存里分窗——逐票查库要发几千次请求。
|
||
marks = ",".join(["%s"] * len(dotted))
|
||
try:
|
||
df = db.read_mysql(
|
||
"factor",
|
||
f"SELECT ts_code, report_date, quarter, org_name, eps FROM gp_report_rc "
|
||
f"WHERE ts_code IN ({marks}) AND report_date > %s AND report_date <= %s "
|
||
f"AND eps IS NOT NULL AND quarter IS NOT NULL",
|
||
tuple(dotted) + (start.isoformat(), end.isoformat()))
|
||
except Exception as e: # noqa: BLE001
|
||
print(f" （券商研报明细表读取失败，丙路整体缺席: {e!r}）")
|
||
return {}
|
||
|
||
# 分票、分窗、分财年地堆起来：{票: {财年: {"now": {机构: (日期, 每股收益)}, "prev": ...}}}
|
||
box: dict = defaultdict(lambda: defaultdict(lambda: {"now": {}, "prev": {}}))
|
||
for r in df.itertuples():
|
||
d = sources._ymd(r.report_date) # noqa: SLF001 —— 同仓自用
|
||
if not d:
|
||
continue
|
||
win = "now" if d > mid.isoformat() else "prev"
|
||
k = common.to_prefix(str(r.ts_code).strip())
|
||
q = str(r.quarter).strip()
|
||
org = str(r.org_name or "").strip() or "未署名"
|
||
slot = box[k][q][win]
|
||
# 同一家机构在窗口里发了多篇，只留最近一篇（模块说明第三条）。
|
||
if org not in slot or d > slot[org][0]:
|
||
slot[org] = (d, float(r.eps))
|
||
|
||
for k, by_q in box.items():
|
||
# 两个窗口都有料的财年里，取行数最多的那个作可比口径（模块说明第二条）。
|
||
usable = [(q, v) for q, v in by_q.items() if v["now"] and v["prev"]]
|
||
if not usable:
|
||
continue
|
||
q, v = max(usable, key=lambda kv: len(kv[1]["now"]) + len(kv[1]["prev"]))
|
||
now = {"eps": st.median([x[1] for x in v["now"].values()]), "firms": len(v["now"])}
|
||
prev = {"eps": st.median([x[1] for x in v["prev"].values()]), "firms": len(v["prev"])}
|
||
s = ls.from_broker(now, prev, as_of=ds)
|
||
s["refs"] = [{**(s["refs"][0] if s["refs"] else {}), "quarter": q}]
|
||
out[k] = s
|
||
return out
|
||
|
||
|
||
def snapshot_of(day: str) -> dict:
|
||
"""行业观点快照表里这个计划日的那一版，按簇键索引；表没建或没有当日行返回空字典。"""
|
||
try:
|
||
df = db.read_mysql(
|
||
"factor",
|
||
f"SELECT * FROM {config.JUDGEMENT_SNAPSHOT_TABLE} WHERE plan_date = %s", (day,))
|
||
except Exception as e: # noqa: BLE001
|
||
print(f" （行业观点快照表读取失败，乙路整体缺席: {e!r}）")
|
||
return {}
|
||
return {str(r["cluster_key"]): dict(r) for _, r in df.iterrows()} if not df.empty else {}
|
||
|
||
|
||
def main(ds: str | None = None) -> None:
|
||
ds = ds or (dt.date.today() - dt.timedelta(days=1)).isoformat()
|
||
print(f"=== 逻辑状态四态覆盖率读数 · 数据日 {ds} ===\n")
|
||
|
||
codes = db.read_pg("SELECT DISTINCT ts_code FROM v_factor_stock_daily "
|
||
"WHERE trade_date = %s", (ds,))["ts_code"].tolist()
|
||
codes = [common.to_prefix(str(c).strip()) for c in codes]
|
||
print(f"当日有行情的票 {len(codes)} 只\n")
|
||
|
||
claims = sources.logic_claims(codes, ds, per_stock=200)
|
||
brokers = broker_paths(codes, ds)
|
||
# 要的是这个计划日当天那一版快照（带迁移与陈旧两列），不是它之前的那一版——
|
||
# load_previous 读的是严格早于计划日的，用它会在快照首日读到空。当天没有就退回上一版。
|
||
snaps = snapshot_of(ds) or judgement.load_previous(ds)
|
||
print(f"甲路取到 {len(claims)} 只票的论断；丙路算得出 {len(brokers)} 只票；"
|
||
f"乙路快照 {len(snaps)} 个簇\n")
|
||
|
||
# 乙路按环节名对上主题：候选卡按环节，产业研判按主题聚簇，两者不在一个命名空间，
|
||
# 这里只做同名匹配，对不上的票乙路就是缺失。这一路的天花板本来就低（实测 1.4%）。
|
||
by_seg = {(r.get("segment_name") or r.get("subject_name") or "").strip(): r
|
||
for r in snaps.values()}
|
||
seg_of = {}
|
||
try:
|
||
d = db.read_pg("SELECT ts_code, target FROM v_factor_transmission_moved "
|
||
"WHERE scan_date = %s", (ds,))
|
||
for r in d.itertuples():
|
||
seg_of[common.to_prefix(str(r.ts_code).strip())] = str(r.target)
|
||
except Exception as e: # noqa: BLE001
|
||
print(f" （传导视图读取失败，乙路整体缺席: {e!r}）")
|
||
|
||
per_path = {ls.PATH_CLAIM: Counter(), ls.PATH_JUDGE: Counter(),
|
||
ls.PATH_BROKER: Counter()}
|
||
states, whys = Counter(), Counter()
|
||
samples: dict = {}
|
||
for k in codes:
|
||
a = ls.from_claims(claims.get(k), ds, stale_days=config.LOGIC_STALE_DAYS)
|
||
b = ls.from_judgement(by_seg.get(seg_of.get(k, "")))
|
||
c = brokers.get(k) or ls.signal(ls.PATH_BROKER, ls.SIG_NONE,
|
||
why="两个窗口里算不出可比的预测")
|
||
per_path[ls.PATH_CLAIM][a["signal"]] += 1
|
||
per_path[ls.PATH_JUDGE][b["signal"]] += 1
|
||
per_path[ls.PATH_BROKER][c["signal"]] += 1
|
||
r = ls.compose([a, b, c, ls.from_events()])
|
||
states[r["state"]] += 1
|
||
if r["why"]:
|
||
whys[f"{r['state']}·{r['why']}"] += 1
|
||
# 逻辑存疑与逻辑强化都是稀有且会触发动作的状态，全部列出来逐条看得见；
|
||
# 无法判断那两类只留一条样例，否则几千行没法读。
|
||
if r["state"] in (ls.STATE_STRONG, ls.STATE_DOUBT):
|
||
samples.setdefault(r["state"], []).append((k, r["reasons"][:3]))
|
||
elif r["state"] == ls.STATE_UNKNOWN and r["why"] == ls.WHY_CONFLICT:
|
||
samples.setdefault("无法判断·证据矛盾", []).append((k, r["reasons"][:3]))
|
||
|
||
print("每路各自的信号分布")
|
||
for path, cnt in per_path.items():
|
||
tot = sum(cnt.values())
|
||
line = "、".join(f"{s} {n}（{n / tot:.1%}）" for s, n in cnt.most_common())
|
||
print(f" {path}：{line}")
|
||
print(f" {ls.PATH_EVENT}：缺失 {len(codes)}（100.0%，无数据源）\n")
|
||
|
||
print("合成后的四态分布")
|
||
for s, n in states.most_common():
|
||
print(f" {s} {n}（{n / len(codes):.1%}）")
|
||
if whys:
|
||
print("\n 无法判断的子因")
|
||
for w, n in whys.most_common():
|
||
print(f" {w} {n}")
|
||
for tag, rows in samples.items():
|
||
cap = len(rows) if tag in (ls.STATE_STRONG, ls.STATE_DOUBT) else 2
|
||
print(f"\n{tag}（{len(rows)} 只{'，全部列出' if cap == len(rows) else '，列前两只'}）")
|
||
for k, rs in rows[:cap]:
|
||
print(f" {k}")
|
||
for x in rs:
|
||
print(f" {x}")
|
||
|
||
|
||
if __name__ == "__main__":
|
||
main(sys.argv[1] if len(sys.argv) > 1 else None)
|