选股计划排序轴换成公司质地,传导降为雷达门与带内次级键(台账 063)
按 docs/选股计划调整方案_2026-09-17.md 落地。三份读数(09-02 §1.8、08-28 周报、09-11 抽查)一致:传导越强超额越负,传导分的公式奖励的正是被证明为负的"未动补涨"。所以把传导从排序主轴换成雷达门加带内次级键,公司质地成为排序主轴。改动全在计划装配层,四个开关全设旧值即逐字回旧。 一、排序轴换质地(plan.py)。新增 _rerank_by_axis 按(质地带、被指向、组内分、原始分、代码)词典序重排、_rerank 按 config.PLAN_RANK_AXIS 分派;组内分从原始分反解、档位不进次序;被指向以候选卡门槛一 gates.pointed 为准。调用点先用 _rerank_by_quality 算旧序拿 rank_old 再算新序。_pick 与 pick_rows 加 uncapped_no_theme:质地轴下无传导行免主题限额、不计数,否则主轴被限额推翻。_row 加 quality_band、rank_axis、pointed、rank_old 四键,score/score_adj/tier 值与含义不变。返回字典质地轴下加 rank_axis、rank_rule、quality_band_counts。encoding 追加一句。 二、候选卡内序与质地门(card.py)。新增 quality_band 函数(好三中二差零,无评析与证据不足与超期一百二十天归一),plan.py 与 card.py 共用它。judge 加 stale_days 参数、返回加 quality_band;质地门"差降关注"改用质地带为零并带上一百二十天守卫(超期报告不再降档)。sort_key 在判决与涨幅之间插入质地带、受 CARD_SORT_BY_QUALITY 控,关掉逐字回旧。三门槛一确认三风险一字未动。 三、入池与评析名单(pool.py、review_targets.py)。pool 把选行抽成纯函数 select_plan_rows:rank 源取主榜新序里被指向的前 POOL_TOP、_pool_source 记 rank_top,备用开关 POOL_RANK_KEEP_OLD 并进旧序强传导前若干、去重、受 POOL_MAX 兜底;tier 与 candidate 逐字保留。strategy_context_entry 加 quality_band、rank_axis、rank_old 三键。review_targets 新增纯函数 pick_main_targets:both 取原始分前一百与新序前五十的并集、去重按新序,score 逐字回旧;评析覆盖与排序轴解耦。 四、周报对照(plan_review.py)。第一节加主榜新序前二十与旧序前二十两份名单(rank_old 缺的旧快照跳过);档位分组外加档位乘质地带交叉分组,看传导在质地档内还剩多少信息量。plan_reconcile 应答加 rank_axis: transmission 一键,说明逐票对账的"主榜第 N 名"仍按原始分序,文案不动。 五、单测、开关、文档。新 test_plan_rank_axis.py 十例(含四开关设旧值逐字回旧、四带序、带内次级键、确定性、超期守卫、入池 rank、评析并集、卡内序、主题限额例外、备用开关),本地 ALL OK。test_plan_quality_rank.py 接线匹配串改成新调用点,test_plan_snapshot.py 加一条质地轴主题限额对照用例。config 加 PLAN_RANK_AXIS、CARD_SORT_BY_QUALITY、REVIEW_TARGETS_AXIS 三个开关与备用开关 POOL_RANK_KEEP_OLD,POOL_SOURCE 默认改 rank。.env.example、README 计划字段、台账 063、入池对接说明、下游对接说明一并改。 没有改的是什么、为什么不合并。分数编码 factors.py 一字未动——它注册在外部平台当因子,改它会污染平台的因子历史;次序改动全部放在装配层。api.py 不碰。三门槛一确认三风险不动,质地门保持"差降关注"。持仓管理系统按上游名次排是协同项,在另一仓库单独审批,不阻塞本次上线(它的看多分桶已把好公司排在前面)。 实测读数。开发机 python3 test_plan_rank_axis.py 与 python3 test_plan_quality_rank.py 均 ALL OK;PLAN_RANK_AXIS=transmission CARD_SORT_BY_QUALITY=0 POOL_SOURCE=tier REVIEW_TARGETS_AXIS=score 下 test_plan_rank_axis.py 仍 ALL OK(逐字回旧那一例断言与 09-14 旧序逐字相同)。test_card、test_pool_logic、test_plan_snapshot 全绿。既有的 test_review_targets、test_plan_review_window、test_plan_logic_state、test_plan_cache、test_plan_reviews_live 因缺 psycopg/fastapi 或本机 Python 3.9 不支持联合类型注解而只能在部署机跑,非本次引入。 Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
parent
6138c01482
commit
4d7ca557b2
10
.env.example
10
.env.example
|
|
@ -91,9 +91,15 @@ MONGO_DB=stock_predictions
|
||||||
# PMS_MYSQL_DB=
|
# PMS_MYSQL_DB=
|
||||||
|
|
||||||
# --- 质地调整分(2026-09-14 工作包四;台账 058):计划装配时把公司质地并进组内名次,
|
# --- 质地调整分(2026-09-14 工作包四;台账 058):计划装配时把公司质地并进组内名次,
|
||||||
# 好的票同档内前移、差的后移,不跨档。默认 8,设 0 即关(关掉后计划输出逐字回旧)---
|
# 好的票同档内前移、差的后移,不跨档。默认 8,设 0 即关。2026-09-17 起只在 transmission 轴下起作用 ---
|
||||||
# PLAN_QUALITY_BONUS=8
|
# PLAN_QUALITY_BONUS=8
|
||||||
|
|
||||||
|
# --- 排序主轴换成公司质地(2026-09-17 方案;台账 063)。机器上不填任何键就是新行为;
|
||||||
|
# 回退时把这三个加下面的 POOL_SOURCE 一起设回旧值即逐字回旧,不用回代码、不用重启 ---
|
||||||
|
# PLAN_RANK_AXIS=quality # 默认 quality(质地带为主键,传导降为带内次级键);transmission=回 09-14 旧轴
|
||||||
|
# CARD_SORT_BY_QUALITY=1 # 候选卡内序先质地带;0=回"先判决再涨幅"
|
||||||
|
# REVIEW_TARGETS_AXIS=both # 评析名单=原始分前100 ∪ 新序前50;score=只按原始分前100(旧)
|
||||||
|
|
||||||
# --- 入池口径(默认与 PMS 候选一致:强传导主榜前 20;改动记得与 PMS 页面的
|
# --- 入池口径(默认与 PMS 候选一致:强传导主榜前 20;改动记得与 PMS 页面的
|
||||||
# PMS_PLAN_TOP_N / PMS_PLAN_TIERS 保持一致)---
|
# PMS_PLAN_TOP_N / PMS_PLAN_TIERS 保持一致)---
|
||||||
# POOL_GROUP_CODE=AKG_PLAN
|
# POOL_GROUP_CODE=AKG_PLAN
|
||||||
|
|
@ -103,6 +109,8 @@ MONGO_DB=stock_predictions
|
||||||
# POOL_TIERS=强传导
|
# POOL_TIERS=强传导
|
||||||
# POOL_THEME_CAP=5
|
# POOL_THEME_CAP=5
|
||||||
# POOL_MAX=60
|
# POOL_MAX=60
|
||||||
|
# POOL_SOURCE=rank # 默认 rank(主榜新序里被指向的前 POOL_TOP,与 PMS 候选同序);tier=强传导前 POOL_TOP(旧);candidate=候选优先
|
||||||
|
# POOL_RANK_KEEP_OLD=0 # 备用:rank 源下把旧序强传导前 POOL_TOP 也并进池,只在"池含 PMS 候选"不变式核不过时开
|
||||||
|
|
||||||
# --- 写完池子后触发决策系统的增量补扫(只补当天没分析过的票;留空=不触发,
|
# --- 写完池子后触发决策系统的增量补扫(只补当天没分析过的票;留空=不触发,
|
||||||
# 当晚 22:30 全量扫兜底)。key 填决策系统 .env 的 XXL_TRIGGER_KEY ---
|
# 当晚 22:30 全量扫兜底)。key 填决策系统 .env 的 XXL_TRIGGER_KEY ---
|
||||||
|
|
|
||||||
|
|
@ -189,7 +189,7 @@ docker compose exec akg-factor-bridge python run.py freeze --date 2026-07-24
|
||||||
|---|---|
|
|---|---|
|
||||||
| `GET /health` | 存活探针;2026-09-03 起带 `pool_top`(POOL_TOP)与 `pool_max`(POOL_MAX),供 PMS 池深探针比对 |
|
| `GET /health` | 存活探针;2026-09-03 起带 `pool_top`(POOL_TOP)与 `pool_max`(POOL_MAX),供 PMS 池深探针比对 |
|
||||||
| `GET /plan/dates?limit=30` | 有计划文件的日期清单 |
|
| `GET /plan/dates?limit=30` | 有计划文件的日期清单 |
|
||||||
| `GET /plan?date=&format=json\|md&top=&obs_top=` | 当日选股计划(基座今日页 `/plan/today` 与 PMS 下单候选都读它)。顶层 `regime`(区制)与 `market`(两市成交额、广度、融资、恐贪四项)都只从当日快照读,快照缺失为空;每行的判决类字段 2026-09-03 起多 `basis`(判决依据)与 `logic`(因果论断带出处) |
|
| `GET /plan?date=&format=json\|md&top=&obs_top=` | 当日选股计划(基座今日页 `/plan/today` 与 PMS 下单候选都读它)。顶层 `regime`(区制)与 `market`(两市成交额、广度、融资、恐贪四项)都只从当日快照读,快照缺失为空;每行的判决类字段 2026-09-03 起多 `basis`(判决依据)与 `logic`(因果论断带出处);2026-09-17 起名次改按质地带优先,每行多 `quality_band`(质地带:好 3 中 2 无评析 1 差 0)、`rank_axis`(quality/transmission)、`pointed`(是否被传导指向)、`rank_old`(旧序名次,供周报四周对照),开关 `PLAN_RANK_AXIS` 设 transmission 即逐字回旧 |
|
||||||
| `POST /plan/refresh?date=` | 重生成当日计划文件 |
|
| `POST /plan/refresh?date=` | 重生成当日计划文件 |
|
||||||
| `GET /plan/verdict?codes=300750,SH600438&date=` | 逐票『计划判决』(2026-08-18 PMS 对接加):decision(main/observe/reject/absent)+ 与命令行 `plan_reconcile` 逐字同口径的 verdict_text + score/rank/tier/赛道/图谱证据。单票数据异常或代码形态认不出只报该票 error,不崩整批 |
|
| `GET /plan/verdict?codes=300750,SH600438&date=` | 逐票『计划判决』(2026-08-18 PMS 对接加):decision(main/observe/reject/absent)+ 与命令行 `plan_reconcile` 逐字同口径的 verdict_text + score/rank/tier/赛道/图谱证据。单票数据异常或代码形态认不出只报该票 error,不崩整批 |
|
||||||
| `GET\|POST /api/v1/xxl/daily-build?key=&steps=&date=` | XXL-JOB 平台触发盘前链(build→plan→push-pool,白名单步骤、单实例互斥、回调结案);`.env` 不配 `XXL_TRIGGER_KEY` 则整组禁用 |
|
| `GET\|POST /api/v1/xxl/daily-build?key=&steps=&date=` | XXL-JOB 平台触发盘前链(build→plan→push-pool,白名单步骤、单实例互斥、回调结案);`.env` 不配 `XXL_TRIGGER_KEY` 则整组禁用 |
|
||||||
|
|
|
||||||
40
card.py
40
card.py
|
|
@ -71,10 +71,31 @@ def _num(v) -> float | None:
|
||||||
return x
|
return x
|
||||||
|
|
||||||
|
|
||||||
|
def quality_band(cr, stale_days: int) -> int:
|
||||||
|
"""公司质地评析的排序带(2026-09-17):好三、中二、差零;无评析、证据不足、报告超期都归一。
|
||||||
|
无评析放在中之下、差之上——不知道不等于差,但不该排在已核实的好公司前面。报告日龄超过
|
||||||
|
stale_days 的一律降到第一带(与逻辑状态第五路、持仓管理系统的基本面立场同一条新鲜度守卫)。
|
||||||
|
plan.py 与 card.py 共用这一个函数,是质地带、质地门、卡内序三处的唯一判定口径。"""
|
||||||
|
if not isinstance(cr, dict):
|
||||||
|
return 1
|
||||||
|
age = cr.get("age_days")
|
||||||
|
if isinstance(age, (int, float)) and age > stale_days:
|
||||||
|
return 1 # 报告超期:无论好中差都归第一带,不再据它升降或降档
|
||||||
|
o = cr.get("overall")
|
||||||
|
if o == "好":
|
||||||
|
return 3
|
||||||
|
if o == "中":
|
||||||
|
return 2
|
||||||
|
if o == "差":
|
||||||
|
return 0
|
||||||
|
return 1 # 无评析、证据不足
|
||||||
|
|
||||||
|
|
||||||
def judge(ev: dict[str, Any], *, start_pct: float = 3.0,
|
def judge(ev: dict[str, Any], *, start_pct: float = 3.0,
|
||||||
accum_max_age: int = 30, neg_tol: float = 0.0,
|
accum_max_age: int = 30, neg_tol: float = 0.0,
|
||||||
logic_stale_days: int = 90, require_started: bool = True,
|
logic_stale_days: int = 90, require_started: bool = True,
|
||||||
company_review: dict | None = None, quality_gate: bool = True) -> dict[str, Any]:
|
company_review: dict | None = None, quality_gate: bool = True,
|
||||||
|
stale_days: int = 120) -> dict[str, Any]:
|
||||||
"""一只票的证据字典 -> 判决字典。
|
"""一只票的证据字典 -> 判决字典。
|
||||||
|
|
||||||
ev 里认的键(缺键按缺失处理,不报错):
|
ev 里认的键(缺键按缺失处理,不报错):
|
||||||
|
|
@ -203,7 +224,10 @@ def judge(ev: dict[str, Any], *, start_pct: float = 3.0,
|
||||||
# 质地好不升档;证据不足、没报告不动。三门槛一确认三风险本身不动。
|
# 质地好不升档;证据不足、没报告不动。三门槛一确认三风险本身不动。
|
||||||
company_gate = ""
|
company_gate = ""
|
||||||
cr = company_review if isinstance(company_review, dict) else None
|
cr = company_review if isinstance(company_review, dict) else None
|
||||||
if quality_gate and cr and cr.get("overall") == "差" and verdict == VERDICT_CANDIDATE:
|
qband = quality_band(company_review, stale_days)
|
||||||
|
# 质地门"差降关注"(2026-09-17 与排序轴共用 quality_band,带一百二十天守卫):判决收敛只降不升——
|
||||||
|
# 质地带为零(差且报告未超期)且判决为候选,降为关注交人裁决;超期报告的带被守卫抬到一,不再降档。
|
||||||
|
if quality_gate and cr and qband == 0 and verdict == VERDICT_CANDIDATE:
|
||||||
verdict = VERDICT_WATCH
|
verdict = VERDICT_WATCH
|
||||||
why_q = ";".join((cr.get("quality_reasons") or [])[:2]) or "基座代码合成的质地档为差"
|
why_q = ";".join((cr.get("quality_reasons") or [])[:2]) or "基座代码合成的质地档为差"
|
||||||
company_gate = f"公司质地评为差({why_q}),候选降为关注交人裁决"
|
company_gate = f"公司质地评为差({why_q}),候选降为关注交人裁决"
|
||||||
|
|
@ -211,7 +235,7 @@ def judge(ev: dict[str, Any], *, start_pct: float = 3.0,
|
||||||
|
|
||||||
return {"verdict": verdict, "reasons": reasons, "missing": missing, "risk": risk,
|
return {"verdict": verdict, "reasons": reasons, "missing": missing, "risk": risk,
|
||||||
"gates": gates, "confirm": confirm, "failed_gates": failed, "basis": basis,
|
"gates": gates, "confirm": confirm, "failed_gates": failed, "basis": basis,
|
||||||
"company_gate": company_gate,
|
"company_gate": company_gate, "quality_band": qband,
|
||||||
"logic": logic_lines(ev.get("logic")),
|
"logic": logic_lines(ev.get("logic")),
|
||||||
"logic_quality": logic_quality(ev.get("logic"))}
|
"logic_quality": logic_quality(ev.get("logic"))}
|
||||||
|
|
||||||
|
|
@ -245,10 +269,16 @@ def logic_lines(claims, limit: int = 3) -> list[str]:
|
||||||
|
|
||||||
|
|
||||||
def sort_key(row: dict[str, Any]) -> tuple:
|
def sort_key(row: dict[str, Any]) -> tuple:
|
||||||
"""卡内序:先判决(候选 < 关注 < 仅展示),再按数据日涨幅降序。不算总分。"""
|
"""卡内序:先判决(候选 < 关注 < 仅展示),2026-09-17 起在判决与涨幅之间插入质地带降序
|
||||||
|
(开关 CARD_SORT_BY_QUALITY,默认开),最后按数据日涨幅降序。开关关掉时元组与旧实现
|
||||||
|
(先判决再涨幅)逐字相同,逐字回旧。不算总分。"""
|
||||||
|
import config
|
||||||
v = _VERDICT_ORDER.get(row.get("verdict"), 9)
|
v = _VERDICT_ORDER.get(row.get("verdict"), 9)
|
||||||
pct0 = _num(row.get("pct0"))
|
pct0 = _num(row.get("pct0"))
|
||||||
return (v, -(pct0 if pct0 is not None else -1e9))
|
pct_key = -(pct0 if pct0 is not None else -1e9)
|
||||||
|
if config.CARD_SORT_BY_QUALITY:
|
||||||
|
return (v, -int(row.get("quality_band") or 0), pct_key)
|
||||||
|
return (v, pct_key)
|
||||||
|
|
||||||
|
|
||||||
def logic_quality(claims) -> dict | None:
|
def logic_quality(claims) -> dict | None:
|
||||||
|
|
|
||||||
28
config.py
28
config.py
|
|
@ -164,8 +164,19 @@ UPSIDE_NEG_TOLERANCE = float(os.environ.get("UPSIDE_NEG_TOLERANCE", "0"))
|
||||||
# 计划装配时把公司深度评析的质地档并进组内分:质地好加这个分、差减这个分、其余不动。
|
# 计划装配时把公司深度评析的质地档并进组内分:质地好加这个分、差减这个分、其余不动。
|
||||||
# 只影响组内名次,不跨档:主榜档位步长是 20,最大加成 8 远小于 20,好的票在同档内前移、
|
# 只影响组内名次,不跨档:主榜档位步长是 20,最大加成 8 远小于 20,好的票在同档内前移、
|
||||||
# 差的票后移,档位标签仍按原始分解码。默认 8,设 0 即关(关掉后计划输出与改动前逐字一致)。
|
# 差的票后移,档位标签仍按原始分解码。默认 8,设 0 即关(关掉后计划输出与改动前逐字一致)。
|
||||||
|
# 2026-09-17 起只在 PLAN_RANK_AXIS=transmission(旧轴)下起作用;质地轴下质地带才是主键、不再用加成。
|
||||||
PLAN_QUALITY_BONUS = float(os.environ.get("PLAN_QUALITY_BONUS", "8"))
|
PLAN_QUALITY_BONUS = float(os.environ.get("PLAN_QUALITY_BONUS", "8"))
|
||||||
|
|
||||||
|
# --- 计划排序主轴(2026-09-17 方案;台账顺延)------------------------------------
|
||||||
|
# 三份读数(09-02 §1.8、08-28 周报、09-11 抽查)一致:传导越强超额越负,传导分的公式奖励的
|
||||||
|
# 正是被证明为负的"未动补涨"。所以把传导从排序主轴换成"雷达门加带内次级键"。
|
||||||
|
# quality(默认):主榜与观察档按(质地带、被指向、组内分、原始分、代码)词典序,传导只在带内决定先后。
|
||||||
|
# transmission:回 09-14 那套(质地加减分、不跨档),逐字回旧,单测钉住。
|
||||||
|
PLAN_RANK_AXIS = os.environ.get("PLAN_RANK_AXIS", "quality").strip().lower()
|
||||||
|
# 候选卡内序(2026-09-17):默认在"先判决"之后、"再涨幅"之前插入质地带(好票同判决内前移)。
|
||||||
|
# 设 0 回旧序"先判决再涨幅"。质地带与工作包一同一个判定函数,带一百二十天新鲜度守卫。
|
||||||
|
CARD_SORT_BY_QUALITY = os.environ.get("CARD_SORT_BY_QUALITY", "1") == "1"
|
||||||
|
|
||||||
# --- 统一任务调度平台触发(08-03;键名与决策系统保持一致,便于平台侧统一配置)----
|
# --- 统一任务调度平台触发(08-03;键名与决策系统保持一致,便于平台侧统一配置)----
|
||||||
# 空串 = /api/v1/xxl/* 整组端点禁用(安全默认)。回调契约见 xxl.py 模块说明。
|
# 空串 = /api/v1/xxl/* 整组端点禁用(安全默认)。回调契约见 xxl.py 模块说明。
|
||||||
XXL_TRIGGER_KEY = os.environ.get("XXL_TRIGGER_KEY", "")
|
XXL_TRIGGER_KEY = os.environ.get("XXL_TRIGGER_KEY", "")
|
||||||
|
|
@ -225,10 +236,19 @@ REGIME_API_TIMEOUT = float(os.environ.get("REGIME_API_TIMEOUT", "6"))
|
||||||
# 弱势日定义(09-02 拍板,首份周报前锁定、之后不改;改它算新一轮验证):八个指数里弱势个数达到几个。
|
# 弱势日定义(09-02 拍板,首份周报前锁定、之后不改;改它算新一轮验证):八个指数里弱势个数达到几个。
|
||||||
REGIME_WEAK_COUNT = int(os.environ.get("REGIME_WEAK_COUNT", "3"))
|
REGIME_WEAK_COUNT = int(os.environ.get("REGIME_WEAK_COUNT", "3"))
|
||||||
|
|
||||||
# --- 入池与候选卡的联动(09-02;第一步只加字段与切片,PMS 行为不变)------------------
|
# --- 入池与候选卡的联动(09-02 起字段与切片;09-17 起默认改 rank)--------------------
|
||||||
# POOL_SOURCE:现状 = 强传导档前 POOL_TOP(默认,不变);candidate = 候选优先、再按强传导档补足到 POOL_TOP
|
# POOL_SOURCE:
|
||||||
# (第二步,复盘读数齐后拍板再切)。
|
# rank(默认,2026-09-17)= 主榜新序(质地优先)里被指向的前 POOL_TOP。它与 PMS 看到的候选
|
||||||
POOL_SOURCE = os.environ.get("POOL_SOURCE", "tier")
|
# 前缀同一个序,质地好的票才不会永远缺"吸筹评分"这条确认线(入池决定谁拿吸筹评分)。
|
||||||
|
# tier = 强传导档前 POOL_TOP(09-14 前旧行为,回退用)。
|
||||||
|
# candidate = 候选优先、再按强传导档补足到 POOL_TOP(09-02 第二步)。
|
||||||
|
POOL_SOURCE = os.environ.get("POOL_SOURCE", "rank")
|
||||||
|
# 备用开关(2026-09-17):rank 源下把旧序(rank_old)强传导前 POOL_TOP 也并进池、去重,受 POOL_MAX 兜底。
|
||||||
|
# 默认关;只在上线次日"池包含 PMS 候选"这条不变式核不过时打开,交回读数。
|
||||||
|
POOL_RANK_KEEP_OLD = os.environ.get("POOL_RANK_KEEP_OLD", "0") == "1"
|
||||||
|
# 评析目标名单排序轴(2026-09-17):both(默认)= 原始分前一百 ∪ 新序前五十的并集,评析覆盖与排序轴
|
||||||
|
# 解耦(新被指向的公司先有报告,好公司报告保持新鲜);score = 只按原始分前一百(逐字回旧)。
|
||||||
|
REVIEW_TARGETS_AXIS = os.environ.get("REVIEW_TARGETS_AXIS", "both").strip().lower()
|
||||||
# 低优先入池切片:让"门槛全过但缺吸筹评分"的关注票入池、当晚获得评分,否则复盘缺数据是环状依赖。
|
# 低优先入池切片:让"门槛全过但缺吸筹评分"的关注票入池、当晚获得评分,否则复盘缺数据是环状依赖。
|
||||||
# 0 = 不开(默认);受 POOL_MAX 约束,只改 Mongo 池成分,PMS 不读 Mongo 池。
|
# 0 = 不开(默认);受 POOL_MAX 约束,只改 Mongo 池成分,PMS 不读 Mongo 池。
|
||||||
POOL_WATCH_SLICE = int(os.environ.get("POOL_WATCH_SLICE", "0"))
|
POOL_WATCH_SLICE = int(os.environ.get("POOL_WATCH_SLICE", "0"))
|
||||||
|
|
|
||||||
|
|
@ -563,3 +563,19 @@
|
||||||
- **部署与预期。** 本地做完提交,155 拉代码后重启桥容器(先经审批,避开 07:00 到 07:30 与 08:30 到 09:05),重启后预热一次计划接口。次日早七点十分起,入池与评析名单都按新序。持仓管理系统榜单当天会有一波名次跳动,属预期。
|
- **部署与预期。** 本地做完提交,155 拉代码后重启桥容器(先经审批,避开 07:00 到 07:30 与 08:30 到 09:05),重启后预热一次计划接口。次日早七点十分起,入池与评析名单都按新序。持仓管理系统榜单当天会有一波名次跳动,属预期。
|
||||||
|
|
||||||
- **复核日期。** 重启并出一次计划后,按 §5.4 的 curl 看前 50 里质地好的只数是否上升、rank 是否连续、每行是否都带 score_adj。
|
- **复核日期。** 重启并出一次计划后,按 §5.4 的 curl 看前 50 里质地好的只数是否上升、rank 是否连续、每行是否都带 score_adj。
|
||||||
|
|
||||||
|
## 063 · 2026-09-17 · 选股计划排序轴换成公司质地,传导降为雷达门与带内次级键
|
||||||
|
|
||||||
|
排序主轴从传导档位换成公司质地。方案与规格书见 docs/选股计划调整方案_2026-09-17.md(四仓库同步副本)。这条是对 09-03 方案里"不改传导档位、分数编码、候选卡两个旋钮"那一句的有意修订——公式、编码、两个旋钮都守住了,改的是计划装配层的次序。
|
||||||
|
|
||||||
|
- **改动。** 五处,全在选股系统仓库:主榜与观察档按(质地带、被指向、组内分、原始分、代码)词典序重排(plan.py);候选卡内序在判决与涨幅之间插入质地带(card.py);入池改成主榜新序里被指向的前 POOL_TOP(pool.py);评析目标名单改成原始分前一百与新序前五十的并集、与排序轴解耦(review_targets.py);周报四周并列新序前二十与旧序前二十、加档位乘质地带交叉分组(plan_review.py)。分数编码 factors.py 一字未动,三门槛一确认三风险一字未动。持仓管理系统按上游名次排是协同项,另批,不阻塞。
|
||||||
|
|
||||||
|
- **依据。** 三份读数一致:09-02 改进方案 §1.8(强传导五日超额 −2.04、弱 −0.39、无 0.00)、08-28 周报(主榜强传导 −1.97/−4.84,无传导 −1.09/−1.93)、09-11 抽查(主榜前一百有评析的七十只里差五十、中十四、好六)。根因是传导分等于源数乘以"一减已动比例",奖励的正是被证明为负的"未动补涨";而"环节被指向"这一位本身定位不负,所以保留它当雷达门与带内次级键,只把它从排序主轴上撤下来。质地数据源已稳定(三百五十九家,好二十一、中八十七、差二百五十一)。
|
||||||
|
|
||||||
|
- **开关与回退。** 四个开关 PLAN_RANK_AXIS、CARD_SORT_BY_QUALITY、POOL_SOURCE、REVIEW_TARGETS_AXIS 全设旧值即逐字回旧,下一次装配即生效,不用重启、不用回代码。回退也记台账。备用开关 POOL_RANK_KEEP_OLD 只在上线次日"池包含 PMS 候选"这条不变式核不过时打开。
|
||||||
|
|
||||||
|
- **预期。** 候选单与前三十名以质地好、中为主;四周内新序前二十相对旧序前二十的五日与十日超额都不更差、且候选单跑赢全池的占比不低于旧序。
|
||||||
|
|
||||||
|
- **验证读数。** 开发机 python3 test_plan_rank_axis.py 与 test_plan_quality_rank.py 均 ALL OK(含四开关设旧值逐字回旧那一例)。上线次日按方案第六节看候选单质地列、前十二名质地分布、入池来源、"池含候选"不变式、评析名单两路并集五处读数。
|
||||||
|
|
||||||
|
- **复核日期。** 09-21 收盘后 155 拉代码(盘前链子进程读新代码不用重启;接口容器建议同窗重启一次并预热,避开七点到七点半与八点半到九点零五)。09-22 首份新序计划,09-25 首份对照周报,上线后四周(10-16 前后)按预期那两条判据决定保留或切回。
|
||||||
|
|
|
||||||
|
|
@ -123,3 +123,11 @@ docker compose exec -T akg-factor-bridge python run.py push-pool
|
||||||
自动维护:入池、留池、出池由 pool.push() 每天在盘前链里维护,持仓永不出池,掉榜但形态没恶化的票留在池里继续观察,形态恶化的进回收站。名单不需要手工维护。
|
自动维护:入池、留池、出池由 pool.push() 每天在盘前链里维护,持仓永不出池,掉榜但形态没恶化的票留在池里继续观察,形态恶化的进回收站。名单不需要手工维护。
|
||||||
|
|
||||||
参数在哪:POOL_TOP 和 POOL_MAX 在 akg-factor-bridge 的 .env 里;PMS_PLAN_TOP_N 在 tradingSystem 的 config/settings.py 里,也可被 PMS 的 param_store 覆盖。改完 .env 要重建容器让新值生效,见第 5 节部署,注意 restart 不重读 env。
|
参数在哪:POOL_TOP 和 POOL_MAX 在 akg-factor-bridge 的 .env 里;PMS_PLAN_TOP_N 在 tradingSystem 的 config/settings.py 里,也可被 PMS 的 param_store 覆盖。改完 .env 要重建容器让新值生效,见第 5 节部署,注意 restart 不重读 env。
|
||||||
|
|
||||||
|
## 8. 入池来源改成质地新序里被指向的前若干(2026-09-17,台账 063)
|
||||||
|
|
||||||
|
入池口径从"强传导档前 POOL_TOP"改成"主榜新序(公司质地优先)里被传导指向的前 POOL_TOP"(开关 `POOL_SOURCE`,默认 `rank`;`tier` 回旧、`candidate` 保留)。夜间推理量不变——仍是 POOL_TOP 只。变的只是这批是谁:从"强传导前五十"变成"质地优先且被指向的前五十"。
|
||||||
|
|
||||||
|
为什么必须与主榜同序:入池决定谁拿得到吸筹评分,而吸筹评分又是候选卡的确认线。池子若与持仓管理系统看到的候选前缀不是同一个序,质地好的票就永远缺确认、永远进不了候选。
|
||||||
|
|
||||||
|
上线次日要核一条不变式:持仓管理系统的候选都在池里(方案第六节有命令,预期差集为空)。核不过就打开备用开关 `POOL_RANK_KEEP_OLD`(默认关),把旧序强传导前 POOL_TOP 也并进池、去重,受 POOL_MAX 兜底,并把读数交回。回退到旧行为把 `POOL_SOURCE` 设回 `tier` 即可。
|
||||||
|
|
|
||||||
|
|
@ -56,3 +56,19 @@
|
||||||
测算约为现规模的三分之一),届时另行通知。
|
测算约为现规模的三分之一),届时另行通知。
|
||||||
- 传导史自 2026-07 起为纯实时积累,无回填历史;系统各环节的判定依据均可
|
- 传导史自 2026-07 起为纯实时积累,无回填历史;系统各环节的判定依据均可
|
||||||
审计(每只候选可回溯到具体的图谱边、异动台账与研报行)。
|
审计(每只候选可回溯到具体的图谱边、异动台账与研报行)。
|
||||||
|
|
||||||
|
## 名次改按公司质地(2026-09-17 更新,台账 063)
|
||||||
|
|
||||||
|
- **排序主轴从传导换成了公司质地。** 主榜、观察档、候选单的名次现在先看公司质地
|
||||||
|
(个股深度评析算出的好、中、差档),传导只在同一质地档内决定先后,并继续当候选卡
|
||||||
|
门槛一("所在环节被产业链指向")。原因:一个月与两周的复盘都显示传导越强、随后的
|
||||||
|
超额越负——传导对"看哪个环节"有信息量,但"在环节里挑还没动的补涨"这一步是负的。
|
||||||
|
- **每行多四个键**:`quality_band`(质地带:好 3、中 2、无评析 1、差 0)、`rank_axis`
|
||||||
|
(quality 或 transmission)、`pointed`(是否被传导指向)、`rank_old`(旧序名次,仅供四周对照)。
|
||||||
|
原有的 `score`、`tier` 值与含义不变。
|
||||||
|
- **持仓管理系统请注意前缀效应**:质地差但强传导的票,可能落到你要的三百只主榜前缀之外。
|
||||||
|
这与"质地差降关注"同向,是预期结果,不要把"强传导票变少"当故障。逐票对账接口
|
||||||
|
`/plan/verdict` 的"主榜第 N 名"仍按原始分序(应答里带 `rank_axis: transmission` 说明),
|
||||||
|
与计划里按质地排的 `rank` 会不一致,属正常。
|
||||||
|
- 回退:把选股系统 `.env` 的 `PLAN_RANK_AXIS` 设回 `transmission`(连同另三个开关),
|
||||||
|
下一次装配即逐字回到旧序。
|
||||||
|
|
|
||||||
114
plan.py
114
plan.py
|
|
@ -134,6 +134,42 @@ def _rerank_by_quality(ranked, cards: dict, bonus: float):
|
||||||
return ranked.reindex(order)
|
return ranked.reindex(order)
|
||||||
|
|
||||||
|
|
||||||
|
def _rerank_by_axis(ranked, cards: dict, axis: str, stale_days: int, has_tier: bool = True):
|
||||||
|
"""质地轴重排(2026-09-17):按(质地带降序、被指向降序、组内分降序、原始分降序、代码升序)词典序。
|
||||||
|
只改顺序、不改分值——档位标签仍按原始分解码。传导强弱档位不再进次序,只保留"被指向"这一位。
|
||||||
|
|
||||||
|
"被指向"以候选卡门槛一为准(cards[k]['gates']['pointed']),不看传导视图的证据行——已启动成员在
|
||||||
|
那张视图里没有证据行,只看证据行会把它误判成没被指向。组内分从原始分反解:主榜分减二百再减档位
|
||||||
|
乘二十(档位取 int((score-190)//20));观察档没有档位,带内直接按原始分。空榜原样返回。"""
|
||||||
|
if ranked.empty:
|
||||||
|
return ranked
|
||||||
|
|
||||||
|
def _key(item):
|
||||||
|
k = item[0]
|
||||||
|
s = float(item[1])
|
||||||
|
c = cards.get(k) or {}
|
||||||
|
band = c.get("quality_band")
|
||||||
|
if band is None:
|
||||||
|
band = card.quality_band(c.get("company_review"), stale_days)
|
||||||
|
pointed = bool((c.get("gates") or {}).get("pointed"))
|
||||||
|
if has_tier:
|
||||||
|
tier = int((s - 190.0) // 20)
|
||||||
|
group = s - 200.0 - tier * 20.0 # 剥掉档位:传导档位不进次序
|
||||||
|
else:
|
||||||
|
group = s # 观察档无档位,带内直接按原始分
|
||||||
|
return (-int(band), not pointed, -group, -s, k) # 每层升序排,负号即降序
|
||||||
|
|
||||||
|
order = [k for k, _ in sorted(ranked.items(), key=_key)]
|
||||||
|
return ranked.reindex(order)
|
||||||
|
|
||||||
|
|
||||||
|
def _rerank(ranked, cards: dict, has_tier: bool):
|
||||||
|
"""按 config.PLAN_RANK_AXIS 分派:quality(默认)走质地轴;transmission 逐字回旧走 09-14 的质地加成。"""
|
||||||
|
if config.PLAN_RANK_AXIS == "quality":
|
||||||
|
return _rerank_by_axis(ranked, cards, "quality", config.COMPANY_REVIEW_STALE_DAYS, has_tier)
|
||||||
|
return _rerank_by_quality(ranked, cards, config.PLAN_QUALITY_BONUS)
|
||||||
|
|
||||||
|
|
||||||
def _val(series: pd.Series, k: str):
|
def _val(series: pd.Series, k: str):
|
||||||
v = series.get(k)
|
v = series.get(k)
|
||||||
return None if v is None or pd.isna(v) else float(v)
|
return None if v is None or pd.isna(v) else float(v)
|
||||||
|
|
@ -376,7 +412,8 @@ def _assemble_cards(ds: str, codes: list, ev: dict, upside: pd.Series,
|
||||||
neg_tol=config.UPSIDE_NEG_TOLERANCE,
|
neg_tol=config.UPSIDE_NEG_TOLERANCE,
|
||||||
logic_stale_days=config.LOGIC_STALE_DAYS,
|
logic_stale_days=config.LOGIC_STALE_DAYS,
|
||||||
require_started=config.CARD_REQUIRE_STARTED,
|
require_started=config.CARD_REQUIRE_STARTED,
|
||||||
company_review=company.get(k), quality_gate=config.CARD_COMPANY_QUALITY_GATE)
|
company_review=company.get(k), quality_gate=config.CARD_COMPANY_QUALITY_GATE,
|
||||||
|
stale_days=config.COMPANY_REVIEW_STALE_DAYS)
|
||||||
cards[k] = {
|
cards[k] = {
|
||||||
**j, "logic_state": state,
|
**j, "logic_state": state,
|
||||||
"valuation": valuation.get(k), "valuation_text": card.valuation_view(valuation.get(k)),
|
"valuation": valuation.get(k), "valuation_text": card.valuation_view(valuation.get(k)),
|
||||||
|
|
@ -470,22 +507,34 @@ def collect(date: str | None = None, top: int = 20, obs_top: int = 10,
|
||||||
cards, segments_pointed = _assemble_cards(
|
cards, segments_pointed = _assemble_cards(
|
||||||
ds, list(main.index) + list(obs.index), ev, upside, mkt_days, risk=risk)
|
ds, list(main.index) + list(obs.index), ev, upside, mkt_days, risk=risk)
|
||||||
|
|
||||||
# 质地调整分(2026-09-14 工作包四):拿到评析后,就在主榜与观察档这两个分数序列上重排一次。
|
# 排序轴(2026-09-17):拿到评析后在主榜与观察档这两个分数序列上重排一次。
|
||||||
# 下面的 _pick 出接口、_full_rows 出复盘快照、快照又喂逐票逻辑状态入库与评析目标名单——
|
# 下面的 _pick 出接口、_full_rows 出复盘快照、快照又喂逐票逻辑状态入库与评析目标名单——
|
||||||
# 四个出口都从这两个序列取,所以只在这里改一次,四处才同序。只改顺序、不改分值。
|
# 四个出口都从这两个序列取,所以只在这里改一次,四处才同序。只改顺序、不改分值。
|
||||||
main = _rerank_by_quality(main, cards, config.PLAN_QUALITY_BONUS)
|
# 质地轴下先用 09-14 的旧序(_rerank_by_quality)算一遍全量名次拿 rank_old(供周报四周对照),再算新序。
|
||||||
obs = _rerank_by_quality(obs, cards, config.PLAN_QUALITY_BONUS)
|
axis = config.PLAN_RANK_AXIS
|
||||||
|
main_old = _rerank_by_quality(main, cards, config.PLAN_QUALITY_BONUS)
|
||||||
|
obs_old = _rerank_by_quality(obs, cards, config.PLAN_QUALITY_BONUS)
|
||||||
|
rank_old_map: dict = {}
|
||||||
|
rank_old_map.update({k: i for i, k in enumerate(main_old.index, 1)})
|
||||||
|
rank_old_map.update({k: i for i, k in enumerate(obs_old.index, 1)})
|
||||||
|
main = _rerank(main, cards, True)
|
||||||
|
obs = _rerank(obs, cards, False)
|
||||||
|
|
||||||
def _pick(ranked: pd.Series, n: int):
|
def _pick(ranked: pd.Series, n: int, uncapped_no_theme: bool = False):
|
||||||
"""分数从高到低取 n 条;每个传导主题最多 theme_cap 条(0=不设限)——
|
"""分数从高到低取 n 条;每个传导主题最多 theme_cap 条(0=不设限)——
|
||||||
传导目标是环节级、同环节成员共享同一条证据,不限额会被少数环节刷屏。"""
|
传导目标是环节级、同环节成员共享同一条证据,不限额会被少数环节刷屏。
|
||||||
|
uncapped_no_theme(质地轴下为真):没有传导主题的行不计数、不受主题限额。质地序前列有不少
|
||||||
|
质地好的无传导票,若仍受五只上限,第六只起就被质地中的被指向票顶掉,主轴被限额悄悄推翻;
|
||||||
|
无传导本就不是板块,限额防的是单板块刷屏。有主题的行照旧每主题最多 theme_cap 条。"""
|
||||||
out, cnt = [], {}
|
out, cnt = [], {}
|
||||||
for k, s in ranked.items():
|
for k, s in ranked.items():
|
||||||
e = ev.get(k)
|
e = ev.get(k)
|
||||||
theme = e[0] if e else "(无传导)"
|
theme = e[0] if e else "(无传导)"
|
||||||
if theme_cap and cnt.get(theme, 0) >= theme_cap:
|
exempt = uncapped_no_theme and e is None
|
||||||
continue
|
if not exempt:
|
||||||
cnt[theme] = cnt.get(theme, 0) + 1
|
if theme_cap and cnt.get(theme, 0) >= theme_cap:
|
||||||
|
continue
|
||||||
|
cnt[theme] = cnt.get(theme, 0) + 1
|
||||||
out.append((k, s))
|
out.append((k, s))
|
||||||
if len(out) >= n:
|
if len(out) >= n:
|
||||||
break
|
break
|
||||||
|
|
@ -508,6 +557,16 @@ def collect(date: str | None = None, top: int = 20, obs_top: int = 10,
|
||||||
float(s) + _quality_bonus(c.get("company_review"), config.PLAN_QUALITY_BONUS), 2),
|
float(s) + _quality_bonus(c.get("company_review"), config.PLAN_QUALITY_BONUS), 2),
|
||||||
"evidence": evidence,
|
"evidence": evidence,
|
||||||
"heat": _val(heat, k), "upside": _val(upside, k)}
|
"heat": _val(heat, k), "upside": _val(upside, k)}
|
||||||
|
# 排序轴四键(2026-09-17):质地带、主轴名、是否被指向(候选卡门槛一)、旧序全量名次。
|
||||||
|
# rank_old 只在质地轴下填(由 09-14 旧序再算一遍得到),供周报四周对照,四周后可停。
|
||||||
|
qb = c.get("quality_band")
|
||||||
|
if qb is None:
|
||||||
|
qb = card.quality_band(c.get("company_review"), config.COMPANY_REVIEW_STALE_DAYS)
|
||||||
|
r["quality_band"] = int(qb)
|
||||||
|
r["rank_axis"] = config.PLAN_RANK_AXIS
|
||||||
|
r["pointed"] = bool((c.get("gates") or {}).get("pointed"))
|
||||||
|
if config.PLAN_RANK_AXIS == "quality":
|
||||||
|
r["rank_old"] = rank_old_map.get(k)
|
||||||
if with_tier:
|
if with_tier:
|
||||||
r["tier"] = _tier_label(s)
|
r["tier"] = _tier_label(s)
|
||||||
if c:
|
if c:
|
||||||
|
|
@ -635,7 +694,7 @@ def collect(date: str | None = None, top: int = 20, obs_top: int = 10,
|
||||||
|
|
||||||
card_counts = {v: sum(1 for c in cards.values() if c.get("verdict") == v)
|
card_counts = {v: sum(1 for c in cards.values() if c.get("verdict") == v)
|
||||||
for v in (card.VERDICT_CANDIDATE, card.VERDICT_WATCH, card.VERDICT_SHOW)}
|
for v in (card.VERDICT_CANDIDATE, card.VERDICT_WATCH, card.VERDICT_SHOW)}
|
||||||
return {
|
result = {
|
||||||
"date": ds,
|
"date": ds,
|
||||||
"generated_at": generated_at,
|
"generated_at": generated_at,
|
||||||
"plan_version": version.git_short_rev(),
|
"plan_version": version.git_short_rev(),
|
||||||
|
|
@ -645,13 +704,14 @@ def collect(date: str | None = None, top: int = 20, obs_top: int = 10,
|
||||||
"heat_date": hd,
|
"heat_date": hd,
|
||||||
"theme_cap": theme_cap,
|
"theme_cap": theme_cap,
|
||||||
"main": [_row(i, k, s, True)
|
"main": [_row(i, k, s, True)
|
||||||
for i, (k, s) in enumerate(_pick(main, top), 1)],
|
for i, (k, s) in enumerate(_pick(main, top, uncapped_no_theme=(axis == "quality")), 1)],
|
||||||
"observe": [_row(i, k, s, False)
|
"observe": [_row(i, k, s, False)
|
||||||
for i, (k, s) in enumerate(_pick(obs, obs_top), 1)],
|
for i, (k, s) in enumerate(_pick(obs, obs_top, uncapped_no_theme=(axis == "quality")), 1)],
|
||||||
"changes": changes,
|
"changes": changes,
|
||||||
"gate_on": bool(config.ENABLE_TRACK_GATE),
|
"gate_on": bool(config.ENABLE_TRACK_GATE),
|
||||||
"encoding": "主榜分=200+传导档位×20+组内分(还没热、还便宜);"
|
"encoding": "主榜分=200+传导档位×20+组内分(还没热、还便宜);"
|
||||||
"观察档分=100+0.6z(传导)+0.4z(−热度)",
|
"观察档分=100+0.6z(传导)+0.4z(−热度);"
|
||||||
|
"名次按质地带优先,传导只在带内作次级键(开关 PLAN_RANK_AXIS)",
|
||||||
# ---- 候选卡(2026-09-02):分数与档位之外的另一份产物,不受 top / theme_cap 裁剪 ----
|
# ---- 候选卡(2026-09-02):分数与档位之外的另一份产物,不受 top / theme_cap 裁剪 ----
|
||||||
"card_params": {"start_pct": config.CARD_START_PCT,
|
"card_params": {"start_pct": config.CARD_START_PCT,
|
||||||
"require_started": config.CARD_REQUIRE_STARTED,
|
"require_started": config.CARD_REQUIRE_STARTED,
|
||||||
|
|
@ -668,6 +728,19 @@ def collect(date: str | None = None, top: int = 20, obs_top: int = 10,
|
||||||
"segments_pointed": segments_pointed,
|
"segments_pointed": segments_pointed,
|
||||||
"_full": {"main": _full_rows(main, True), "observe": _full_rows(obs, False)},
|
"_full": {"main": _full_rows(main, True), "observe": _full_rows(obs, False)},
|
||||||
}
|
}
|
||||||
|
# 排序轴三个只读键(2026-09-17,只在质地轴下加):主轴名、一句话次序规则、主榜与观察档各带只数。
|
||||||
|
if axis == "quality":
|
||||||
|
def _band_counts(rows: list) -> dict:
|
||||||
|
out: dict = {}
|
||||||
|
for r in rows:
|
||||||
|
b = r.get("quality_band")
|
||||||
|
out[b] = out.get(b, 0) + 1
|
||||||
|
return out
|
||||||
|
result["rank_axis"] = axis
|
||||||
|
result["rank_rule"] = "质地带(好3中2无评析1差0)降序 → 带内被指向优先 → 组内分 → 原始分 → 代码"
|
||||||
|
result["quality_band_counts"] = {"main": _band_counts(result["_full"]["main"]),
|
||||||
|
"observe": _band_counts(result["_full"]["observe"])}
|
||||||
|
return result
|
||||||
|
|
||||||
|
|
||||||
def _fmt_pct(v) -> str:
|
def _fmt_pct(v) -> str:
|
||||||
|
|
@ -938,19 +1011,26 @@ def snapshot_fresh_enough(snap: dict) -> tuple:
|
||||||
return True, ""
|
return True, ""
|
||||||
|
|
||||||
|
|
||||||
def pick_rows(rows, n: int, theme_cap: int) -> list:
|
def pick_rows(rows, n: int, theme_cap: int, uncapped_no_theme=None) -> list:
|
||||||
"""从**已按分数排好序**的行里取 n 条,每个传导主题最多 theme_cap 条(0=不设限)。
|
"""从**已按分数排好序**的行里取 n 条,每个传导主题最多 theme_cap 条(0=不设限)。
|
||||||
|
|
||||||
与 collect 里那个同名的内部函数是同一套规则,差别只在取主题的来源:那边回查证据映射,
|
与 collect 里那个同名的内部函数是同一套规则,差别只在取主题的来源:那边回查证据映射,
|
||||||
这边直接读行里的 evidence.theme —— 装配时就把主题写进每一行了,快照里也带着。
|
这边直接读行里的 evidence.theme —— 装配时就把主题写进每一行了,快照里也带着。
|
||||||
两处规则必须一致,test_plan_snapshot.py 拿真实快照逐行比对钉住这件事。
|
两处规则必须一致,test_plan_snapshot.py 拿真实快照逐行比对钉住这件事。
|
||||||
|
|
||||||
|
uncapped_no_theme(2026-09-17):质地轴下没有传导主题的行免限额、不计数(与 _pick 同口径)。
|
||||||
|
缺省 None 时从行的 rank_axis 推断(质地轴即为真),周报重建旧序时显式传 False。
|
||||||
"""
|
"""
|
||||||
|
if uncapped_no_theme is None:
|
||||||
|
uncapped_no_theme = bool(rows) and (rows[0].get("rank_axis") == "quality")
|
||||||
out, cnt = [], {}
|
out, cnt = [], {}
|
||||||
for r in rows:
|
for r in rows:
|
||||||
theme = ((r.get("evidence") or {}).get("theme")) or "(无传导)"
|
theme = ((r.get("evidence") or {}).get("theme")) or "(无传导)"
|
||||||
if theme_cap and cnt.get(theme, 0) >= theme_cap:
|
exempt = uncapped_no_theme and theme == "(无传导)"
|
||||||
continue
|
if not exempt:
|
||||||
cnt[theme] = cnt.get(theme, 0) + 1
|
if theme_cap and cnt.get(theme, 0) >= theme_cap:
|
||||||
|
continue
|
||||||
|
cnt[theme] = cnt.get(theme, 0) + 1
|
||||||
out.append(r)
|
out.append(r)
|
||||||
if len(out) >= n:
|
if len(out) >= n:
|
||||||
break
|
break
|
||||||
|
|
|
||||||
|
|
@ -214,6 +214,9 @@ def verdict(k: str, L: dict) -> dict:
|
||||||
"score": (round(float(sc), 1) if sc is not None else None),
|
"score": (round(float(sc), 1) if sc is not None else None),
|
||||||
"rank": rank,
|
"rank": rank,
|
||||||
"rank_total": rank_total,
|
"rank_total": rank_total,
|
||||||
|
# 逐票对账的"主榜第 N 名"仍按原始分序(语义即计划里的 rank_old),与计划的质地序 rank 会不一致;
|
||||||
|
# 加这一键写明这里是传导(原始分)轴,避免与计划的 rank 混淆(2026-09-17)。
|
||||||
|
"rank_axis": "transmission",
|
||||||
"tier": tier,
|
"tier": tier,
|
||||||
"upside": (round(float(up), 4) if up is not None else None),
|
"upside": (round(float(up), 4) if up is not None else None),
|
||||||
"heat": (round(float(ht), 2) if ht is not None else None),
|
"heat": (round(float(ht), 2) if ht is not None else None),
|
||||||
|
|
|
||||||
|
|
@ -382,6 +382,13 @@ def run(since: str, until: str | None, horizons: tuple, start: str, out_dir: str
|
||||||
tier_of = {r["code"]: r.get("tier") for r in main_rows}
|
tier_of = {r["code"]: r.get("tier") for r in main_rows}
|
||||||
verdict_of = {r["code"]: r.get("verdict") for r in main_rows + obs_rows}
|
verdict_of = {r["code"]: r.get("verdict") for r in main_rows + obs_rows}
|
||||||
seg_codes = sorted({r["code"] for r in main_rows + obs_rows if r.get("evidence")})
|
seg_codes = sorted({r["code"] for r in main_rows + obs_rows if r.get("evidence")})
|
||||||
|
# 过渡期四周对照(2026-09-17):新序前二十(main_rows 已按 rank 排)与旧序前二十(按 rank_old)。
|
||||||
|
# rank_old 只在质地轴下有;旧快照缺该键就跳过这两份名单。
|
||||||
|
band_of = {r["code"]: r.get("quality_band") for r in main_rows}
|
||||||
|
main_new20 = [r["code"] for r in main_rows[:20]]
|
||||||
|
main_old20 = ([r["code"] for r in sorted(
|
||||||
|
main_rows, key=lambda r: r.get("rank_old") if r.get("rank_old") is not None else 1 << 60)[:20]]
|
||||||
|
if main_rows and main_rows[0].get("rank_old") is not None else None)
|
||||||
prod, prod_note = pms_roster(day)
|
prod, prod_note = pms_roster(day)
|
||||||
if not prod:
|
if not prod:
|
||||||
# 无 PMS 快照的日子不再退化为桥重算(那不是 PMS 拿到的名单):交付名单当日剔除并注记(台账 010)
|
# 无 PMS 快照的日子不再退化为桥重算(那不是 PMS 拿到的名单):交付名单当日剔除并注记(台账 010)
|
||||||
|
|
@ -419,6 +426,9 @@ def run(since: str, until: str | None, horizons: tuple, start: str, out_dir: str
|
||||||
"强传导交付名单": prod, "候选单": cands, "关注单": watch, "环节名单": seg_codes,
|
"强传导交付名单": prod, "候选单": cands, "关注单": watch, "环节名单": seg_codes,
|
||||||
"主榜等权": main_codes, "观察档等权": obs_codes,
|
"主榜等权": main_codes, "观察档等权": obs_codes,
|
||||||
"全池等权": list(ret.dropna().index),
|
"全池等权": list(ret.dropna().index),
|
||||||
|
# 过渡期四周并列(2026-09-17):新序前二十与旧序前二十,看排序轴换掉后前列超额的变化。
|
||||||
|
"主榜新序前二十": main_new20,
|
||||||
|
**({"主榜旧序前二十": main_old20} if main_old20 else {}),
|
||||||
# 2026-09-03:PMS 账本三份与择时看多一份,与其余名单同口径算收益
|
# 2026-09-03:PMS 账本三份与择时看多一份,与其余名单同口径算收益
|
||||||
**ledger, TIMING_LIST: bullish,
|
**ledger, TIMING_LIST: bullish,
|
||||||
}
|
}
|
||||||
|
|
@ -434,6 +444,16 @@ def run(since: str, until: str | None, horizons: tuple, start: str, out_dir: str
|
||||||
if s:
|
if s:
|
||||||
rows.append({"date": day, "h": h, "list": "主榜", "group": f"档位={tier}",
|
rows.append({"date": day, "h": h, "list": "主榜", "group": f"档位={tier}",
|
||||||
"regime_post": regime_post, "regime_pre": regime_pre, **s})
|
"regime_post": regime_post, "regime_pre": regime_pre, **s})
|
||||||
|
# 档位×质地带交叉分组(2026-09-17):看传导在质地档内还剩多少信息量
|
||||||
|
_band_label = {3: "好", 2: "中", 1: "无评析", 0: "差"}
|
||||||
|
for tier in ("强传导", "弱传导", "无传导"):
|
||||||
|
for b in (3, 2, 1, 0):
|
||||||
|
codes = [c for c, t in tier_of.items() if t == tier and band_of.get(c) == b]
|
||||||
|
s = summarize(ret, codes, base_all, base_main, caps, cap_base)
|
||||||
|
if s:
|
||||||
|
rows.append({"date": day, "h": h, "list": "主榜",
|
||||||
|
"group": f"档位×质地={tier}×{_band_label[b]}",
|
||||||
|
"regime_post": regime_post, "regime_pre": regime_pre, **s})
|
||||||
for v in ("候选", "关注", "仅展示"):
|
for v in ("候选", "关注", "仅展示"):
|
||||||
codes = [c for c, vv in verdict_of.items() if vv == v]
|
codes = [c for c, vv in verdict_of.items() if vv == v]
|
||||||
s = summarize(ret, codes, base_all, base_main, caps, cap_base)
|
s = summarize(ret, codes, base_all, base_main, caps, cap_base)
|
||||||
|
|
|
||||||
59
pool.py
59
pool.py
|
|
@ -168,9 +168,53 @@ def strategy_context_entry(r: dict, plan_date: str, seg_ratio=None) -> dict:
|
||||||
"pct0": card_.get("pct0"),
|
"pct0": card_.get("pct0"),
|
||||||
"accum_state": accum.get("state"),
|
"accum_state": accum.get("state"),
|
||||||
"accum_age": accum.get("age"),
|
"accum_age": accum.get("age"),
|
||||||
|
# 排序轴三键(2026-09-17):质地带、主轴名、旧序全量名次。只加字段,择时决策系统读池只取代码列。
|
||||||
|
"quality_band": r.get("quality_band"),
|
||||||
|
"rank_axis": r.get("rank_axis"),
|
||||||
|
"rank_old": r.get("rank_old"),
|
||||||
"segment_started_ratio": (seg_ratio or {}).get(theme) if theme else None}
|
"segment_started_ratio": (seg_ratio or {}).get(theme) if theme else None}
|
||||||
|
|
||||||
|
|
||||||
|
def select_plan_rows(main_rows, top, source, tiers, keep_old=False, candidates=None):
|
||||||
|
"""从主榜行里选入池行(纯函数,离线可测)。入池决定谁拿吸筹评分(候选卡确认线),所以池子
|
||||||
|
要与持仓管理系统看到的候选前缀同一个序,否则质地好的票永远缺确认。
|
||||||
|
|
||||||
|
rank(2026-09-17 默认):主榜新序(质地优先)里被指向的前 top,_pool_source 记 rank_top。
|
||||||
|
keep_old 为真时再并进按 rank_old 排的强传导前 top、去重、_pool_source 记 rank_old_strong,
|
||||||
|
总数受 POOL_MAX 兜底——只在上线次日"池包含候选"这条不变式核不过时打开。
|
||||||
|
tier:档位白名单里的前 top(09-14 前旧行为,回退用),逐字保留。
|
||||||
|
candidate:候选优先、再按档位白名单补足到 top(09-02 第二步,需传 candidates),逐字保留。
|
||||||
|
"""
|
||||||
|
import config
|
||||||
|
tier_rows = [r for r in main_rows if not tiers or r.get("tier") in tiers]
|
||||||
|
if source == "rank":
|
||||||
|
rows = [r for r in main_rows if r.get("pointed")][:top]
|
||||||
|
for r in rows:
|
||||||
|
r["_pool_source"] = "rank_top"
|
||||||
|
if keep_old:
|
||||||
|
have = {r["code"] for r in rows}
|
||||||
|
old_strong = sorted((r for r in tier_rows if r.get("rank_old") is not None),
|
||||||
|
key=lambda r: r["rank_old"])[:top]
|
||||||
|
for r in old_strong:
|
||||||
|
if r["code"] not in have:
|
||||||
|
r["_pool_source"] = "rank_old_strong"
|
||||||
|
rows.append(r)
|
||||||
|
have.add(r["code"])
|
||||||
|
rows = rows[:config.POOL_MAX]
|
||||||
|
return rows
|
||||||
|
if source == "candidate":
|
||||||
|
cand = list(candidates or [])
|
||||||
|
seen = {r["code"] for r in cand}
|
||||||
|
rows = (cand + [r for r in tier_rows if r["code"] not in seen])[:top]
|
||||||
|
for r in rows:
|
||||||
|
r["_pool_source"] = "candidate" if r.get("verdict") == "候选" else "tier_top"
|
||||||
|
return rows
|
||||||
|
rows = tier_rows[:top]
|
||||||
|
for r in rows:
|
||||||
|
r["_pool_source"] = "tier_top"
|
||||||
|
return rows
|
||||||
|
|
||||||
|
|
||||||
def build_recycle_docs(d: dict, group_id: str, group_name: str, org_id: str,
|
def build_recycle_docs(d: dict, group_id: str, group_name: str, org_id: str,
|
||||||
now: dt.datetime) -> list:
|
now: dt.datetime) -> list:
|
||||||
"""回收站文档,字段照抄现有格式(group_id/group_name/org_id/removal_batch/
|
"""回收站文档,字段照抄现有格式(group_id/group_name/org_id/removal_batch/
|
||||||
|
|
@ -290,18 +334,9 @@ def push(date: str | None = None, top: int | None = None,
|
||||||
theme_cap=config.POOL_THEME_CAP)
|
theme_cap=config.POOL_THEME_CAP)
|
||||||
tiers = config.POOL_TIERS
|
tiers = config.POOL_TIERS
|
||||||
ds = data["date"]
|
ds = data["date"]
|
||||||
tier_rows = [r for r in data["main"] if not tiers or r.get("tier") in tiers]
|
# 选行抽成纯函数(2026-09-17):rank 源取新序里被指向的前 top,tier/candidate 逐字保留。
|
||||||
if config.POOL_SOURCE == "candidate":
|
plan_rows = select_plan_rows(data["main"], top, config.POOL_SOURCE, tiers,
|
||||||
# 第二步(2026-09-02 方案第 2.2 节第五项,复盘读数齐后拍板再切):候选优先,
|
keep_old=config.POOL_RANK_KEEP_OLD, candidates=data.get("candidates"))
|
||||||
# 再按强传导档补足到 top;默认 POOL_SOURCE=tier 不走这里,行为与现状一致。
|
|
||||||
cand = list(data.get("candidates") or [])
|
|
||||||
seen = {r["code"] for r in cand}
|
|
||||||
plan_rows = (cand + [r for r in tier_rows if r["code"] not in seen])[:top]
|
|
||||||
else:
|
|
||||||
plan_rows = tier_rows[:top]
|
|
||||||
for r in plan_rows:
|
|
||||||
r["_pool_source"] = ("candidate" if config.POOL_SOURCE == "candidate"
|
|
||||||
and r.get("verdict") == "候选" else "tier_top")
|
|
||||||
plan_codes = [r["code"] for r in plan_rows]
|
plan_codes = [r["code"] for r in plan_rows]
|
||||||
|
|
||||||
# 2. 持仓(读不到直接抛,整轮不动池子)与旧池子
|
# 2. 持仓(读不到直接抛,整轮不动池子)与旧池子
|
||||||
|
|
|
||||||
|
|
@ -12,7 +12,8 @@ import config
|
||||||
import db
|
import db
|
||||||
|
|
||||||
TABLE = os.environ.get("REVIEW_TARGETS_TABLE", "t_akg_review_targets")
|
TABLE = os.environ.get("REVIEW_TARGETS_TABLE", "t_akg_review_targets")
|
||||||
TOP_MAIN = int(os.environ.get("REVIEW_TARGETS_MAIN", "100")) # 主榜前一百只
|
TOP_MAIN = int(os.environ.get("REVIEW_TARGETS_MAIN", "100")) # 主榜按原始分前一百只(跟着雷达走)
|
||||||
|
TOP_MAIN_RANK = int(os.environ.get("REVIEW_TARGETS_MAIN_RANK", "50")) # 主榜按新序前五十只(好公司报告保持新鲜)
|
||||||
TOP_OBS = int(os.environ.get("REVIEW_TARGETS_OBS", "20")) # 观察档前二十只
|
TOP_OBS = int(os.environ.get("REVIEW_TARGETS_OBS", "20")) # 观察档前二十只
|
||||||
COLUMNS = ("plan_date", "ts_code", "source", "rank", "name")
|
COLUMNS = ("plan_date", "ts_code", "source", "rank", "name")
|
||||||
_CREATE = """
|
_CREATE = """
|
||||||
|
|
@ -63,10 +64,36 @@ def save(ds: str, rows: list, conn_factory=None) -> None:
|
||||||
conn.commit()
|
conn.commit()
|
||||||
|
|
||||||
|
|
||||||
|
def pick_main_targets(main_rows: list, axis: str, top_score: int, top_rank: int) -> list:
|
||||||
|
"""评析目标名单的主榜取法(纯函数,2026-09-17)。评析覆盖不能被排序轴锁死——只取新序前列的话,
|
||||||
|
新被指向的公司永远排在已评析公司后面、永远拿不到报告,所以名单要与排序轴解耦。
|
||||||
|
|
||||||
|
both(默认):按 score 降序前 top_score(跟着雷达走)∪ 按 rank 升序前 top_rank(好公司报告保持新鲜),
|
||||||
|
去重后按 rank 升序排。score:只按 score 降序前 top_score(逐字回旧)。"""
|
||||||
|
def _score(r):
|
||||||
|
return -(r.get("score") or 0.0)
|
||||||
|
|
||||||
|
def _rank(r):
|
||||||
|
return r.get("rank") if r.get("rank") is not None else 1 << 60
|
||||||
|
|
||||||
|
if axis != "both":
|
||||||
|
return sorted(main_rows, key=_score)[:top_score]
|
||||||
|
seen: set = set()
|
||||||
|
union: list = []
|
||||||
|
for r in sorted(main_rows, key=_score)[:top_score] + sorted(main_rows, key=_rank)[:top_rank]:
|
||||||
|
code = r.get("code")
|
||||||
|
if code not in seen:
|
||||||
|
seen.add(code)
|
||||||
|
union.append(r)
|
||||||
|
union.sort(key=_rank)
|
||||||
|
return union
|
||||||
|
|
||||||
|
|
||||||
def persist(ds: str, snap: dict, write=None) -> dict:
|
def persist(ds: str, snap: dict, write=None) -> dict:
|
||||||
"""目标是主榜前 TOP_MAIN 只加观察档前 TOP_OBS 只(快照里的 main / observe 已按名次排),约一百二十家,
|
"""目标是主榜(按 REVIEW_TARGETS_AXIS 取,both 是原始分前 TOP_MAIN ∪ 新序前 TOP_MAIN_RANK)加观察档前
|
||||||
与方案说的"一两百家"一致;07:10 计划默认只展示 20 加 10,那份太窄。"""
|
TOP_OBS 只(快照里的 main / observe 已是新序全部行,每行带 score 与 rank),约一百二十到一百五十家。"""
|
||||||
rows = build_rows(ds, (snap.get("main") or [])[:TOP_MAIN], (snap.get("observe") or [])[:TOP_OBS])
|
main_sel = pick_main_targets(snap.get("main") or [], config.REVIEW_TARGETS_AXIS, TOP_MAIN, TOP_MAIN_RANK)
|
||||||
|
rows = build_rows(ds, main_sel, (snap.get("observe") or [])[:TOP_OBS])
|
||||||
(write or save)(ds, rows)
|
(write or save)(ds, rows)
|
||||||
n_main = sum(1 for r in rows if r["source"] == "main")
|
n_main = sum(1 for r in rows if r["source"] == "main")
|
||||||
print(f" 评析目标名单 {ds}:主榜 {n_main},观察 {len(rows) - n_main},写入 {TABLE}")
|
print(f" 评析目标名单 {ds}:主榜 {n_main},观察 {len(rows) - n_main},写入 {TABLE}")
|
||||||
|
|
|
||||||
|
|
@ -104,8 +104,9 @@ def test_wiring():
|
||||||
print("[不变式四的接线:collect 两序列都重排、_row 落 score_adj]")
|
print("[不变式四的接线:collect 两序列都重排、_row 落 score_adj]")
|
||||||
src = open(os.path.join(os.path.dirname(os.path.abspath(__file__)), "plan.py"),
|
src = open(os.path.join(os.path.dirname(os.path.abspath(__file__)), "plan.py"),
|
||||||
encoding="utf-8").read()
|
encoding="utf-8").read()
|
||||||
i_main = src.find("main = _rerank_by_quality(main, cards,")
|
# 2026-09-17:调用点改成先算旧序拿 rank_old(main_old/obs_old),再按轴分派算新序,匹配串随之更新。
|
||||||
i_obs = src.find("obs = _rerank_by_quality(obs, cards,")
|
i_main = src.find("main_old = _rerank_by_quality(main, cards,")
|
||||||
|
i_obs = src.find("obs_old = _rerank_by_quality(obs, cards,")
|
||||||
t("主榜序列被重排", i_main > 0)
|
t("主榜序列被重排", i_main > 0)
|
||||||
t("观察档序列被重排", i_obs > 0)
|
t("观察档序列被重排", i_obs > 0)
|
||||||
# 重排必须在 _pick 与 _full_rows 定义之前——四个出口才同序
|
# 重排必须在 _pick 与 _full_rows 定义之前——四个出口才同序
|
||||||
|
|
|
||||||
|
|
@ -0,0 +1,205 @@
|
||||||
|
"""选股计划排序轴换成公司质地(2026-09-17):传导降为雷达门与带内次级键。
|
||||||
|
|
||||||
|
离线,不连库不起服务。跑法:python3 test_plan_rank_axis.py,预期最后一行是 ALL OK。
|
||||||
|
逐字回旧那一例是 case1:四个开关都设旧值时,装配重排输出与 09-14 的 _rerank_by_quality 逐字节相同。
|
||||||
|
附录丙十例一一对应本文件的十个 test_*。
|
||||||
|
"""
|
||||||
|
import os
|
||||||
|
import sys
|
||||||
|
import types
|
||||||
|
|
||||||
|
# 离线不连库:装不上的驱动打空壳,pandas 本机装得上就用真的(重排要在真 Series 上验)。
|
||||||
|
for _n in ("pandas", "pymysql", "psycopg", "pymongo"):
|
||||||
|
if _n not in sys.modules:
|
||||||
|
try:
|
||||||
|
__import__(_n)
|
||||||
|
except Exception: # noqa: BLE001
|
||||||
|
_m = types.ModuleType(_n)
|
||||||
|
if _n == "pandas":
|
||||||
|
_m.DataFrame = type("DataFrame", (), {})
|
||||||
|
sys.modules[_n] = _m
|
||||||
|
|
||||||
|
import pandas as pd # noqa: E402 —— 重排就在 pd.Series 上做,直接用真的 pandas
|
||||||
|
|
||||||
|
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
|
||||||
|
|
||||||
|
import card # noqa: E402
|
||||||
|
import config # noqa: E402
|
||||||
|
import plan # noqa: E402
|
||||||
|
import pool # noqa: E402
|
||||||
|
import review_targets # noqa: E402
|
||||||
|
|
||||||
|
FAILED = []
|
||||||
|
|
||||||
|
|
||||||
|
def t(name, cond, extra=""):
|
||||||
|
print((" ok " if cond else " FAIL ") + name + ((" " + str(extra)) if not cond else ""))
|
||||||
|
if not cond:
|
||||||
|
FAILED.append(name)
|
||||||
|
|
||||||
|
|
||||||
|
def _cr(overall, age_days=None):
|
||||||
|
if overall is None:
|
||||||
|
return None
|
||||||
|
d = {"overall": overall}
|
||||||
|
if age_days is not None:
|
||||||
|
d["age_days"] = age_days
|
||||||
|
return d
|
||||||
|
|
||||||
|
|
||||||
|
def _cards(spec: dict) -> dict:
|
||||||
|
"""spec:{代码: (质地档, 被指向[, 报告日龄])}。装成 collect 里 cards 的形状(只放本改动要读的两路)。"""
|
||||||
|
out = {}
|
||||||
|
for k, v in spec.items():
|
||||||
|
overall, pointed = v[0], v[1]
|
||||||
|
age = v[2] if len(v) > 2 else None
|
||||||
|
out[k] = {"company_review": _cr(overall, age), "gates": {"pointed": pointed}}
|
||||||
|
return out
|
||||||
|
|
||||||
|
|
||||||
|
def test_1_replay_transmission():
|
||||||
|
print("[case1 四开关设旧值:逐字回旧到 _rerank_by_quality]")
|
||||||
|
ranked = pd.Series({"A": 242.0, "B": 238.0, "C": 222.0, "D": 218.0})
|
||||||
|
cards = _cards({"A": ("差", True), "B": ("好", False), "C": ("好", True), "D": ("差", False)})
|
||||||
|
old = config.PLAN_RANK_AXIS
|
||||||
|
try:
|
||||||
|
config.PLAN_RANK_AXIS = "transmission"
|
||||||
|
out = plan._rerank(ranked, cards, True)
|
||||||
|
finally:
|
||||||
|
config.PLAN_RANK_AXIS = old
|
||||||
|
ref = plan._rerank_by_quality(ranked, cards, config.PLAN_QUALITY_BONUS)
|
||||||
|
t("序列顺序逐字相同", list(out.index) == list(ref.index), list(out.index))
|
||||||
|
t("分值逐字相同", list(out.values) == list(ref.values))
|
||||||
|
|
||||||
|
|
||||||
|
def test_2_bands_order():
|
||||||
|
print("[case2 质地带:好 > 中 > 无评析 > 差]")
|
||||||
|
ranked = pd.Series({"g": 210.0, "m": 210.0, "n": 210.0, "b": 210.0})
|
||||||
|
cards = _cards({"g": ("好", False), "m": ("中", False), "n": (None, False), "b": ("差", False)})
|
||||||
|
out = plan._rerank_by_axis(ranked, cards, "quality", 120, True)
|
||||||
|
t("好中无差", list(out.index) == ["g", "m", "n", "b"], list(out.index))
|
||||||
|
|
||||||
|
|
||||||
|
def test_3_within_band():
|
||||||
|
print("[case3 同带内:被指向优先,再按组内分]")
|
||||||
|
ranked = pd.Series({"p_lo": 213.0, "p_hi": 217.0, "np_hi": 218.0})
|
||||||
|
cards = _cards({"p_lo": ("好", True), "p_hi": ("好", True), "np_hi": ("好", False)})
|
||||||
|
out = plan._rerank_by_axis(ranked, cards, "quality", 120, True)
|
||||||
|
t("被指向优先且带内按组内分", list(out.index) == ["p_hi", "p_lo", "np_hi"], list(out.index))
|
||||||
|
|
||||||
|
|
||||||
|
def test_4_deterministic():
|
||||||
|
print("[case4 确定性:两次相同、末位键代码升序]")
|
||||||
|
ranked = pd.Series({"B": 210.0, "A": 210.0})
|
||||||
|
cards = _cards({"A": ("好", True), "B": ("好", True)})
|
||||||
|
o1 = list(plan._rerank_by_axis(ranked, cards, "quality", 120, True).index)
|
||||||
|
o2 = list(plan._rerank_by_axis(ranked, cards, "quality", 120, True).index)
|
||||||
|
t("两次装配顺序相同", o1 == o2, o1)
|
||||||
|
t("末位键代码升序", o1 == ["A", "B"], o1)
|
||||||
|
|
||||||
|
|
||||||
|
def test_5_stale_demotes():
|
||||||
|
print("[case5 超期一百二十天的好公司落到第一带]")
|
||||||
|
ranked = pd.Series({"fresh": 210.0, "mid": 210.0, "stale": 210.0})
|
||||||
|
cards = _cards({"fresh": ("好", False, 10), "mid": ("中", False, 10), "stale": ("好", False, 200)})
|
||||||
|
out = plan._rerank_by_axis(ranked, cards, "quality", 120, True)
|
||||||
|
t("超期好票落到无评析带(排到中之后)", list(out.index) == ["fresh", "mid", "stale"], list(out.index))
|
||||||
|
|
||||||
|
|
||||||
|
def test_6_pool_rank_source():
|
||||||
|
print("[case6 入池 rank:只取被指向的,数量等于 top,按新序]")
|
||||||
|
main_rows = [
|
||||||
|
{"code": "a", "pointed": True, "tier": "强传导", "rank_old": 3},
|
||||||
|
{"code": "b", "pointed": False, "tier": "强传导", "rank_old": 1},
|
||||||
|
{"code": "c", "pointed": True, "tier": "弱传导", "rank_old": 5},
|
||||||
|
{"code": "d", "pointed": True, "tier": "无传导", "rank_old": 2},
|
||||||
|
]
|
||||||
|
rows = pool.select_plan_rows(main_rows, 2, "rank", set())
|
||||||
|
t("只取被指向", all(r.get("pointed") for r in rows), [r["code"] for r in rows])
|
||||||
|
t("数量等于 top", len(rows) == 2, len(rows))
|
||||||
|
t("按主榜新序取前二", [r["code"] for r in rows] == ["a", "c"], [r["code"] for r in rows])
|
||||||
|
t("_pool_source 记 rank_top", all(r["_pool_source"] == "rank_top" for r in rows))
|
||||||
|
|
||||||
|
|
||||||
|
def test_7_review_targets_union():
|
||||||
|
print("[case7 评析名单两路并集:去重、按新序]")
|
||||||
|
main = [
|
||||||
|
{"code": "600001.SH", "name": "x1", "score": 300, "rank": 1},
|
||||||
|
{"code": "600002.SH", "name": "x2", "score": 260, "rank": 2},
|
||||||
|
{"code": "600003.SH", "name": "x3", "score": 320, "rank": 3},
|
||||||
|
]
|
||||||
|
both = review_targets.pick_main_targets(main, "both", 2, 1)
|
||||||
|
codes = [r["code"] for r in both]
|
||||||
|
# score 降序前二:x3(320)、x1(300);rank 升序前一:x1(rank1)。并集去重={x1,x3},按 rank 排 → x1,x3
|
||||||
|
t("并集去重", set(codes) == {"600001.SH", "600003.SH"}, codes)
|
||||||
|
t("按新序(rank)排", codes == ["600001.SH", "600003.SH"], codes)
|
||||||
|
only = review_targets.pick_main_targets(main, "score", 2, 1)
|
||||||
|
t("score 轴逐字回旧(score 降序前二)", [r["code"] for r in only] == ["600003.SH", "600001.SH"],
|
||||||
|
[r["code"] for r in only])
|
||||||
|
|
||||||
|
|
||||||
|
def test_8_card_sort_toggle():
|
||||||
|
print("[case8 卡内序开关:开则质地优先,关则回先判决再涨幅]")
|
||||||
|
rows = [
|
||||||
|
{"verdict": "候选", "quality_band": 0, "pct0": 5.0},
|
||||||
|
{"verdict": "候选", "quality_band": 3, "pct0": 1.0},
|
||||||
|
]
|
||||||
|
old = config.CARD_SORT_BY_QUALITY
|
||||||
|
try:
|
||||||
|
config.CARD_SORT_BY_QUALITY = True
|
||||||
|
on = sorted(rows, key=card.sort_key)
|
||||||
|
t("开:质地好在前", on[0]["quality_band"] == 3, on[0])
|
||||||
|
config.CARD_SORT_BY_QUALITY = False
|
||||||
|
off = sorted(rows, key=card.sort_key)
|
||||||
|
t("关:涨幅高在前(逐字回旧)", off[0]["pct0"] == 5.0, off[0])
|
||||||
|
finally:
|
||||||
|
config.CARD_SORT_BY_QUALITY = old
|
||||||
|
|
||||||
|
|
||||||
|
def test_9_theme_cap_exemption():
|
||||||
|
print("[case9 主题限额:质地轴下无传导票免限额,有主题仍每主题五只]")
|
||||||
|
no_theme = [{"code": f"n{i}", "evidence": None, "rank_axis": "quality"} for i in range(6)]
|
||||||
|
t("无传导票免五只限额", len(plan.pick_rows(no_theme, 20, 5, uncapped_no_theme=True)) == 6)
|
||||||
|
t("uncapped 为假时无传导仍只进五只", len(plan.pick_rows(no_theme, 20, 5, uncapped_no_theme=False)) == 5)
|
||||||
|
themed = [{"code": f"t{i}", "evidence": {"theme": "锂电"}, "rank_axis": "quality"} for i in range(6)]
|
||||||
|
t("有主题仍受每主题五只", len(plan.pick_rows(themed, 20, 5, uncapped_no_theme=True)) == 5)
|
||||||
|
|
||||||
|
|
||||||
|
def test_10_pool_keep_old():
|
||||||
|
print("[case10 备用开关 keep_old:并进旧序强传导前 top,去重,受 POOL_MAX 兜底]")
|
||||||
|
main_rows = [
|
||||||
|
{"code": "a", "pointed": True, "tier": "强传导", "rank_old": 3},
|
||||||
|
{"code": "b", "pointed": False, "tier": "强传导", "rank_old": 1},
|
||||||
|
{"code": "e", "pointed": False, "tier": "强传导", "rank_old": 2},
|
||||||
|
]
|
||||||
|
rows = pool.select_plan_rows(main_rows, 2, "rank", {"强传导"}, keep_old=True)
|
||||||
|
codes = [r["code"] for r in rows]
|
||||||
|
t("被指向的 a 在", "a" in codes, codes)
|
||||||
|
t("旧序强传导前二并进(b、e)", "b" in codes and "e" in codes, codes)
|
||||||
|
t("同票不重复", len(codes) == len(set(codes)), codes)
|
||||||
|
old_max = config.POOL_MAX
|
||||||
|
try:
|
||||||
|
config.POOL_MAX = 2
|
||||||
|
rows2 = pool.select_plan_rows(main_rows, 2, "rank", {"强传导"}, keep_old=True)
|
||||||
|
t("POOL_MAX 兜底", len(rows2) <= 2, len(rows2))
|
||||||
|
finally:
|
||||||
|
config.POOL_MAX = old_max
|
||||||
|
|
||||||
|
|
||||||
|
def main():
|
||||||
|
for fn in (test_1_replay_transmission, test_2_bands_order, test_3_within_band,
|
||||||
|
test_4_deterministic, test_5_stale_demotes, test_6_pool_rank_source,
|
||||||
|
test_7_review_targets_union, test_8_card_sort_toggle, test_9_theme_cap_exemption,
|
||||||
|
test_10_pool_keep_old):
|
||||||
|
fn()
|
||||||
|
print("-" * 70)
|
||||||
|
if FAILED:
|
||||||
|
print("FAILED %d: %s" % (len(FAILED), "; ".join(FAILED)))
|
||||||
|
return 1
|
||||||
|
print("ALL OK — 排序轴换质地:逐字回旧 / 四带序 / 带内次级键 / 确定性 / 超期守卫 / "
|
||||||
|
"入池 rank / 评析并集 / 卡内序 / 主题限额例外 / 备用开关")
|
||||||
|
return 0
|
||||||
|
|
||||||
|
|
||||||
|
if __name__ == "__main__":
|
||||||
|
sys.exit(main())
|
||||||
|
|
@ -50,6 +50,12 @@ def test_pick_rules():
|
||||||
t("没有主题的行归到同一个桶,不当成各自独立",
|
t("没有主题的行归到同一个桶,不当成各自独立",
|
||||||
[r["code"] for r in plan.pick_rows(
|
[r["code"] for r in plan.pick_rows(
|
||||||
[_row("X", None, 1), _row("Y", None, 2), _row("Z", "光伏", 3)], 3, 1)] == ["X", "Z"])
|
[_row("X", None, 1), _row("Y", None, 2), _row("Z", "光伏", 3)], 3, 1)] == ["X", "Z"])
|
||||||
|
# 质地轴对照(2026-09-17):同样两只无主题票,行带 rank_axis=quality 时免主题限额,两只都进
|
||||||
|
t("质地轴下无主题行不受主题限额(对照上一条)",
|
||||||
|
[r["code"] for r in plan.pick_rows(
|
||||||
|
[dict(_row("X", None, 1), rank_axis="quality"),
|
||||||
|
dict(_row("Y", None, 2), rank_axis="quality"),
|
||||||
|
dict(_row("Z", "光伏", 3), rank_axis="quality")], 3, 1)] == ["X", "Y", "Z"])
|
||||||
t("要的条数比总行数多就全给", len(plan.pick_rows(rows, 99, 0)) == 5)
|
t("要的条数比总行数多就全给", len(plan.pick_rows(rows, 99, 0)) == 5)
|
||||||
t("要零条时先给一条(照抄原实现的边界:先放进去再判断够没够,不是我们要它这样)",
|
t("要零条时先给一条(照抄原实现的边界:先放进去再判断够没够,不是我们要它这样)",
|
||||||
len(plan.pick_rows(rows, 0, 0)) == 1)
|
len(plan.pick_rows(rows, 0, 0)) == 1)
|
||||||
|
|
|
||||||
Loading…
Reference in New Issue