mirror of
https://ghfast.top/https://github.com/aeroxw/tick-stock-panel.git
synced 2026-09-12 14:24:15 +08:00
shy3130
e0cd625ef4
- 因子平台: /factors 一级页(检验/因子库/编辑器/组合/挖掘), DSL 公式因子(25 算子点选、双语字段、我的因子模板、脏公式守卫), 版本与生命周期, 自动挖掘 L1 统计筛选 - 因子↔策略四条桥: 触发器 Zap 快建因子条件信号、因子一键生成排名策略、自定义信号 AI 提示词接入因子分组、策略回测因子归因(胜/败单入场信号日因子均值, 独立 tab, 双语因子名) - 回测: 统计卡新增盈亏比(≥1 红/<1 绿), 蒙卡回撤合并为中位/95% 双值卡(自适应字号), 高级设置基础过滤与策略编辑器参数对齐(5 组区间) - 信号库独立页 /signals(原设置 tab 迁出), 持仓提醒入导航; 挖掘并入因子页第 5 tab, /mining 旧链接重定向 - 研究线配套: 因子目录 61→77(评分/矩阵双内核), stats_v2(Newey-West/BH-FDR/DSR), enriched 管道与异动/报价服务配套调整 - 文档: README 导航与特性表、features.md 因子平台章节、操作说明书 9.2、factor-platform-plan 执行状态与 §5、二开文档桥接说明; 交流与支持节改版 - 版本 0.2.2 → 0.2.3; 后端全量 1625 passed(1 例环境性跳过), 前端 build 通过
320 lines
12 KiB
Python
320 lines
12 KiB
Python
"""自定义信号 — 用户用「字段 + 运算符 + 值」组合出的布尔信号。
|
||
|
||
职责:
|
||
- 从 data/user_data/custom_signals/*.json 加载信号定义
|
||
- 把每个信号的 conditions 编译成一条 Polars 布尔表达式（AND 组合）
|
||
- 供 pipeline 在 compute_signals / compute_enriched_today 末尾注入为列
|
||
|
||
不知道: 引擎、AI、API、回测、监控。纯函数 + 模块级缓存。
|
||
|
||
设计:
|
||
- 信号列名加前缀 ``csg_`` 避免与内置 ``signal_`` 列冲突。
|
||
- 回测/选股/监控都按列名找信号，因此注入列后零特殊处理即可三处生效。
|
||
- 字段白名单 + 固定运算符集，杜绝任意表达式注入。
|
||
- 第一版只支持 AND（多条件同时满足）。
|
||
"""
|
||
from __future__ import annotations
|
||
|
||
import json
|
||
import logging
|
||
import re
|
||
from pathlib import Path
|
||
|
||
import polars as pl
|
||
|
||
logger = logging.getLogger(__name__)
|
||
|
||
# ── 常量 ────────────────────────────────────────────────
|
||
PREFIX = "csg_" # 自定义信号列名前缀
|
||
ID_RE = re.compile(r"^[a-z0-9_]{1,40}$")
|
||
OPS = {">", ">=", "<", "<=", "==", "!="}
|
||
|
||
# 字段白名单：只允许这些列出现在条件里（防注入）。均为数值型。
|
||
# 与 ENRICHED_COLUMNS 的数值列保持一致，排除 symbol/date/name 等非数值列。
|
||
ALLOWED_FIELDS: frozenset[str] = frozenset({
|
||
# 行情
|
||
"open", "high", "low", "close", "volume", "amount", "turnover_rate",
|
||
"consecutive_limit_ups", "consecutive_limit_downs",
|
||
# 基础
|
||
"prev_close", "change_pct", "change_amount", "amplitude",
|
||
# 均线 / 指数均线
|
||
"ma5", "ma10", "ma20", "ma30", "ma60",
|
||
"ema5", "ema10", "ema20", "ema30", "ema60",
|
||
# MACD / BOLL / KDJ / ATR
|
||
"macd_dif", "macd_dea", "macd_hist",
|
||
"boll_upper", "boll_lower",
|
||
"kdj_k", "kdj_d", "kdj_j",
|
||
"atr_14",
|
||
# 量价 / 极值 / 动量 / 波动率 / RSI
|
||
"vol_ma5", "vol_ma10", "vol_ratio_5d",
|
||
"high_60d", "low_60d",
|
||
"momentum_5d", "momentum_10d", "momentum_20d", "momentum_30d", "momentum_60d",
|
||
"annual_vol_20d",
|
||
"rsi_6", "rsi_14", "rsi_24",
|
||
# 异动偏离 (交易所异动规则口径, 运行时列)
|
||
"deviate_3d", "deviate_10d", "deviate_30d",
|
||
})
|
||
|
||
# 运算符 → Polars 表达式构造器（输入 col_expr, value）
|
||
_OP_BUILDERS = {
|
||
">": lambda c, v: c > v,
|
||
">=": lambda c, v: c >= v,
|
||
"<": lambda c, v: c < v,
|
||
"<=": lambda c, v: c <= v,
|
||
"==": lambda c, v: c == v,
|
||
"!=": lambda c, v: c != v,
|
||
}
|
||
|
||
|
||
def allowed_fields() -> frozenset[str]:
|
||
"""条件可引用字段 = 物化列白名单 并入 注册表因子 (虚拟/自定义/复合)。
|
||
|
||
因子列在历史路径 (compute_signals) 由 materialize_factor_columns 复用
|
||
评分物化管线补算; 盘中单日快照无滚动窗口, 依赖因子的信号被 inject 以
|
||
缺列告警跳过 (与日期偏移条件同样的优雅降级)。
|
||
"""
|
||
from app.factors.registry import all_factors
|
||
|
||
return frozenset(ALLOWED_FIELDS | {spec.id for spec in all_factors()})
|
||
|
||
|
||
def materialize_factor_columns(
|
||
df: pl.DataFrame,
|
||
exprs: dict[str, pl.Expr],
|
||
needed: set[str] | None = None,
|
||
) -> pl.DataFrame:
|
||
"""把信号表达式引用、且 df 缺失的注册表因子列补算出来。
|
||
|
||
复用评分物化路径 (materialize_scoring_columns) — 与检验/评分同一条计算
|
||
逻辑, 不引入第二套实现。非注册表列不在此处理 (缺列仍由 inject 告警跳过)。
|
||
"""
|
||
if df.is_empty() or not exprs:
|
||
return df
|
||
cols = set(df.columns)
|
||
missing: set[str] = set()
|
||
for name, roots in expression_dependencies(exprs).items():
|
||
if needed is not None and name not in needed:
|
||
continue
|
||
missing.update(root for root in roots if root not in cols)
|
||
if not missing:
|
||
return df
|
||
from app.factors.registry import all_factors
|
||
|
||
factor_ids = {spec.id for spec in all_factors()}
|
||
to_compute = missing & factor_ids
|
||
if not to_compute:
|
||
return df
|
||
from app.strategy.scoring import materialize_scoring_columns
|
||
|
||
return materialize_scoring_columns(df, sorted(to_compute))
|
||
|
||
|
||
# ── 持久化（镜像 strategy/config.py 的写法）──────────────
|
||
def _dir(data_dir: Path) -> Path:
|
||
d = data_dir / "user_data" / "custom_signals"
|
||
d.mkdir(parents=True, exist_ok=True)
|
||
return d
|
||
|
||
|
||
def _path(data_dir: Path, signal_id: str) -> Path:
|
||
return _dir(data_dir) / f"{signal_id}.json"
|
||
|
||
|
||
def load_all(data_dir: Path) -> list[dict]:
|
||
"""读取全部自定义信号定义。损坏的文件被跳过。"""
|
||
d = _dir(data_dir)
|
||
out: list[dict] = []
|
||
for f in sorted(d.glob("*.json")):
|
||
try:
|
||
out.append(json.loads(f.read_text(encoding="utf-8")))
|
||
except Exception as e:
|
||
logger.warning("custom signal load failed %s: %s", f.name, e)
|
||
return out
|
||
|
||
|
||
def save_one(data_dir: Path, sig: dict) -> None:
|
||
p = _path(data_dir, sig["id"])
|
||
p.parent.mkdir(parents=True, exist_ok=True)
|
||
p.write_text(json.dumps(sig, ensure_ascii=False, indent=2), encoding="utf-8")
|
||
|
||
|
||
def delete_one(data_dir: Path, signal_id: str) -> bool:
|
||
p = _path(data_dir, signal_id)
|
||
if p.exists():
|
||
p.unlink()
|
||
return True
|
||
return False
|
||
|
||
|
||
# ── 校验 ────────────────────────────────────────────────
|
||
|
||
MAX_DAYS = 60 # 偏移天数上限 (前N日的 N)
|
||
|
||
|
||
def _parse_days(c: dict, key: str, i: int) -> int:
|
||
"""解析并校验条件的天数偏移 (leftDays / rightDays)。返回 0..MAX_DAYS。"""
|
||
raw = c.get(key, 0)
|
||
try:
|
||
n = int(raw)
|
||
except (TypeError, ValueError):
|
||
raise ValueError(f"第 {i+1} 个条件: {key} 必须是整数: {raw!r}")
|
||
if n < 0 or n > MAX_DAYS:
|
||
raise ValueError(f"第 {i+1} 个条件: {key} 必须在 0..{MAX_DAYS} 之间: {n}")
|
||
return n
|
||
|
||
|
||
def _parse_right(right: str) -> tuple[str, object]:
|
||
"""解析右值。返回 ('field', colname) 或 ('const', float)。
|
||
|
||
接受三种形式:
|
||
- 数字 (int / float / 数字字符串) → 常量
|
||
- "field:字段名" → 字段引用
|
||
- 裸字段名 (在白名单内) → 自动视为字段引用
|
||
(AI 生成偶尔漏写 field: 前缀; 白名单字段名不可能是数字, 无歧义)
|
||
"""
|
||
if isinstance(right, (int, float)):
|
||
return ("const", float(right))
|
||
if not isinstance(right, str):
|
||
raise ValueError(f"非法右值: {right!r}")
|
||
allowed = allowed_fields()
|
||
if right.startswith("field:"):
|
||
col = right[len("field:"):]
|
||
if col not in allowed:
|
||
raise ValueError(f"右值字段不在白名单: {col}")
|
||
return ("field", col)
|
||
# 纯数字
|
||
try:
|
||
return ("const", float(right))
|
||
except ValueError:
|
||
pass
|
||
# 裸字段名 — 兜底容错, 仍受白名单约束
|
||
if right in allowed:
|
||
return ("field", right)
|
||
raise ValueError(f"非法右值（应为 field:xxx 或数字）: {right!r}")
|
||
|
||
|
||
def validate(sig: dict) -> None:
|
||
"""校验一个信号定义，非法则抛 ValueError（含中文信息）。"""
|
||
sid = sig.get("id", "")
|
||
if not isinstance(sid, str) or not ID_RE.match(sid):
|
||
raise ValueError(f"信号 id 非法（仅小写字母数字下划线，1-40字符）: {sid!r}")
|
||
if not isinstance(sig.get("name"), str) or not sig["name"].strip():
|
||
raise ValueError("信号 name 不能为空")
|
||
if sig.get("kind") not in ("entry", "exit", "both"):
|
||
raise ValueError("kind 必须是 entry / exit / both")
|
||
conds = sig.get("conditions")
|
||
if not isinstance(conds, list) or len(conds) == 0:
|
||
raise ValueError("conditions 不能为空")
|
||
if len(conds) > 8:
|
||
raise ValueError("conditions 最多 8 条")
|
||
for i, c in enumerate(conds):
|
||
if not isinstance(c, dict):
|
||
raise ValueError(f"第 {i+1} 个条件格式错误")
|
||
left = c.get("left", "")
|
||
if left not in allowed_fields():
|
||
raise ValueError(f"第 {i+1} 个条件: 字段 {left!r} 不在白名单")
|
||
if c.get("op") not in OPS:
|
||
raise ValueError(f"第 {i+1} 个条件: 运算符 {c.get('op')!r} 非法")
|
||
_parse_right(c.get("right")) # 会校验右值字段/数字
|
||
_parse_days(c, "leftDays", i) # 左字段偏移
|
||
_parse_days(c, "rightDays", i) # 右字段偏移
|
||
|
||
|
||
# ── 编译为 Polars 表达式 ─────────────────────────────────
|
||
def column_name(signal_id: str) -> str:
|
||
"""信号 id → DataFrame 列名（加前缀）。"""
|
||
return f"{PREFIX}{signal_id}"
|
||
|
||
|
||
def _col(name: str, days: int = 0) -> pl.Expr:
|
||
"""构造列表达式; days>0 时取 N 个交易日前的值 (按 symbol 分组 shift)。"""
|
||
expr = pl.col(name)
|
||
if days > 0:
|
||
expr = expr.shift(days).over("symbol")
|
||
return expr
|
||
|
||
|
||
def build_expressions(signals: list[dict], allow_shift: bool = True) -> dict[str, pl.Expr]:
|
||
"""把多个自定义信号编译成 {column_name: pl.Expr}。
|
||
|
||
- 只处理 enabled != False 的信号。
|
||
- 单个信号内多条件用 ``&`` 串联（AND）。
|
||
- allow_shift=False 时, 跳过带日期偏移 (leftDays/rightDays>0) 的信号
|
||
(盘中单日快照上 .shift 跨 symbol 语义不正确, 优雅降级)。
|
||
- 编译失败的信号被跳过并告警（不影响其它信号）。
|
||
"""
|
||
out: dict[str, pl.Expr] = {}
|
||
for sig in signals:
|
||
if sig.get("enabled") is False:
|
||
continue
|
||
try:
|
||
conds = sig["conditions"]
|
||
col_name = column_name(sig["id"])
|
||
parts: list[pl.Expr] = []
|
||
for c in conds:
|
||
left_days = int(c.get("leftDays", 0) or 0)
|
||
right_days = int(c.get("rightDays", 0) or 0)
|
||
# 盘中路径不支持偏移 → 跳过整个信号
|
||
if not allow_shift and (left_days > 0 or right_days > 0):
|
||
raise ValueError("盘中实时路径不支持日期偏移条件, 已跳过")
|
||
left = c["left"]
|
||
op = c["op"]
|
||
kind, val = _parse_right(c["right"])
|
||
right_expr = _col(val, right_days) if kind == "field" else val
|
||
parts.append(_OP_BUILDERS[op](_col(left, left_days), right_expr))
|
||
combined = parts[0]
|
||
for p in parts[1:]:
|
||
combined = combined & p
|
||
out[col_name] = combined
|
||
except Exception as e:
|
||
logger.warning("custom signal compile failed %s: %s", sig.get("id"), e)
|
||
return out
|
||
|
||
|
||
def expression_dependencies(exprs: dict[str, pl.Expr] | None = None) -> dict[str, frozenset[str]]:
|
||
"""返回自定义信号列到根字段的依赖映射。"""
|
||
source = exprs if exprs is not None else {}
|
||
return {name: frozenset(_expr_root_columns(expr)) for name, expr in source.items()}
|
||
|
||
|
||
def inject(
|
||
df: pl.DataFrame,
|
||
exprs: dict[str, pl.Expr],
|
||
needed: set[str] | None = None,
|
||
) -> pl.DataFrame:
|
||
"""把编译好的信号表达式作为列加入 df。
|
||
|
||
``needed=None`` 保持历史全量语义；传入集合时只注入被请求的自定义信号。
|
||
缺失依赖会明确告警，避免回测静默丢失信号。
|
||
"""
|
||
if df.is_empty() or not exprs:
|
||
return df
|
||
cols = set(df.columns)
|
||
add: dict[str, pl.Expr] = {}
|
||
for name, expr in exprs.items():
|
||
if needed is not None and name not in needed:
|
||
continue
|
||
# 提取该表达式引用的所有字段列，缺失则跳过（避免运行时报错）
|
||
required = _expr_root_columns(expr)
|
||
if required.issubset(cols):
|
||
add[name] = expr
|
||
else:
|
||
logger.warning(
|
||
"custom signal %s missing dependencies: %s",
|
||
name,
|
||
sorted(required - cols),
|
||
)
|
||
if add:
|
||
df = df.with_columns([e.alias(n) for n, e in add.items()])
|
||
return df
|
||
|
||
|
||
def _expr_root_columns(expr: pl.Expr) -> set[str]:
|
||
"""尽力提取表达式里出现的列名。失败则返回空集（保守跳过）。"""
|
||
try:
|
||
# Polars 的 meta.root_names() 返回表达式引用的根列名
|
||
names = expr.meta.root_names()
|
||
return set(names)
|
||
except Exception:
|
||
return set()
|