mirror of
https://github.com/shy3130/tick-stock-panel.git
synced 2026-11-02 10:26:13 +00:00
fix(backtest): 矩阵回测接入扩展时序数据
扩展数值列按设计不落 enriched parquet (只存 14 列), 此前矩阵回测引用 ext_* 字段直接报 fields unavailable — 叠加条件/评分引用扩展数据的策略 无法回测, 扩展数据失去回测意义。 修复: 引擎在矩阵加载后按矩阵自身 (日期×标的) 骨架现场注入时序模式数值 扩展字段 (attach_matrix_ext_fields, 与策略页帧路径同一台 PIT 安全 join, 缺失补空值=条件不命中), 与财务因子 attach 同构; parquet 矩阵缓存不受 影响。快照模式与字符串扩展字段仍拒绝 (快照回看历史=未来函数), 引用时 前置校验给出明确报错。 实测: 隔夜溢价精选·迭代优化 + 人气排名 20~100 叠加条件, 回测 1.1s 跑通, 交易数 29→22 证明过滤真实生效; 新增 7 个单测 (对齐/NaN/拦截/叠加联动/ 端到端正反例), 全量回归 3131 通过。 Signed-off-by: shy3130 <shy3130@qq.com>
This commit is contained in:
shy3130
@@ -520,6 +520,22 @@ class BacktestEngine:
|
||||
| set(feature_plan.instrument_columns)
|
||||
| set(feature_plan.matrix_columns)
|
||||
)
|
||||
# 扩展时序数值字段不进 parquet 解析 (enriched 只落 14 存储列), 矩阵构建后
|
||||
# 按 (日期, 标的) 对齐现场注入 — 与财务因子 attach 同构。快照模式/字符串
|
||||
# 字段/未注册名在此拦截 (快照回看历史 = 未来函数)。
|
||||
from app.factors.ext_factors import numeric_timeseries_ext_columns
|
||||
|
||||
ext_wanted = frozenset(c for c in field_columns if c.startswith("ext_"))
|
||||
if ext_wanted:
|
||||
invalid_ext = ext_wanted - numeric_timeseries_ext_columns(
|
||||
self.repo.store.data_dir
|
||||
)
|
||||
if invalid_ext:
|
||||
raise ValueError(
|
||||
f"扩展字段不可用于矩阵回测 (需为时序模式的数值扩展字段): "
|
||||
f"{sorted(invalid_ext)}"
|
||||
)
|
||||
field_columns = field_columns - ext_wanted
|
||||
cache_root = (
|
||||
self.repo.store.data_dir / ".backtest_matrix_cache"
|
||||
if settings.backtest_matrix_disk_cache_enabled
|
||||
@@ -530,6 +546,7 @@ class BacktestEngine:
|
||||
if cache_profile is not None
|
||||
else frozenset(field_columns)
|
||||
)
|
||||
cache_fields = frozenset(cache_fields) - ext_wanted
|
||||
cache_max_bytes = (
|
||||
cache_profile.max_disk_bytes
|
||||
if cache_profile is not None
|
||||
@@ -575,6 +592,14 @@ class BacktestEngine:
|
||||
self.repo.store.data_dir if self.repo is not None else None,
|
||||
fundamental_names,
|
||||
)
|
||||
if ext_wanted:
|
||||
from app.factors.ext_factors import attach_matrix_ext_fields
|
||||
|
||||
market = attach_matrix_ext_fields(
|
||||
market,
|
||||
self.repo.store.data_dir if self.repo is not None else None,
|
||||
ext_wanted,
|
||||
)
|
||||
return market
|
||||
except EnrichedGenerationUnavailableError:
|
||||
if attempt + 1 >= attempts:
|
||||
|
||||
@@ -28,6 +28,7 @@ import contextlib
|
||||
import logging
|
||||
import re
|
||||
from pathlib import Path
|
||||
from typing import Any
|
||||
|
||||
import polars as pl
|
||||
|
||||
@@ -76,6 +77,21 @@ def _numeric_fields(config) -> list:
|
||||
return [f for f in config.fields if f.dtype in _NUMERIC_DTYPES]
|
||||
|
||||
|
||||
def numeric_timeseries_ext_columns(data_dir: Path | None = None) -> frozenset[str]:
|
||||
"""时序模式数值扩展字段的列名集合 — 回测矩阵可注入的白名单。
|
||||
|
||||
快照模式 (mode=snapshot) 刻意排除: 快照是"最新值", 回看历史注入即未来函数,
|
||||
帧路径的 include_snapshot=False 同理只放行时序。
|
||||
"""
|
||||
root = _resolve_dir(data_dir)
|
||||
return frozenset(
|
||||
ext_column_name(cfg.id, f.name)
|
||||
for cfg in _load_configs(root)
|
||||
if cfg.mode == "timeseries"
|
||||
for f in _numeric_fields(cfg)
|
||||
)
|
||||
|
||||
|
||||
def _signal_fields(config) -> list:
|
||||
"""帧 join / 信号条件可用的字段 (数值 + 字符串)。"""
|
||||
return [f for f in config.fields if f.dtype in _SIGNAL_DTYPES]
|
||||
@@ -326,6 +342,56 @@ def attach_ext_columns(
|
||||
return df
|
||||
|
||||
|
||||
def attach_matrix_ext_fields(market: Any, data_dir: Path | None, names: Any) -> Any:
|
||||
"""把时序扩展数值列作为 matrix fields 附加到 (frozen) MarketDataMatrix 副本。
|
||||
|
||||
回测矩阵只读 enriched parquet 的 14 个存储列 (扩展列按设计不落盘, 见
|
||||
pipeline._select_storage_cols); 扩展数据在这里按矩阵自身的 (日期×标的)
|
||||
骨架现场 join — 与策略页帧路径同一台 attach_ext_columns (PIT 安全:
|
||||
include_snapshot=False, 只进时序模式)。left join 保持骨架行序, 直接
|
||||
reshape 为 (T, A); 缺失 (标的, 日期) 为 NaN, 与叠加条件掩码"NaN 不命中"
|
||||
及帧路径 fill_null(False) 同语义。财务因子的 attach_matrix_fundamental_fields
|
||||
是同构先例。
|
||||
"""
|
||||
import dataclasses
|
||||
from types import MappingProxyType
|
||||
|
||||
import numpy as np
|
||||
|
||||
wanted = sorted({str(n) for n in names})
|
||||
if not wanted:
|
||||
return market
|
||||
labels = list(market.timestamp_labels)
|
||||
symbols = list(market.symbols)
|
||||
n_t, n_a = len(labels), len(symbols)
|
||||
if n_t == 0 or n_a == 0:
|
||||
return market
|
||||
skeleton = pl.DataFrame({
|
||||
"date": np.repeat(np.array(labels, dtype=object), n_a),
|
||||
"symbol": np.tile(np.array(symbols, dtype=object), n_t),
|
||||
})
|
||||
enriched = attach_ext_columns(skeleton, include_snapshot=False, data_dir=data_dir)
|
||||
missing = [c for c in wanted if c not in enriched.columns]
|
||||
if missing:
|
||||
raise ValueError(
|
||||
f"扩展字段数据不可用: {missing} — 检查扩展数据是否已拉取、字段是否仍存在"
|
||||
)
|
||||
extra: dict[str, np.ndarray] = {}
|
||||
for name in wanted:
|
||||
arr = (
|
||||
enriched[name]
|
||||
.cast(pl.Float64)
|
||||
.fill_null(float("nan"))
|
||||
.to_numpy()
|
||||
.astype(np.float32)
|
||||
)
|
||||
arr = arr.reshape(n_t, n_a)
|
||||
arr.flags.writeable = False
|
||||
extra[name] = arr
|
||||
merged = {**dict(market.fields), **extra}
|
||||
return dataclasses.replace(market, fields=MappingProxyType(merged))
|
||||
|
||||
|
||||
def invalidate_ext_caches(data_dir: Path | None = None, *, keep_strategy_cache: bool = False) -> None:
|
||||
"""扩展数据/配置变更后的失效入口 (写入端自动调用)。
|
||||
|
||||
|
||||
@@ -0,0 +1,295 @@
|
||||
"""矩阵回测的扩展数据字段注入 — 对齐 / 缺失补空 / 模式拦截 / 叠加条件联动。
|
||||
|
||||
背景: enriched parquet 只落 14 个存储列 (扩展列按设计不落盘), 扩展时序数值
|
||||
字段由引擎在矩阵加载后按 (日期, 标的) 骨架现场 join (attach_matrix_ext_fields),
|
||||
与财务因子 attach_matrix_fundamental_fields 同构。设计对齐 docs/features.md
|
||||
「第三方数据接入」与 strategy.py 叠加条件矩阵列的注入约定。
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
from datetime import date
|
||||
from pathlib import Path
|
||||
from types import SimpleNamespace
|
||||
|
||||
import numpy as np
|
||||
import polars as pl
|
||||
import pytest
|
||||
|
||||
from app.backtest.engine import BacktestEngine
|
||||
from app.factors.ext_factors import (
|
||||
numeric_timeseries_ext_columns,
|
||||
)
|
||||
from app.tickflow.repository import DataStore, KlineRepository
|
||||
|
||||
SYMS = ["600001.SH", "600002.SH"]
|
||||
DATES = [date(2026, 1, 5), date(2026, 1, 6), date(2026, 1, 7), date(2026, 1, 8)]
|
||||
EXT_COL = "ext_ext_demo_rank" # ext_column_name("ext_demo", "rank")
|
||||
SNAP_COL = "ext_ext_snapdemo_heat" # 快照模式数值列 — 必须被拒
|
||||
STR_COL = "ext_ext_demo_sname" # 字符串列 — 必须被拒
|
||||
|
||||
|
||||
def _write_enriched(data_dir: Path) -> None:
|
||||
for offset, d in enumerate(DATES):
|
||||
rows = []
|
||||
for i, sym in enumerate(SYMS):
|
||||
close = 10.0 + i + offset
|
||||
rows.append({
|
||||
"symbol": sym, "date": d,
|
||||
"open": close, "high": close * 1.01, "low": close * 0.99, "close": close,
|
||||
"volume": 1000.0, "amount": close * 100000.0,
|
||||
"raw_close": close, "raw_high": close * 1.01, "raw_low": close * 0.99,
|
||||
"turnover_rate": 1.0 + i,
|
||||
"consecutive_limit_ups": 0, "consecutive_limit_downs": 0,
|
||||
})
|
||||
part = data_dir / "kline_daily_enriched" / f"date={d.isoformat()}"
|
||||
part.mkdir(parents=True)
|
||||
pl.DataFrame(rows).write_parquet(part / "part.parquet")
|
||||
inst = data_dir / "instruments"
|
||||
inst.mkdir(parents=True)
|
||||
pl.DataFrame({
|
||||
"symbol": SYMS, "name": ["甲", "乙"],
|
||||
"total_shares": [1e9, 1e9], "float_shares": [1e9, 1e9],
|
||||
}).write_parquet(inst / "part.parquet")
|
||||
|
||||
|
||||
def _write_ext_config(data_dir: Path, config_id: str, mode: str, fields: list[dict]) -> None:
|
||||
cfg_dir = data_dir / "ext_data" / config_id
|
||||
cfg_dir.mkdir(parents=True)
|
||||
(cfg_dir / "config.json").write_text(json.dumps({
|
||||
"id": config_id, "label": config_id, "mode": mode, "fields": fields,
|
||||
"symbol_map": {"type": "mapped", "col": "symbol"},
|
||||
}, ensure_ascii=False), encoding="utf-8")
|
||||
|
||||
|
||||
def _write_ext_timeseries(data_dir: Path, rows: list[dict]) -> None:
|
||||
"""rows: [{symbol, date, rank, sname}] — 缺行即缺失组合 (期望 NaN)。"""
|
||||
by_date: dict[str, list[dict]] = {}
|
||||
for r in rows:
|
||||
by_date.setdefault(r["date"], []).append(r)
|
||||
ts = data_dir / "ext_data" / "ext_demo" / "timeseries"
|
||||
ts.mkdir(parents=True)
|
||||
for d, day_rows in by_date.items():
|
||||
part = ts / f"date={d}"
|
||||
part.mkdir(parents=True)
|
||||
pl.DataFrame(day_rows).write_parquet(part / "part.parquet")
|
||||
|
||||
|
||||
@pytest.fixture()
|
||||
def data_dir(tmp_path: Path) -> Path:
|
||||
_write_enriched(tmp_path)
|
||||
_write_ext_config(tmp_path, "ext_demo", "timeseries", [
|
||||
{"name": "rank", "dtype": "int", "label": "排名"},
|
||||
{"name": "sname", "dtype": "string", "label": "名称"},
|
||||
])
|
||||
_write_ext_config(tmp_path, "ext_snapdemo", "snapshot", [
|
||||
{"name": "heat", "dtype": "float", "label": "热度"},
|
||||
])
|
||||
# 数据布局: 600001 全 4 天有排名; 600002 只有前 2 天 (后 2 天缺失 → NaN);
|
||||
# 1/5 顺带带字符串列验证只注入数值。
|
||||
rows = []
|
||||
for k, d in enumerate(DATES):
|
||||
# 600001: 首日 15 (区间外), 其余 50 (20<50<=100 命中叠加区间)
|
||||
rows.append({"symbol": SYMS[0], "date": d.isoformat(), "rank": 15 if k == 0 else 50, "sname": "甲"})
|
||||
for d in DATES[:2]:
|
||||
rows.append({"symbol": SYMS[1], "date": d.isoformat(), "rank": 150, "sname": "乙"})
|
||||
_write_ext_timeseries(tmp_path, rows)
|
||||
return tmp_path
|
||||
|
||||
|
||||
def _engine(data_dir: Path) -> BacktestEngine:
|
||||
return BacktestEngine(KlineRepository(DataStore(data_dir)))
|
||||
|
||||
|
||||
def _plan(matrix_extra: frozenset[str]) -> SimpleNamespace:
|
||||
return SimpleNamespace(
|
||||
execution_backend="matrix_native",
|
||||
base_columns=frozenset({"turnover_rate"}),
|
||||
instrument_columns=frozenset(),
|
||||
matrix_columns=matrix_extra,
|
||||
fundamental_columns=frozenset(),
|
||||
)
|
||||
|
||||
|
||||
def test_numeric_timeseries_columns_filter(data_dir: Path):
|
||||
cols = numeric_timeseries_ext_columns(data_dir)
|
||||
assert EXT_COL in cols
|
||||
assert SNAP_COL not in cols # 快照模式排除 (回看历史 = 未来函数)
|
||||
assert STR_COL not in cols # 字符串列不进矩阵
|
||||
|
||||
|
||||
def test_matrix_load_attaches_ext_fields_aligned(data_dir: Path):
|
||||
engine = _engine(data_dir)
|
||||
market = engine.load_market_data_matrix_for_backtest(
|
||||
None, DATES[0], DATES[-1], _plan(frozenset({EXT_COL})),
|
||||
)
|
||||
assert EXT_COL in market.fields
|
||||
arr = np.asarray(market.fields[EXT_COL])
|
||||
assert arr.shape == (len(DATES), len(SYMS))
|
||||
labels = list(market.timestamp_labels)
|
||||
syms = list(market.symbols)
|
||||
i0, i1 = labels.index(DATES[0].isoformat()), syms.index(SYMS[0])
|
||||
j1 = syms.index(SYMS[1])
|
||||
assert arr[i0, i0 - i0] == pytest.approx(15.0) # 600001 首日排名 15
|
||||
assert arr[labels.index(DATES[-1].isoformat()), j1] != arr[labels.index(DATES[-1].isoformat()), j1] # 600002 末日缺失 → NaN
|
||||
assert arr[labels.index(DATES[1].isoformat()), j1] == pytest.approx(150.0)
|
||||
# 基础列照常来自 parquet
|
||||
assert "turnover_rate" in market.fields
|
||||
# 数组只读 (frozen 副本约定, 同财务因子)
|
||||
with pytest.raises(ValueError):
|
||||
arr[0, 0] = 1.0
|
||||
|
||||
|
||||
def test_matrix_load_rejects_snapshot_or_unknown_ext(data_dir: Path):
|
||||
engine = _engine(data_dir)
|
||||
with pytest.raises(ValueError, match=r"不可用于矩阵回测"):
|
||||
engine.load_market_data_matrix_for_backtest(
|
||||
None, DATES[0], DATES[-1], _plan(frozenset({SNAP_COL})),
|
||||
)
|
||||
with pytest.raises(ValueError, match=r"不可用于矩阵回测"):
|
||||
engine.load_market_data_matrix_for_backtest(
|
||||
None, DATES[0], DATES[-1], _plan(frozenset({"ext_ext_demo_ranks"})), # 拼错的名
|
||||
)
|
||||
|
||||
|
||||
def test_overlay_mask_on_ext_field(data_dir: Path):
|
||||
"""端到端语义: 叠加条件在注入后的 ext 字段上评估, NaN 不命中。"""
|
||||
from app.backtest.strategy import StrategyBacktestService
|
||||
|
||||
engine = _engine(data_dir)
|
||||
market = engine.load_market_data_matrix_for_backtest(
|
||||
None, DATES[0], DATES[-1], _plan(frozenset({EXT_COL})),
|
||||
)
|
||||
mask, err = StrategyBacktestService._overlay_matrix_mask(market, [
|
||||
{"left": EXT_COL, "op": ">", "right": "20", "leftDays": 0, "rightDays": 0},
|
||||
{"left": EXT_COL, "op": "<=", "right": "100", "leftDays": 0, "rightDays": 0},
|
||||
])
|
||||
assert err is None and mask is not None
|
||||
arr = np.asarray(market.fields[EXT_COL])
|
||||
# 掩码与数组逐元素一致: 15 命中 (20<15? 不 — 15>20 为假!)
|
||||
np.testing.assert_array_equal(mask, (arr > 20) & (arr <= 100))
|
||||
# 正命中样本: 600001 第 2~4 天 (rank=50) 命中, 首日 15 与 600002 全部不命中
|
||||
assert mask[1, 0] and mask[2, 0] and mask[3, 0]
|
||||
assert not mask[0, 0] and not mask.all(axis=0)[1]
|
||||
|
||||
|
||||
def test_attach_requires_existing_data(data_dir: Path):
|
||||
"""配置存在但数据全空 → 明确报错而非静默全 NaN (fail-closed)。"""
|
||||
empty_dir = data_dir / "ext_data" / "ext_empty"
|
||||
empty_dir.mkdir(parents=True)
|
||||
(empty_dir / "config.json").write_text(json.dumps({
|
||||
"id": "ext_empty", "label": "e", "mode": "timeseries",
|
||||
"fields": [{"name": "v", "dtype": "float"}],
|
||||
"symbol_map": {"type": "mapped", "col": "symbol"},
|
||||
}), encoding="utf-8")
|
||||
engine = _engine(data_dir)
|
||||
with pytest.raises(ValueError, match=r"扩展字段数据不可用"):
|
||||
engine.load_market_data_matrix_for_backtest(
|
||||
None, DATES[0], DATES[-1], _plan(frozenset({"ext_ext_empty_v"})),
|
||||
)
|
||||
|
||||
|
||||
# ---------------------------------------------------------------
|
||||
# 端到端: 矩阵策略 + 扩展叠加条件真实回测 (回归: SignalMatrix 只读修复)
|
||||
# ---------------------------------------------------------------
|
||||
|
||||
_ALWAYS_ENTRY_STRATEGY = '''import numpy as np
|
||||
from app.backtest.matrix import make_signal_matrix
|
||||
|
||||
META = {
|
||||
"id": "ext_overlay_probe",
|
||||
"name": "ext叠加探针",
|
||||
"asset_types": ["stock"],
|
||||
"timeframes": ["1d"],
|
||||
"params": [],
|
||||
"scoring": {},
|
||||
"required_features": ["open", "high", "low", "close", "volume"],
|
||||
}
|
||||
EXECUTION_BACKEND = "matrix_native"
|
||||
ENTRY_SIGNALS = []
|
||||
EXIT_SIGNALS = []
|
||||
STOP_LOSS = None
|
||||
MAX_HOLD_DAYS = 1
|
||||
|
||||
class AlwaysEntry:
|
||||
def required_fields(self):
|
||||
return frozenset({"open", "high", "low", "close", "volume"})
|
||||
def required_warmup_bars(self, params):
|
||||
return 1
|
||||
def compute_signals(self, market, params):
|
||||
entry = np.ones(market.shape, dtype=np.uint8)
|
||||
return make_signal_matrix(market.shape, entry=entry)
|
||||
|
||||
MATRIX_STRATEGY = AlwaysEntry()
|
||||
'''
|
||||
|
||||
|
||||
def _write_strategy(data_dir: Path) -> None:
|
||||
sd = data_dir / "strategies" / "custom"
|
||||
sd.mkdir(parents=True, exist_ok=True)
|
||||
(sd / "ext_overlay_probe.py").write_text(_ALWAYS_ENTRY_STRATEGY, encoding="utf-8")
|
||||
|
||||
|
||||
def _service_for(data_dir: Path):
|
||||
from app.backtest.strategy import StrategyBacktestService, StrategyEngine
|
||||
|
||||
_write_strategy(data_dir)
|
||||
engine = StrategyEngine(strategy_dirs=[data_dir / "strategies" / "custom"])
|
||||
return StrategyBacktestService(BacktestEngine(KlineRepository(DataStore(data_dir))), engine)
|
||||
|
||||
|
||||
def _ext_overlay_conditions() -> list[dict]:
|
||||
return [{"left": EXT_COL, "op": "<=", "right": "100", "leftDays": 0, "rightDays": 0}]
|
||||
|
||||
|
||||
def test_matrix_backtest_runs_with_ext_overlay(data_dir: Path):
|
||||
"""扩展叠加条件端到端: 全命中版本正常出信号 (回归 SignalMatrix 只读修复)。"""
|
||||
from app.backtest.strategy import StrategyBacktestConfig
|
||||
|
||||
service = _service_for(data_dir)
|
||||
res = service.run(StrategyBacktestConfig(
|
||||
strategy_id="ext_overlay_probe", symbols=None,
|
||||
start=DATES[0], end=DATES[-1],
|
||||
overrides={
|
||||
"basic_filter": {"enabled": False}, # 合成小盘股过不了默认市值过滤
|
||||
"overlay_filter": _ext_overlay_conditions(),
|
||||
},
|
||||
matching="close_t", fees_pct=0, slippage_bps=0, max_positions=2,
|
||||
))
|
||||
assert res.error is None, res.error
|
||||
assert (res.stats.get("n_trades") or 0) > 0
|
||||
|
||||
|
||||
def test_matrix_backtest_ext_overlay_can_block_all(data_data_dir=None):
|
||||
"""反证过滤真实生效: 永不满足的叠加条件 → 无买入信号。"""
|
||||
import tempfile
|
||||
|
||||
from app.backtest.strategy import StrategyBacktestConfig
|
||||
with tempfile.TemporaryDirectory() as td:
|
||||
data_dir = Path(td)
|
||||
_write_enriched(data_dir)
|
||||
_write_ext_config(data_dir, "ext_demo", "timeseries", [
|
||||
{"name": "rank", "dtype": "int", "label": "排名"},
|
||||
{"name": "sname", "dtype": "string", "label": "名称"},
|
||||
])
|
||||
rows = [
|
||||
{"symbol": SYMS[0], "date": d.isoformat(), "rank": 15, "sname": "甲"}
|
||||
for d in DATES
|
||||
] + [
|
||||
{"symbol": SYMS[1], "date": d.isoformat(), "rank": 150, "sname": "乙"}
|
||||
for d in DATES[:2]
|
||||
]
|
||||
_write_ext_timeseries(data_dir, rows)
|
||||
service = _service_for(data_dir)
|
||||
res = service.run(StrategyBacktestConfig(
|
||||
strategy_id="ext_overlay_probe", symbols=None,
|
||||
start=DATES[0], end=DATES[-1],
|
||||
overrides={
|
||||
"basic_filter": {"enabled": False},
|
||||
"overlay_filter": [
|
||||
{"left": EXT_COL, "op": ">", "right": "1000", "leftDays": 0, "rightDays": 0},
|
||||
],
|
||||
},
|
||||
matching="close_t", fees_pct=0, slippage_bps=0, max_positions=2,
|
||||
))
|
||||
assert res.error is not None and "未产生买入信号" in res.error
|
||||
@@ -60,6 +60,10 @@
|
||||
|
||||
**结果沉淀**:回测结果可**导出 CSV**(概要 / 净值曲线 / 交易明细 / 分标的统计四段),并可**保存候选**(完整参数与指标快照);候选在回测页「候选方案」中查看,支持**一键载入复测**,回填全部参数继续研究。
|
||||
|
||||
**稳健参数筛选(验证 tab)**:参数不再按单窗样本内得分排名——每组参数在 K 个互不重叠验证段上独立回测,按稳健分 R = 0.6×中位 + 0.4×最差 排名,只有在多段行情上都好的参数才能当冠军。配套:主目标白名单(排除胜率/持仓天数等退化目标)、K 段合计最小交易数门槛(fail-closed,不足者降序垫底不静默剔除)、Top-20 全窗 IS 细化(样本内乐观偏差/退化率)、冠军收敛度诊断(各段冠军是否收敛于同一组)、高原判据(冠军网格邻域得分衰减,孤峰=过拟合警报)、运行前验证窗预览。冠军参数可一键应用(走策略配置 override,策略页可还原)。旧单窗样本内排名路径已移除。
|
||||
|
||||
**滚动样本外检验(验证 tab)**:每折在训练区间网格优化选参、紧邻测试区间做样本外(OOS)验证,滚动前移;核心产出是 OOS 拼接复利净值与 IS→OOS 退化(过拟合信号)。补强:每折 OOS 最小交易数门槛(不足标记跳过不混入复利)、运行前折预览(<3 折告警)、参数漂移轨迹(逐折最优值 + 稳定/漂移徽章)、「采用最新折参数」(确认框注明来源折,走策略配置 override)。
|
||||
|
||||
**因子与策略挖掘**:复用已有日频因子和 matrix-native 策略，通过 T-1 市场环境、相关去重和嵌套样本外验证生成研究候选。任务在 spawn worker 中运行，支持持久 run ID、取消、刷新重连和显式发布；自动周度任务默认关闭且永不自动发布。完整口径见 [因子与策略挖掘](./mining.md)。
|
||||
|
||||
**策略设置统一**:回测页「高级策略设置」抽屉与策略页设置弹窗采用同一信息架构(策略条件 / 策略权重 / 风控触发 / 回测范围)。风控与持仓纪律(止损 · 止盈 · 移动止损 · 回撤止盈 · 最长持仓)以策略设置保存的值为默认,回测抽屉内可临时覆盖,也可「应用到策略」持久化;策略页保存的叠加条件随回测请求继承,与扫描 / 监控同一口径;回测范围(资产类型 / 股票池)为回测专属概念。
|
||||
@@ -181,6 +185,9 @@
|
||||
|
||||
详见 [custom-data-source.md](./custom-data-source.md) 与 [plugin-development.md](./plugin-development.md)。
|
||||
|
||||
### 扩展数据进回测
|
||||
数值型**时序模式**扩展字段(人气排名等)可在回测中使用:叠加条件 / 策略评分引用 `ext_*` 字段时,引擎在矩阵加载后按 (日期, 标的) 对齐现场注入(与策略页帧路径同一台 PIT 安全 join,缺失补空值=条件不命中)。快照模式扩展字段与字符串扩展字段不进回测(快照回看历史即未来函数);引用时回测前置校验直接拒绝并说明原因。
|
||||
|
||||
### 全量分钟(盘中增量落盘)
|
||||
|
||||
「全量分钟」是独立能力(`full_minute`),**与其他能力同样可路由**:盘中将全市场当日 1 分钟 K 持续增量落盘 `data/kline_minute/`,分钟策略与分时视图即取即用。接入方式:TickFlow Expert,或声明 `full_minute` 数据集的插件/自定义源(设置 → 数据源 → 全量分钟 路由);配好能力后开启实时行情即自动运行(冷启动全天修复轮 + 稳态增量轮)。
|
||||
|
||||
Reference in New Issue
Block a user