feat(backtest): 支持5m/15m周期回测(ashare适配绕vnpy Interval enum)

ashare_engine override load_data: 5m/15m 直查 dbbardata 转 BarData(MINUTE),
绕过 vnpy Interval enum 限制(原生只认 d/1h/1m)。cta_engine +interval 参数,
5m/15m 映射 MINUTE 过父类校验。api(schema/routes)+orchestrator 透传 interval。
顺带修 cta_engine DB路径污染(yaml NAS路径在Win VPS误解析→改读vt_setting.json)。

测试: 15m 600519一年 3888bars total_return-0.22 sharpe-2.04; 5m 11664bars; 日线未回归(237bars)。

sanguo_api/routes.py

@@ -78,7 +78,8 @@ async def submit_cta(req: CtaBacktestRequest):
         cfg=None,
         benchmark=req.benchmark,
         capital=req.capital,
-        position_pct=req.position_pct
+        position_pct=req.position_pct,
+        interval=req.interval,
     )
     return {"task_id": tid}
 

sanguo_api/schemas.py

@@ -14,6 +14,8 @@ class CtaBacktestRequest(BaseModel):
     benchmark: str = "hs300"
     capital: float = 1_000_000
     position_pct: float = 0.95
+    # K 线周期："d"=日线(默认) / "5m" / "15m"。5m/15m 走 AShareBacktestingEngine 适配。
+    interval: str = "d"
     # A 股费用参数（可选，前端先不暴露，给默认值）
     commission_rate: float = 0.00025  # 万 2.5
     min_commission: float = 5.0  # 最低 5 元

sanguo_backtest/ashare_engine.py

@@ -6,13 +6,15 @@ vnpy 源码零修改，全部覆写在本文件。
 - AShareBacktestingEngine：
   - send_order 覆写 → 做空拦截（SHORT+OPEN 拒单）+ 定寸重算 volume
   - update_daily_close 覆写 → 工厂换 AShareDailyResult（父类在 :647 实例化 DailyResult）
+  - load_data 覆写 → 支持 5m/15m 周期（vnpy Interval enum 不认 '5m'/'15m'）
 """
 from __future__ import annotations
 
 import logging
+from datetime import datetime
 
 from vnpy_ctastrategy.backtesting import BacktestingEngine, DailyResult
-from vnpy.trader.constant import Direction, Offset
+from vnpy.trader.constant import Direction, Offset, Interval, Exchange
 
 logger = logging.getLogger(__name__)
 
@@ -123,6 +125,11 @@ class AShareBacktestingEngine(BacktestingEngine):
         self.stamp_duty_rate: float = 0.0005  # 卖方 0.05%
         self.transfer_fee_rate: float = 0.00001  # 沪市 0.001%
         self.is_sse: bool = False
+        # 5m/15m 适配：vnpy Interval enum 不认 '5m'/'15m'，cta_engine 把 engine.interval
+        # 映射成 Interval.MINUTE 让父类 set_parameters 校验通过，真实 DB interval 字符串
+        # 存此字段供 load_data 自定义路径使用。默认 "d" 走 vnpy 原生日线路径。
+        self.raw_interval: str = "d"
+        self.sqlite_db_path: str | None = None  # 由 cta_engine 注入（_dcfg.data_paths["vnpy_db"]）
 
     def send_order(
         self,
@@ -174,3 +181,93 @@ class AShareBacktestingEngine(BacktestingEngine):
                 transfer_fee_rate=self.transfer_fee_rate,
                 is_sse=self.is_sse,
             )
+
+    def load_data(self) -> None:
+        """覆写父类：支持 5m/15m 周期（绕过 vnpy Interval enum 限制）。
+
+        vnpy Interval enum 只有 MINUTE('1m')/HOUR('1h')/DAILY('d') 等，不认 '5m'/'15m'。
+        父类 load_data 调 ``INTERVAL_DELTA_MAP[self.interval]`` 和
+        ``load_bar_data(..., self.interval, ...)`` 都依赖 enum，传 '15m' 会 ValueError。
+
+        适配思路：self.interval 已被父类 set_parameters 映射成 Interval.MINUTE（enum
+        校验通过），真实 DB interval 存 self.raw_interval。当 raw_interval 是 5m/15m
+        时，直接 sqlite 查 dbbardata 表，把每行转 BarData(interval=Interval.MINUTE)，
+        绕过 vnpy database_manager 的 enum 限制；其余周期（d/1m 等）走 vnpy 原生路径。
+        """
+        if self.raw_interval not in ("5m", "15m"):
+            super().load_data()
+            return
+        self._load_intraday_data()
+
+    def _load_intraday_data(self) -> None:
+        """5m/15m 直查 SQLite → BarData(MINUTE)，绕过 vnpy enum 限制。"""
+        from vnpy.trader.object import BarData
+        import sqlite3
+
+        self.output(f"开始加载 {self.raw_interval} 历史数据（ashare 适配）")
+
+        if not self.end:
+            self.end = datetime.now()
+        if self.start >= self.end:
+            self.output("起始日期必须小于结束日期")
+            return
+
+        db_path = self.sqlite_db_path
+        if not db_path:
+            try:
+                from vnpy.trader.setting import SETTINGS
+                db_path = SETTINGS.get("database.database")
+            except Exception:
+                db_path = None
+        if not db_path:
+            raise RuntimeError(
+                "5m/15m 回测需要 sqlite_db_path（或 SETTINGS['database.database']），"
+                "cta_engine 应在 load_data 前注入。"
+            )
+
+        symbol, exchange_str = self.vt_symbol.split(".")
+        # peewee DateTimeField 存的是 ISO 字符串；start/end 用 datetime 比较即可
+        # （SQLite 会把参数转成可比较的字符串形式）。
+        conn = sqlite3.connect(db_path)
+        try:
+            cur = conn.execute(
+                "SELECT datetime, volume, turnover, open_interest, "
+                "open_price, high_price, low_price, close_price "
+                "FROM dbbardata "
+                "WHERE symbol=? AND exchange=? AND interval=? "
+                "AND datetime>=? AND datetime<=? "
+                "ORDER BY datetime",
+                (symbol, exchange_str, self.raw_interval, self.start, self.end),
+            )
+            rows = cur.fetchall()
+        finally:
+            conn.close()
+
+        exchange = Exchange(exchange_str)
+        bars: list[BarData] = []
+        for dt, vol, turnover, oi, o, h, l, c in rows:
+            if isinstance(dt, str):
+                try:
+                    dt = datetime.fromisoformat(dt)
+                except ValueError:
+                    continue
+            # vnpy_sqlite save 时 convert_tz 改成 UTC，回测时按本地时间跑即可
+            # （日线回测也是直接读 DB datetime，行为一致）。
+            bars.append(BarData(
+                symbol=symbol,
+                exchange=exchange,
+                datetime=dt,
+                interval=Interval.MINUTE,  # 5m/15m 不在 enum，统一标 MINUTE
+                volume=float(vol or 0),
+                turnover=float(turnover or 0),
+                open_interest=float(oi or 0),
+                open_price=float(o or 0),
+                high_price=float(h or 0),
+                low_price=float(l or 0),
+                close_price=float(c or 0),
+                gateway_name="sqlite",
+            ))
+
+        self.history_data.clear()
+        self.history_data.extend(bars)
+        self.output(f"历史数据加载完成，数据量：{len(self.history_data)}")

sanguo_backtest/cta_engine.py

@@ -67,6 +67,27 @@ def guess_exchange(symbol: str) -> Exchange:
     return Exchange("SSE")
 
 
+def _read_vt_setting_db() -> str | None:
+    """从 vnpy 原生 vt_setting.json 直接读 database.database 路径。
+
+    vnpy SETTINGS 字典会被 datareader.read_index_daily 等代码覆写为 yaml 的
+    NAS Linux 路径（Windows VPS 上不存在但有 0 字节空文件，os.path.exists 误判），
+    导致连续多次回测时第二次起路径污染。vt_setting.json 是机器本地配置，权威。
+    """
+    try:
+        import json
+        from pathlib import Path
+        # vnpy 约定：~/.vntrader/vt_setting.json
+        home = Path.home() / ".vntrader" / "vt_setting.json"
+        if not home.exists():
+            return None
+        with open(home, "r", encoding="utf-8") as f:
+            data = json.load(f)
+        return data.get("database.database")
+    except Exception:
+        return None
+
+
 def run_cta_backtest(
     strategy_class,
     symbol: str,
@@ -83,6 +104,7 @@ def run_cta_backtest(
     min_commission: float = 5.0,
     stamp_duty_rate: float = 0.0005,
     transfer_fee_rate: float = 0.00001,
+    interval: str = "d",
 ) -> BacktestResult:
     """
     Run CTA strategy backtest using AShareBacktestingEngine (vnpy 子类化).
@@ -104,6 +126,9 @@ def run_cta_backtest(
         min_commission: 单笔最低佣金（默认 5 元）
         stamp_duty_rate: 印花税率卖方（默认 0.0005）
         transfer_fee_rate: 过户费率沪市（默认 0.00001）
+        interval: K 线周期，"d"=日线(默认) / "5m" / "15m"。
+            5m/15m 走 AShareBacktestingEngine 适配层（直查 dbbardata interval='5m'/'15m'，
+            绕过 vnpy Interval enum 不认非标周期的限制）。
 
     Returns:
         BacktestResult: Result object with backtest statistics and status
@@ -127,10 +152,18 @@ def run_cta_backtest(
         # Create and configure A-share backtesting engine
         engine = AShareBacktestingEngine()
 
+        # 周期映射：vnpy Interval enum 只有 '1m'/'1h'/'d' 等，不认 '5m'/'15m'。
+        # 5m/15m 时把 enum 传 MINUTE 让父类 set_parameters 校验通过，真实 DB interval
+        # 字符串存 engine.raw_interval 给 AShareBacktestingEngine.load_data 自定义路径用。
+        if interval in ("5m", "15m"):
+            engine_interval = "1m"  # Interval.MINUTE.value
+        else:
+            engine_interval = interval  # "d" / "1m" / "1h" 等 vnpy 原生支持的值
+
         # Set parameters with A-share specific values
         engine.set_parameters(
             vt_symbol=vt_symbol,
-            interval="d",  # Interval.DAILY.value — vnpy enum uses "d" not "1d"
+            interval=engine_interval,
             start=start_dt,
             end=end_dt,
             rate=commission_rate,  # 佣金率（AShareDailyResult 用自身 commission_rate，此处仅保持一致）
@@ -139,6 +172,9 @@ def run_cta_backtest(
             pricetick=0.01,  # Minimum price tick (0.01 yuan for A-shares)
             capital=capital,  # Starting capital
         )
+        # raw_interval 在 5m/15m 时触发 ashare_engine 适配路径；其它周期与 engine.interval
+        # 一致，走 vnpy 原生 load_data。
+        engine.raw_interval = interval
 
         # A 股适配参数（定寸 + 费用）
         engine.position_pct = position_pct
@@ -155,13 +191,32 @@ def run_cta_backtest(
         # doesn't inherit main-process SETTINGS, so set before engine.load_data.
         # _dcfg is also reused by the metrics branch (benchmark data_paths) since the
         # cfg param can be None when called via the API.
+        #
+        # 路径解析优先级：vnpy 原生 vt_setting.json(机器相关，正确) >
+        # yaml data_paths.vnpy_db（模板可能是 NAS Linux 路径，Windows VPS 上不存在）。
+        # 任一不存在时回退到另一个，避免 VPS 上 yaml 写 NAS 路径导致回测加载 0 数据。
+        #
+        # 不能直接用 SETTINGS.get("database.database")：它会被 datareader.read_index_daily
+        # （基准加载）覆写为 yaml NAS 路径，污染后续调用。这里每次重新读 vt_setting.json
+        # 原始值（machine-truth），不依赖被污染的 SETTINGS 缓存。
         _dcfg = None
         try:
             from vnpy.trader.setting import SETTINGS
             from sanguo_data.config import load_config, find_config_path
             _dcfg = load_config(find_config_path())
             SETTINGS["database.name"] = "sqlite"
-            SETTINGS["database.database"] = _dcfg.data_paths["vnpy_db"]
+            _vt_db = _read_vt_setting_db()  # 从 vt_setting.json 直接读（不依赖 SETTINGS dict）
+            _yaml_db = _dcfg.data_paths.get("vnpy_db")
+            if _vt_db and os.path.exists(_vt_db):
+                _resolved_db = _vt_db
+            elif _yaml_db and os.path.exists(_yaml_db):
+                _resolved_db = _yaml_db
+            else:
+                # 都不存在时保留 vnpy SETTINGS（让下游报出真实错误，而不是被 yaml 覆盖）
+                _resolved_db = _vt_db or _yaml_db
+            SETTINGS["database.database"] = _resolved_db
+            # 5m/15m 适配：ashare_engine._load_intraday_data 直查 SQLite 需要 DB 物理路径。
+            engine.sqlite_db_path = _resolved_db
         except Exception as e:
             logging.warning("vnpy 数据库配置加载失败（回测可能无法加载历史数据）: %s", e)
 

sanguo_orchestrator/runner.py

@@ -32,7 +32,8 @@ class Orchestrator:
 
     async def submit_cta(self, strategy_class, symbol: str, params: dict,
                        start: str, end: str, cfg, benchmark: str = "hs300",
-                       capital: float = 1_000_000, position_pct: float = 0.95) -> str:
+                       capital: float = 1_000_000, position_pct: float = 0.95,
+                       interval: str = "d") -> str:
         """Submit a CTA backtesting task asynchronously"""
         # Stable uuid up front → reused as the persisted DB task_id, so runner-id ==
         # DB task_id (durable across restarts; previously used id(params) memory addr).
@@ -48,6 +49,7 @@ class Orchestrator:
             benchmark=benchmark,
             capital=capital,
             position_pct=position_pct,
+            interval=interval,
         )
         await self._notify_stage(task_id, "排队中")
 
@@ -55,7 +57,7 @@ class Orchestrator:
         fut: Future = self.pool.submit_work(
             task_id, _cta_worker, spec["strategy_class"], spec["symbol"],
             spec["params"], spec["start"], spec["end"], spec["cfg"], spec["benchmark"],
-            self.db_path, task_id, spec["capital"], spec["position_pct"]
+            self.db_path, task_id, spec["capital"], spec["position_pct"], spec["interval"]
         )
 
         task = self.pool.get_task(task_id)
@@ -194,10 +196,10 @@ class Orchestrator:
 
 
 # Module-level worker functions (must be top-level for ProcessPoolExecutor pickle)
-def _cta_worker(strategy_class, symbol: str, params: dict, start: str, end: str, cfg, benchmark: str, db_path: str, task_id: str, capital: float = 1_000_000, position_pct: float = 0.95) -> any:
+def _cta_worker(strategy_class, symbol: str, params: dict, start: str, end: str, cfg, benchmark: str, db_path: str, task_id: str, capital: float = 1_000_000, position_pct: float = 0.95, interval: str = "d") -> any:
     """Worker for CTA backtest (lazy import, spawn-friendly)"""
     from sanguo_backtest.cta_engine import run_cta_backtest
-    return run_cta_backtest(strategy_class, symbol, params, start, end, cfg, db_path, benchmark=benchmark, task_id=task_id, capital=capital, position_pct=position_pct)
+    return run_cta_backtest(strategy_class, symbol, params, start, end, cfg, db_path, benchmark=benchmark, task_id=task_id, capital=capital, position_pct=position_pct, interval=interval)
 
 
 def _opt_worker(strategy_class, symbol: str, grid: dict, start: str, end: str, cfg, db_path: str, task_id: str) -> any: