feat(strategy): filter_history 列依赖 AST 推导 — 消灭回测全量兜底 (P1a)

- feature_deps: 追踪 pl.col/cols 常量引用推导输入列; 动态列名/pl.all/cs.* 等不可证明形态一律返回 None, 保持全量兜底 (fail-closed, 宁慢不错列) - engine: REQUIRED_FEATURES 缺失且存在动态过滤时加载期自动推导 - bench_feature_resolution: legacy(空依赖) vs 推导的解析宽度对比基准 - 实测: 看板策略等 4 个无声明策略 37→3~5 指标列 (宽度比 7~12x); short_term_top10 动态列名循环正确保持兜底 Signed-off-by: shy3130 <shy3130@qq.com>
2026-11-02 10:26:13 +00:00 · 2026-10-26 13:39:39 +08:00
parent e9b938a3ac
commit 8cedb67c36
4 changed files with 380 additions and 2 deletions
@@ -0,0 +1,120 @@
+"""P0/P1a 基准: 动态 filter_history 策略的依赖解析宽度对比。
+
+对同一策略比较两种依赖声明下的 ResolvedFeaturePlan:
+  legacy   — required_features 为空 (P1a 之前的行为, 触发全量指标兜底)
+  inferred — 引擎 AST 推导结果 (P1a 之后, 不可推导的策略无此行)
+
+记录解析耗时与 indicator/base 列宽。数据加载 IO 大致随列宽线性增长,
+宽度比即回测数据准备阶段的近似提速上界 (非端到端回测耗时)。
+
+用法:
+  cd backend && uv run python scripts/bench_feature_resolution.py [strategy_id ...]
+  不传 id 时扫描全部 filter_history 策略。--json PATH 另存结果。
+"""
+from __future__ import annotations
+
+import argparse
+import dataclasses
+import json
+import logging
+import sys
+import time
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
+
+# legacy 臂会故意触发全量兜底 WARNING, 基准输出只保留表格。
+logging.getLogger("app.backtest.strategy").setLevel(logging.ERROR)
+
+from app.backtest.strategy import StrategyDependencyResolver  # noqa: E402
+from app.config import settings  # noqa: E402
+from app.strategy.engine import StrategyEngine  # noqa: E402
+
+
+def _build_engine() -> StrategyEngine:
+    dirs = [
+        Path(__file__).resolve().parents[1] / "app" / "strategy" / "research",
+        settings.data_dir / "strategies" / "custom",
+        settings.data_dir / "strategies" / "ai",
+        settings.data_dir / "strategies" / "composite",
+    ]
+    return StrategyEngine(strategy_dirs=[d for d in dirs if d.exists()])
+
+
+def _resolve_quiet(resolver: StrategyDependencyResolver, strategy, required: frozenset):
+    """给定 required_features 解析 FeaturePlan, 返回 (耗时ms, plan)。"""
+    patched = dataclasses.replace(strategy, required_features=required)
+    t0 = time.perf_counter()
+    plan = resolver.resolve(
+        patched,
+        params={p.get("id"): p.get("default") for p in strategy.meta.get("params", [])},
+        basic_filter=strategy.basic_filter or {"enabled": False},
+        entry_signals=strategy.entry_signals,
+        exit_signals=strategy.exit_signals,
+    )
+    elapsed_ms = (time.perf_counter() - t0) * 1000
+    return elapsed_ms, plan
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("ids", nargs="*", help="策略 id; 缺省扫描全部动态策略")
+    parser.add_argument("--json", dest="json_path", default=None)
+    args = parser.parse_args()
+
+    engine = _build_engine()
+    resolver = StrategyDependencyResolver()
+    strategies = [
+        s for s in engine.list_strategies_full() if s.filter_history_fn is not None
+    ] if hasattr(engine, "list_strategies_full") else []
+    if not strategies:
+        # list 接口名差异兜底: 直接遍历注册表
+        strategies = [
+            s for s in engine._strategies.values() if s.filter_history_fn is not None
+        ]
+    if args.ids:
+        strategies = [s for s in strategies if s.meta.get("id") in set(args.ids)]
+
+    rows: list[dict] = []
+    for s in strategies:
+        sid = s.meta.get("id", "<unknown>")
+        declared = bool(s.required_features)
+        legacy_ms, legacy_plan = _resolve_quiet(resolver, s, frozenset())
+        inferred_ms, inferred_plan = _resolve_quiet(resolver, s, s.required_features)
+        legacy_w = len(legacy_plan.indicator_columns)
+        inferred_w = len(inferred_plan.indicator_columns)
+        rows.append({
+            "strategy": sid,
+            "declared_required_features": declared,
+            "legacy": {
+                "resolve_ms": round(legacy_ms, 2),
+                "indicator_cols": legacy_w,
+                "base_cols": len(legacy_plan.base_columns),
+            },
+            "inferred": {
+                "resolve_ms": round(inferred_ms, 2),
+                "indicator_cols": inferred_w,
+                "base_cols": len(inferred_plan.base_columns),
+            },
+            "indicator_width_ratio": round(legacy_w / inferred_w, 2) if inferred_w else None,
+        })
+
+    print(f"{'strategy':36} {'依赖':4} {'legacy列宽':>9} {'当前列宽':>8} {'宽度比':>6}")
+    for r in rows:
+        # 依赖列非空 = 显式声明或 P1a 推导; 空的唯一情形是不可证明 (保持全量兜底)。
+        has_deps = "有" if r["declared_required_features"] else "无"
+        ratio = r["indicator_width_ratio"]
+        print(
+            f"{r['strategy']:36} {has_deps:4} "
+            f"{r['legacy']['indicator_cols']:>9} {r['inferred']['indicator_cols']:>8} "
+            f"{(f'{ratio:.1f}x') if ratio else '—':>6}"
+        )
+    if args.json_path:
+        Path(args.json_path).write_text(
+            json.dumps({"rows": rows}, ensure_ascii=False, indent=2), encoding="utf-8"
+        )
+        print(f"\n已写入 {args.json_path}")
+
+
+if __name__ == "__main__":
+    main()