rebuild(screener): add controlled formulas and rolling backtests

next/backend/bootstrap/container.py

@@ -87,8 +87,10 @@ def build_container(settings: Settings) -> ApplicationContainer:
     accounts = AccountService(database, account_repository, PasswordHasher(), cipher)
     memberships = MembershipService(database, account_repository)
     model_pool = ModelPoolService(database, model_pool_repository, cipher)
-    screener = ScreenerService(database, screener_repository, market_repository, gateway)
     llm = LLMGateway(database, LLMRepository(), memberships, model_pool)
+    screener = ScreenerService(
+        database, screener_repository, market_repository, gateway, llm
+    )
     mentor = MentorService(
         database,
         MentorRepository(),

next/backend/database/migrations/m0012_screener_backtests.py

@@ -0,0 +1,31 @@
+from __future__ import annotations
+
+import sqlite3
+
+from backend.database.migrations.runner import Migration
+
+
+def upgrade(connection: sqlite3.Connection) -> None:
+    connection.execute(
+        """
+        CREATE TABLE screener_run_backtests (
+            run_id INTEGER PRIMARY KEY
+                REFERENCES screener_runs(id) ON DELETE CASCADE,
+            payload_json TEXT NOT NULL,
+            created_at TEXT NOT NULL
+        )
+        """
+    )
+
+
+def downgrade(connection: sqlite3.Connection) -> None:
+    connection.execute("DROP TABLE screener_run_backtests")
+
+
+MIGRATION = Migration(
+    version=12,
+    name="create_screener_backtests",
+    signature="screener:v2:persistent-rolling-backtests",
+    upgrade=upgrade,
+    downgrade=downgrade,
+)

next/backend/database/migrations/registry.py

@@ -9,6 +9,7 @@ from backend.database.migrations.m0008_mentor_llm import MIGRATION as MENTOR_LLM
 from backend.database.migrations.m0009_heaven import MIGRATION as HEAVEN
 from backend.database.migrations.m0010_review import MIGRATION as REVIEW
 from backend.database.migrations.m0011_operations import MIGRATION as OPERATIONS
+from backend.database.migrations.m0012_screener_backtests import MIGRATION as SCREENER_BACKTESTS
 from backend.database.migrations.runner import Migration
 
 MIGRATIONS: tuple[Migration, ...] = (
@@ -23,4 +24,5 @@ MIGRATIONS: tuple[Migration, ...] = (
     HEAVEN,
     REVIEW,
     OPERATIONS,
+    SCREENER_BACKTESTS,
 )

next/backend/features/screener/backtest.py

@@ -0,0 +1,84 @@
+from __future__ import annotations
+
+from typing import Any
+
+from backend.features.screener.engine import execute_formula
+
+MINIMUM_STABLE_SAMPLES = 20
+FORWARD_TRADING_DAYS = 3
+
+
+def rolling_backtest(
+    snapshots: list[dict[str, Any]], formula: dict[str, Any]
+) -> dict[str, Any]:
+    returns: list[float] = []
+    evaluated_dates: list[str] = []
+    for index in range(max(0, len(snapshots) - FORWARD_TRADING_DAYS)):
+        current = snapshots[index]
+        future = snapshots[index + FORWARD_TRADING_DAYS]
+        outcome = execute_formula(current["rows"], formula, current["coverage"])
+        if outcome["status"] not in {"completed", "no_signal"}:
+            continue
+        evaluated_dates.append(str(current["trade_date"]))
+        future_closes = {
+            str(row.get("identifier")): _positive(row.get("close"))
+            for row in future["rows"]
+        }
+        for candidate in outcome["items"]:
+            entry = _positive(candidate.get("close"))
+            future_close = future_closes.get(str(candidate.get("identifier")))
+            if entry is None or future_close is None:
+                continue
+            returns.append((future_close / entry - 1) * 100)
+    sample_size = len(returns)
+    stable = sample_size >= MINIMUM_STABLE_SAMPLES
+    period_start = evaluated_dates[0] if evaluated_dates else None
+    period_end = evaluated_dates[-1] if evaluated_dates else None
+    return {
+        "sample_size": sample_size,
+        "evaluated_dates": len(evaluated_dates),
+        "stable": stable,
+        "win_rate": (
+            round(sum(value > 0 for value in returns) / sample_size * 100, 1)
+            if stable
+            else None
+        ),
+        "average_return_3d": (
+            round(sum(returns) / sample_size, 2) if stable else None
+        ),
+        "period_start": period_start,
+        "period_end": period_end,
+        "message": (
+            "历史估计已达到最低样本要求"
+            if stable
+            else (
+                f"小样本：当前{sample_size}个有效样本，"
+                f"满{MINIMUM_STABLE_SAMPLES}个后显示胜率与平均收益"
+            )
+        ),
+    }
+
+
+def attach_historical_estimate(
+    run: dict[str, Any], backtest: dict[str, Any] | None
+) -> dict[str, Any]:
+    run["backtest"] = backtest
+    if not backtest or not backtest.get("stable"):
+        return run
+    win_rate = float(backtest["win_rate"])
+    for item in run.get("items") or []:
+        score = item.get("score_display")
+        item["historical_estimate"] = (
+            round(win_rate * 0.65 + float(score) * 0.35, 1)
+            if isinstance(score, (int, float))
+            else None
+        )
+    return run
+
+
+def _positive(value: Any) -> float | None:
+    try:
+        number = float(value)
+    except (TypeError, ValueError):
+        return None
+    return number if number > 0 else None

next/backend/features/screener/compiler.py

@@ -0,0 +1,113 @@
+from __future__ import annotations
+
+import hashlib
+import json
+from typing import Any
+
+from backend.features.screener.catalog import CatalogError, factor_catalog, validate_formula
+
+PROMPT_VERSION = "screener:formula-compiler:v1"
+
+
+def compile_messages(description: str) -> list[dict[str, str]]:
+    factors = factor_catalog()["factors"]
+    schema = {
+        "universe": {"exclude_st": True, "listed_days_min": 120},
+        "filters": [{"field": "amount_billion", "op": ">=", "value": 1}],
+        "score": [{"field": "return_20d", "weight": 1.0, "direction": "desc"}],
+        "limit": 30,
+        "min_score": 0.5,
+    }
+    return [
+        {
+            "role": "system",
+            "content": (
+                "你是受控选股公式编译器。只输出一个JSON对象，不得输出Markdown、解释、股票或代码。"
+                "只能使用给定因子；filters运算符仅限 >、>=、<、<=、==、!=、between、in；"
+                "score权重必须为0至1小数且总和等于1；direction仅限asc或desc；"
+                "limit为1至50，min_score为0至1。无法完全表达时选择最接近的已知因子，不得创造字段。"
+                f"\nJSON结构：{json.dumps(schema, ensure_ascii=False, separators=(',', ':'))}"
+                f"\n可用因子：{json.dumps(factors, ensure_ascii=False, separators=(',', ':'))}"
+            ),
+        },
+        {"role": "user", "content": description},
+    ]
+
+
+def description_key(description: str) -> str:
+    digest = hashlib.sha256(description.encode("utf-8")).hexdigest()[:16]
+    return f"formula:{digest}"
+
+
+def parse_compiled_formula(content: str) -> dict[str, Any]:
+    text = content.strip()
+    if text.startswith("```"):
+        lines = text.splitlines()
+        if len(lines) < 3 or lines[-1].strip() != "```":
+            raise CatalogError("模型返回的公式格式无效")
+        text = "\n".join(lines[1:-1]).strip()
+    try:
+        raw = json.loads(text)
+    except json.JSONDecodeError as exc:
+        raise CatalogError("模型未返回有效JSON公式") from exc
+    if not isinstance(raw, dict):
+        raise CatalogError("模型返回的公式必须是对象")
+    return normalize_custom_formula(raw)
+
+
+def normalize_custom_formula(raw: dict[str, Any]) -> dict[str, Any]:
+    universe = raw.get("universe")
+    filters = raw.get("filters")
+    scores = raw.get("score")
+    if not isinstance(universe, dict) or not isinstance(filters, list) or not isinstance(
+        scores, list
+    ):
+        raise CatalogError("模型返回的公式结构不完整")
+    exclude_st = universe.get("exclude_st", True)
+    listed_days = universe.get("listed_days_min", 120)
+    if not isinstance(exclude_st, bool) or not isinstance(listed_days, int):
+        raise CatalogError("股票范围设置无效")
+    normalized_scores = []
+    for score in scores:
+        if not isinstance(score, dict):
+            raise CatalogError("评分因子结构无效")
+        normalized_scores.append(
+            {
+                "field": score.get("field"),
+                "weight": score.get("weight"),
+                "direction": score.get("direction", "desc"),
+            }
+        )
+    if any(not isinstance(item, dict) for item in filters):
+        raise CatalogError("筛选条件结构无效")
+    numeric_weights = [item["weight"] for item in normalized_scores]
+    if all(
+        isinstance(value, (int, float)) and not isinstance(value, bool)
+        for value in numeric_weights
+    ):
+        total = sum(float(value) for value in numeric_weights)
+        if abs(total - 100) <= 0.0001:
+            for item in normalized_scores:
+                item["weight"] = float(item["weight"]) / 100
+    minimum = raw.get("min_score")
+    if isinstance(minimum, (int, float)) and not isinstance(minimum, bool) and 1 < minimum <= 100:
+        minimum = float(minimum) / 100
+    formula = {
+        "universe": {
+            "exclude_st": exclude_st,
+            "listed_days_min": listed_days,
+        },
+        "filters": [
+            {
+                "field": item.get("field"),
+                "op": item.get("op"),
+                "value": item.get("value"),
+            }
+            for item in filters
+        ],
+        "score": normalized_scores,
+        "limit": raw.get("limit"),
+        "min_score": minimum,
+    }
+    validate_formula(formula)
+    return formula

next/backend/features/screener/repository.py

@@ -88,6 +88,28 @@ class ScreenerRepository:
             (snapshot_id,),
         ).fetchone()
 
+    def factor_snapshots_through(
+        self, connection: sqlite3.Connection, through: str, limit: int
+    ) -> tuple[sqlite3.Row, ...]:
+        return tuple(
+            connection.execute(
+                """
+                SELECT * FROM (
+                    SELECT snapshots.*,
+                        ROW_NUMBER() OVER (
+                            PARTITION BY trade_date ORDER BY id DESC
+                        ) AS revision_rank
+                    FROM screener_factor_snapshots AS snapshots
+                    WHERE trade_date <= ?
+                )
+                WHERE revision_rank = 1
+                ORDER BY trade_date DESC, id DESC
+                LIMIT ?
+                """,
+                (through, limit),
+            ).fetchall()
+        )
+
     def factor_rows(self, connection: sqlite3.Connection, snapshot_id: int) -> list[dict[str, Any]]:
         return [
             json.loads(str(row["payload_json"]))
@@ -180,6 +202,30 @@ class ScreenerRepository:
             ),
         )
 
+    def save_backtest(
+        self,
+        connection: sqlite3.Connection,
+        run_id: int,
+        payload: dict[str, Any],
+    ) -> None:
+        connection.execute(
+            """
+            INSERT OR IGNORE INTO screener_run_backtests (
+                run_id, payload_json, created_at
+            ) VALUES (?, ?, ?)
+            """,
+            (run_id, _json(payload), _now()),
+        )
+
+    def backtest(
+        self, connection: sqlite3.Connection, run_id: int
+    ) -> dict[str, Any] | None:
+        row = connection.execute(
+            "SELECT payload_json FROM screener_run_backtests WHERE run_id = ?",
+            (run_id,),
+        ).fetchone()
+        return json.loads(str(row["payload_json"])) if row else None
+
     def latest_runs(
         self,
         connection: sqlite3.Connection,

next/backend/features/screener/routes.py

@@ -15,6 +15,8 @@ from backend.features.accounts.auth import (
 )
 from backend.features.screener.schemas import (
     CustomStrategyInput,
+    FormulaCompileInput,
+    FormulaCompileResponse,
     IdentifierResponse,
     MessageResponse,
     ScreenerCatalogResponse,
@@ -24,6 +26,7 @@ from backend.features.screener.schemas import (
 )
 from backend.features.screener.service import ScreenerError
 from backend.http.errors import AppError
+from backend.llm.gateway import LLMGatewayError
 
 router = APIRouter(prefix="/screener", tags=["screener"])
 
@@ -60,6 +63,15 @@ def save_custom(
     return _call(request, "save_custom", principal.user.id, payload.name, payload.formula)
 
 
+@router.post("/formula/compile", response_model=FormulaCompileResponse)
+def compile_formula(
+    payload: FormulaCompileInput,
+    request: Request,
+    principal: SmartWritePrincipal,
+) -> dict:
+    return _call(request, "compile_formula", principal, payload.description)
+
+
 @router.delete("/custom/{strategy_id}", response_model=MessageResponse)
 def delete_custom(
     request: Request,
@@ -110,3 +122,6 @@ def _call(request: Request, method: str, *args):
         return getattr(request.app.state.container.screener, method)(*args)
     except (ScreenerError, MarketDataUnavailable, ProviderError, DataQualityError) as exc:
         raise AppError("screener_unavailable", str(exc), 409) from exc
+    except LLMGatewayError as exc:
+        status = 403 if exc.code in {"membership_required", "quota_exhausted"} else 503
+        raise AppError(exc.code, str(exc), status) from exc

next/backend/features/screener/schemas.py

@@ -38,6 +38,15 @@ class CustomStrategyInput(BaseModel):
     formula: dict[str, Any]
 
 
+class FormulaCompileInput(BaseModel):
+    description: str = Field(min_length=5, max_length=500)
+
+
+class FormulaCompileResponse(BaseModel):
+    formula: dict[str, Any]
+    prompt_version: str
+
+
 class TrackInput(BaseModel):
     run_id: int = Field(gt=0)
     identifier: str = Field(min_length=1, max_length=40)

next/backend/features/screener/service.py

@@ -9,12 +9,20 @@ from backend.data.contracts import SnapshotState
 from backend.data.gateway import DataGateway
 from backend.data.repository import MarketRepository
 from backend.database.connection import Database
+from backend.features.accounts.models import Principal
+from backend.features.screener.backtest import attach_historical_estimate, rolling_backtest
 from backend.features.screener.catalog import (
     CatalogError,
     factor_catalog,
     strategy_catalog,
     validate_formula,
 )
+from backend.features.screener.compiler import (
+    PROMPT_VERSION,
+    compile_messages,
+    description_key,
+    parse_compiled_formula,
+)
 from backend.features.screener.engine import execute_formula
 from backend.features.screener.factors import build_factor_snapshot
 from backend.features.screener.repository import (
@@ -23,6 +31,7 @@ from backend.features.screener.repository import (
     decode_run,
     decode_track,
 )
+from backend.llm.gateway import LLMGateway
 
 SHANGHAI = ZoneInfo("Asia/Shanghai")
 PHASE_REGIMES = {
@@ -47,11 +56,13 @@ class ScreenerService:
         repository: ScreenerRepository,
         market_repository: MarketRepository,
         gateway: DataGateway,
+        llm: LLMGateway,
     ) -> None:
         self._database = database
         self._repository = repository
         self._market_repository = market_repository
         self._gateway = gateway
+        self._llm = llm
 
     def catalog(self) -> dict[str, Any]:
         factors = factor_catalog()
@@ -89,10 +100,16 @@ class ScreenerService:
                 decode_run(row)
                 for row in self._repository.latest_runs(connection, "curated", through)
             ]
-            custom_runs = [
-                decode_run(row)
-                for row in self._repository.latest_runs(connection, "custom", through, user_id)
-            ]
+            custom_runs = []
+            for row in self._repository.latest_runs(connection, "custom", through, user_id):
+                decoded = decode_run(row)
+                if decoded is not None:
+                    custom_runs.append(
+                        attach_historical_estimate(
+                            decoded,
+                            self._repository.backtest(connection, int(row["id"])),
+                        )
+                    )
         return {
             "trade_date": through,
             "message": "",
@@ -161,6 +178,28 @@ class ScreenerService:
                 self._repository.save_custom_strategy(connection, user_id, normalized, formula)
             )
 
+    def compile_formula(self, principal: Principal, description: str) -> dict[str, Any]:
+        normalized = " ".join(description.split())
+        if not 5 <= len(normalized) <= 500:
+            raise ScreenerError("自然语言条件应为5至500个字符")
+        call = self._llm.prepare(
+            principal,
+            feature="screener_formula",
+            prompt_version=PROMPT_VERSION,
+            business_id=description_key(normalized),
+            input_chars=len(normalized),
+        )
+        content = "".join(
+            event.content
+            for event in self._llm.stream(call, compile_messages(normalized))
+            if event.type == "delta"
+        )
+        try:
+            formula = parse_compiled_formula(content)
+        except CatalogError as exc:
+            raise ScreenerError(f"公式转换失败：{exc}") from exc
+        return {"formula": formula, "prompt_version": PROMPT_VERSION}
+
     def delete_custom(self, user_id: int, strategy_id: int) -> None:
         with self._database.transaction() as connection:
             if not self._repository.delete_custom_strategy(connection, user_id, strategy_id):
@@ -184,7 +223,7 @@ class ScreenerService:
         result = self._run(int(snapshot["id"]), str(snapshot["trade_date"]), strategy, user_id)
         if result is None:
             raise ScreenerError("自定义策略执行失败")
-        return result
+        return self._with_backtest(result, strategy["formula"], str(snapshot["trade_date"]))
 
     def tracks(self, user_id: int) -> list[dict[str, Any]]:
         with self._database.read() as connection:
@@ -274,6 +313,32 @@ class ScreenerService:
                 ).fetchone()
             )
 
+    def _with_backtest(
+        self,
+        run: dict[str, Any],
+        formula: dict[str, Any],
+        through: str,
+    ) -> dict[str, Any]:
+        run_id = int(run["id"])
+        with self._database.read() as connection:
+            existing = self._repository.backtest(connection, run_id)
+        if existing is None:
+            with self._database.read() as connection:
+                rows = self._repository.factor_snapshots_through(connection, through, 63)
+                history = [
+                    {
+                        "trade_date": str(row["trade_date"]),
+                        "coverage": json.loads(str(row["coverage_json"])),
+                        "rows": self._repository.factor_rows(connection, int(row["id"])),
+                    }
+                    for row in reversed(rows)
+                ]
+            calculated = rolling_backtest(history, formula)
+            with self._database.transaction() as connection:
+                self._repository.save_backtest(connection, run_id, calculated)
+                existing = self._repository.backtest(connection, run_id)
+        return attach_historical_estimate(run, existing)
+
     def _market_snapshot(self, trade_date: str) -> dict[str, Any]:
         with self._database.read() as connection:
             row = self._market_repository.latest_summary(connection, trade_date)

next/docs/evidence/stage-15/acceptance.md

@@ -8,8 +8,8 @@
 
 ## 减法结果
 
-- 旧运行时代码：55个文件、61,793行；当前重建运行时代码：216个职责文件、24,677行。
-- 运行时代码净减少37,116行，约60.1%；迁移/备份工具、测试和文档未混入运行时比较。
+- 旧运行时代码：55个文件、61,793行；当前重建运行时代码：219个职责文件、25,053行。
+- 运行时代码净减少36,740行，约59.5%；迁移/备份工具、测试和文档未混入运行时比较。
 - 唯一浏览器API、数据网关、LLM网关、弹窗Host、设计令牌和移动规则均通过扫描。
 - LLM网关原有3个账户领域具体类型反向导入已改为最小Protocol，未增加第二套服务。
 - 超过章程建议行数的算法、Provider和CSS已逐项登记保留原因及拆分触发条件，见`../../final/subtraction-audit.md`。
@@ -17,7 +17,7 @@
 ## 维护与回退演练
 
 - 已在真实迁移库副本执行schema 10 -> 9 -> 10，完整性及外键检查通过；明确验证数据回退只能在副本执行。
-- 演练发现并修复`tools.database status`对已有migration误报未知版本的问题；状态查询保持只读；运行治理迁移后当前版本为`schema_version=11`。
+- 演练发现并修复`tools.database status`对已有migration误报未知版本的问题；状态查询保持只读；加入选股回测持久化迁移后当前版本为`schema_version=12`。
 - 人工维护入口、常见改动路径、数据源标准、故障定位与复杂度红线见`../../final/maintenance-guide.md`。
 - NAS切换、观察和回退动作见`../../final/cutover-checklist.md`；正式容器未变更。
 
@@ -28,12 +28,12 @@
 - 生产响应统一设置CSP、`nosniff`、拒绝Frame、Permissions Policy、Referrer Policy；HTTPS增加HSTS。
 - CSP保留Vue动态宽度样式所需的`style-src 'unsafe-inline'`，脚本仍只允许同源。
 - 静态哈希资源长期缓存，SPA入口不缓存；未知API不被SPA接管。
-- 真实迁移库副本市场摘要中位数20.90ms、P95 23.55ms；前端生产JS 296.90KB、CSS 92.08KB（未压缩）。
+- 真实迁移库副本市场摘要中位数20.90ms、P95 23.55ms；前端生产JS 298.87KB、CSS 92.40KB（未压缩）。
 
 ## 最终门禁
 
 - Ruff：通过。
-- pytest：104项通过。
+- pytest：106项通过。
 - Vue TypeScript：通过。
 - Vitest：3个文件、7项通过。
 - Vite生产构建：通过。

next/docs/final/completion-audit.md

@@ -21,14 +21,7 @@
 - 数据库迁移、旧库只读迁移、备份、校验恢复、生产静态文件服务和安全响应头已有自动测试。
 - PC 统一 Shell、设计令牌、日夜主题和独立移动端布局规则已经建立。
 
-## 阻断“完整迁移”的产品缺口
-
-### P0：用户可见功能缺失
-
-1. **自定义选股完整能力**：缺少自然语言转换受控公式和滚动回测；精选策略详情尚未完整显示评分权重、
-   执行频率和风险等级。
-
-### P1：验收和交互覆盖不足
+## 阻断“完整迁移”的验收缺口
 
 1. 固定案例 `C06`、`C08` 当前为未实现，不能标记通过。
 2. `C05`、`U01` 的主题无白闪需要首帧/加载态直接断言，不能只依赖夜间截图。
@@ -59,7 +52,13 @@
 - 状态栏读取真实后台任务与最后成功行情状态，不再显示固定占位文字。
 - 模型池中的每个模型可由管理员独立测试；测试复用统一LLM网关并记录模型、耗时、成功或失败，
   不返回密钥和上游正文，也不计入会员每日额度。
-- 本轮验证为Ruff、104项pytest、Vue类型检查、7项Vitest、生产构建和21项Playwright全部通过。
+- 自然语言选股只允许由统一LLM网关生成白名单JSON公式；未知因子直接拒绝，LLM不参与股票候选计算，
+  用户必须核对生成的因子、权重和过滤条件后再保存。
+- 自定义选股滚动回测只读取已归档因子快照，以入选日至第3个后续交易日收盘计算；少于20个有效样本
+  只提示小样本，达到门槛后才显示胜率和平均3日收益，结果按运行持久化且重复读取不漂移。
+- 精选策略详情直接展示配置中已有的质量、风险、执行频率、适用环境、失效风险、适用阶段和完整评分权重，
+  没有补造无法验证的策略元信息。
+- 本轮验证为Ruff、106项pytest、Vue类型检查、7项Vitest、生产构建和21项Playwright全部通过。
 
 ## 外部环境阻断项
 
@@ -71,7 +70,7 @@
 
 只有以下条件全部满足后，才可恢复“重建完成”的表述：
 
-1. 本文件所有 P0 缺口完成并有直接测试。
+1. 本文件登记的用户可见功能缺口全部完成并有直接测试。
 2. 85 个固定案例逐项标记为“确定性测试”“浏览器直接断言”或“人工视觉验收”；不得使用宽泛截图代替算法证明。
 3. 第 26 节页面覆盖矩阵逐格完成，人工验收项由用户确认。
 4. Docker 构建、启动、持久化重启、备份恢复和回退在实际容器环境通过。

next/frontend/src/pages/screener/CustomPanel.vue

@@ -16,6 +16,7 @@ const props = defineProps<{
   factorGroups: Record<string, string[]>;
   strategies: CustomStrategy[];
   runs: ScreenerRun[];
+  compiledFormula: ScreenerFormula | null;
   locked: boolean;
   busy: boolean;
 }>();
@@ -24,9 +25,11 @@ const emit = defineEmits<{
   run: [strategy: CustomStrategy];
   remove: [strategy: CustomStrategy];
   track: [run: ScreenerRun, candidate: Candidate];
+  compile: [description: string];
 }>();
 
 const name = ref("我的选股策略");
+const naturalLanguage = ref("");
 const selectedFactor = ref("return_20d");
 const scores = ref<FormulaScore[]>([
   { field: "return_20d", weight: 60, direction: "desc" },
@@ -55,6 +58,11 @@ watch(
   },
 );
 
+watch(
+  () => props.compiledFormula,
+  (formula) => { if (formula) applyFormula(formula); },
+);
+
 function addFactor(): void {
   if (!selectedFactor.value || scores.value.some((item) => item.field === selectedFactor.value)) {
     return;
@@ -66,18 +74,22 @@ function addFilter(): void {
   filters.value.push({ field: "amount_billion", op: ">=", value: 1 });
 }
 
-function edit(strategy: CustomStrategy): void {
-  selected.value = strategy.id;
-  name.value = strategy.name;
-  scores.value = strategy.formula.score.map((item) => ({
+function applyFormula(formula: ScreenerFormula): void {
+  scores.value = formula.score.map((item) => ({
     ...item,
     weight: Math.round(item.weight * 100),
   }));
-  filters.value = strategy.formula.filters.map((item) => ({ ...item }));
-  listedDays.value = strategy.formula.universe.listed_days_min;
-  excludeSt.value = strategy.formula.universe.exclude_st;
-  outputLimit.value = strategy.formula.limit;
-  minimumScore.value = Math.round(strategy.formula.min_score * 100);
+  filters.value = formula.filters.map((item) => ({ ...item }));
+  listedDays.value = formula.universe.listed_days_min;
+  excludeSt.value = formula.universe.exclude_st;
+  outputLimit.value = formula.limit;
+  minimumScore.value = Math.round(formula.min_score * 100);
+}
+
+function edit(strategy: CustomStrategy): void {
+  selected.value = strategy.id;
+  name.value = strategy.name;
+  applyFormula(strategy.formula);
 }
 
 function save(): void {
@@ -128,6 +140,10 @@ function save(): void {
     </div>
     <div class="custom-filters">
       <header class="card-header"><h2>过滤与输出</h2></header>
+      <div class="formula-compiler">
+        <label class="field"><span>自然语言转公式</span><textarea v-model="naturalLanguage" class="input" rows="2" maxlength="500" placeholder="例如：近20日走势较强、成交额不少于3亿元，偏重板块强度"></textarea></label>
+        <button class="btn" type="button" :disabled="busy || naturalLanguage.trim().length < 5" @click="emit('compile', naturalLanguage)">{{ busy ? "转换中" : "转换" }}</button>
+      </div>
       <div class="filter-list">
         <div v-for="(item, index) in filters" :key="index">
           <select v-model="item.field" class="select">
@@ -164,6 +180,10 @@ function save(): void {
       <button class="btn" type="button" :disabled="busy" @click="edit(current)">编辑</button>
       <button class="btn" type="button" :disabled="busy" @click="emit('remove', current)">删除</button>
     </div>
+    <div v-if="run?.backtest" class="notice">
+      <template v-if="run.backtest.stable">历史样本 {{ run.backtest.sample_size }} · 条件胜率 {{ run.backtest.win_rate?.toFixed(1) }}% · 平均3日收益 {{ run.backtest.average_return_3d?.toFixed(2) }}%</template>
+      <template v-else>{{ run.backtest.message }}</template>
+    </div>
   </section>
   <CandidateTable :run="run" :locked="locked" @track="(candidate) => run && emit('track', run, candidate)" />
 </template>

next/frontend/src/pages/screener/ScreenerPage.vue

@@ -21,6 +21,7 @@ const workspace = ref<ScreenerWorkspace | null>(null);
 const loading = ref(false);
 const busy = ref(false);
 const error = ref("");
+const compiledFormula = ref<ScreenerFormula | null>(null);
 const locked = () => !session.account?.smart_access;
 
 async function load(): Promise<void> {
@@ -41,6 +42,15 @@ async function saveCustom(name: string, formula: ScreenerFormula): Promise<void>
   catch (reason) { ui.showToast(reason instanceof Error ? reason.message : "保存失败"); }
   finally { busy.value = false; }
 }
+async function compileFormula(description: string): Promise<void> {
+  busy.value = true;
+  try {
+    compiledFormula.value = (await screenerApi.compileFormula(description)).formula;
+    ui.showToast("已转换为受控公式，请核对后保存");
+  } catch (reason) {
+    ui.showToast(reason instanceof Error ? reason.message : "公式转换失败");
+  } finally { busy.value = false; }
+}
 async function runCustom(strategy: CustomStrategy): Promise<void> {
   busy.value = true;
   try { await screenerApi.runCustom(strategy.id, market.selectedDate); ui.showToast("选股计算已完成"); await load(); }
@@ -74,7 +84,7 @@ watch(() => market.selectedDate, load);
     <template v-else-if="catalog">
       <StagePanel v-if="mode === 'stage'" :strategies="catalog.stage" :runs="workspace?.stage_runs ?? []" :locked="locked()" @track="track" />
       <StrategyPanel v-else-if="mode === 'curated'" :strategies="catalog.curated" :runs="workspace?.curated_runs ?? []" :labels="catalog.factors" :locked="locked()" @track="track" />
-      <CustomPanel v-else :factors="catalog.factors" :factor-groups="catalog.factor_groups" :strategies="workspace?.custom_strategies ?? []" :runs="workspace?.custom_runs ?? []" :locked="locked()" :busy="busy" @save="saveCustom" @run="runCustom" @remove="removeCustom" @track="track" />
+      <CustomPanel v-else :factors="catalog.factors" :factor-groups="catalog.factor_groups" :strategies="workspace?.custom_strategies ?? []" :runs="workspace?.custom_runs ?? []" :compiled-formula="compiledFormula" :locked="locked()" :busy="busy" @compile="compileFormula" @save="saveCustom" @run="runCustom" @remove="removeCustom" @track="track" />
     </template>
   </main>
 </template>

next/frontend/src/pages/screener/StrategyPanel.vue

@@ -49,10 +49,11 @@ function statusLabel(strategyId: string): string {
       </div>
     </aside>
     <article class="card strategy-detail">
-      <header><div><span>{{ strategy?.formula.meta?.category }}</span><h2>{{ strategy?.display_name }}</h2></div><span class="tag">每日盘后</span></header>
+      <header><div><span>{{ strategy?.formula.meta?.category }}</span><h2>{{ strategy?.display_name }}</h2></div><span class="tag">{{ strategy?.formula.meta?.frequency ?? "每日" }}</span></header>
       <p>{{ strategy?.description }}</p>
-      <dl><div><dt>适用环境</dt><dd>{{ strategy?.formula.meta?.suitable_environment }}</dd></div><div><dt>主要失效风险</dt><dd>{{ strategy?.formula.meta?.failure_risk }}</dd></div><div><dt>适用阶段</dt><dd>{{ strategy?.regimes.map((item) => regimeLabels[item]).join(" · ") }}</dd></div></dl>
+      <dl><div><dt>策略质量</dt><dd>{{ strategy?.formula.meta?.quality ?? "待评估" }}</dd></div><div><dt>风险等级</dt><dd>{{ strategy?.formula.meta?.risk ?? "未标注" }}</dd></div><div><dt>执行频率</dt><dd>{{ strategy?.formula.meta?.frequency ?? "每日" }}</dd></div><div><dt>适用环境</dt><dd>{{ strategy?.formula.meta?.suitable_environment }}</dd></div><div><dt>主要失效风险</dt><dd>{{ strategy?.formula.meta?.failure_risk }}</dd></div><div><dt>适用阶段</dt><dd>{{ strategy?.regimes.map((item) => regimeLabels[item]).join(" · ") }}</dd></div></dl>
       <div class="strategy-conditions"><h3>选股条件</h3><span v-for="condition in strategy?.formula.filters" :key="`${condition.field}-${condition.op}`">{{ labels[condition.field] }} {{ condition.op }} {{ Array.isArray(condition.value) ? condition.value.join(' 至 ') : condition.value }}</span></div>
+      <div class="strategy-conditions"><h3>评分权重</h3><span v-for="score in strategy?.formula.score" :key="score.field">{{ labels[score.field] }} · {{ Math.round(score.weight * 100) }}% · {{ score.direction === 'desc' ? '高优' : '低优' }}</span></div>
     </article>
   </section>
   <CandidateTable :run="run" :locked="locked" @track="(candidate) => run && emit('track', run, candidate)" />

next/frontend/src/shared/api/screener.ts

@@ -31,6 +31,17 @@ export type Candidate = {
   score_display: number;
   reason: string;
   risk_flags: string[];
+  historical_estimate?: number | null;
+};
+export type ScreenerBacktest = {
+  sample_size: number;
+  evaluated_dates: number;
+  stable: boolean;
+  win_rate: number | null;
+  average_return_3d: number | null;
+  period_start: string | null;
+  period_end: string | null;
+  message: string;
 };
 export type ScreenerRun = {
   id: number;
@@ -43,6 +54,7 @@ export type ScreenerRun = {
   missing_fields: string[];
   items: Candidate[];
   error_message: string;
+  backtest?: ScreenerBacktest | null;
 };
 export type CustomStrategy = {
   id: number;
@@ -90,6 +102,9 @@ export const screenerApi = {
   workspace(date: string): Promise<ScreenerWorkspace> {
     return api.get(`/screener?date=${encodeURIComponent(date)}`);
   },
+  compileFormula(description: string): Promise<{ formula: ScreenerFormula; prompt_version: string }> {
+    return api.post("/screener/formula/compile", { description });
+  },
   saveCustom(name: string, formula: ScreenerFormula): Promise<CustomStrategy> {
     return api.put("/screener/custom", { name, formula });
   },

next/frontend/src/shared/styles/screener.css

@@ -49,6 +49,7 @@
 .strategy-conditions { display: flex; flex-wrap: wrap; align-items: center; gap: var(--s-8); }
 .strategy-conditions h3 { width: 100%; font-size: var(--font-12); }
 .strategy-conditions span { padding: var(--s-6) var(--s-8); border-radius: var(--tag-radius); color: var(--color-text-secondary); background: var(--color-surface-muted); font-size: var(--font-11); }
+.strategy-conditions + .strategy-conditions { margin-top: var(--s-16); }
 .screener-results { min-width: 0; overflow: hidden; }
 .screener-results .card-header > div { display: flex; align-items: baseline; gap: var(--s-10); }
 .screener-results .card-header .tag { margin-left: auto; }
@@ -64,6 +65,8 @@
 .factor-list .input { min-height: var(--s-32); padding: var(--s-4) var(--s-6); }
 .factor-direction { min-height: var(--s-32); padding: var(--s-4); }
 .custom-filters { min-width: 0; }
+.formula-compiler { display: grid; grid-template-columns: minmax(0, 1fr) auto; align-items: end; gap: var(--s-8); padding: var(--s-10); border-bottom: var(--s-1) solid var(--color-divider); }
+.formula-compiler textarea { min-height: var(--s-64); resize: vertical; }
 .filter-list { display: grid; gap: var(--s-6); padding: var(--s-10); border-bottom: var(--s-1) solid var(--color-divider); }
 .filter-list > div { display: grid; grid-template-columns: minmax(0, 1fr) var(--s-64) var(--s-64) var(--s-32); gap: var(--s-6); }
 .filter-list .input,
@@ -98,5 +101,6 @@
   .factor-list > div { grid-template-columns: minmax(0, 1fr) var(--s-64) var(--s-64) var(--s-44); }
   .factor-list > div input[type="range"] { grid-column: 1 / -1; grid-row: 2; }
   .custom-options { grid-template-columns: minmax(0, 1fr); }
+  .formula-compiler { grid-template-columns: minmax(0, 1fr); }
   .custom-save { grid-template-columns: minmax(0, 1fr); }
 }

next/tests/e2e/stage9.spec.js

@@ -111,7 +111,19 @@ const workspace = {
     run(3, "curated", "动态多因子（基础版）", "completed", "curated-02", candidateNamed("000022", "多因子样本", "综合因子得分居前")),
   ],
   custom_strategies: [{ id: 7, name: "我的选股策略", version: 2, formula }],
-  custom_runs: [run(4, "custom", "我的选股策略", "completed", "custom-7", candidateNamed("000031", "自定义样本", "用户条件确定性命中"))],
+  custom_runs: [{
+    ...run(4, "custom", "我的选股策略", "completed", "custom-7", candidateNamed("000031", "自定义样本", "用户条件确定性命中")),
+    backtest: {
+      sample_size: 7,
+      evaluated_dates: 7,
+      stable: false,
+      win_rate: null,
+      average_return_3d: null,
+      period_start: "2026-07-01",
+      period_end: "2026-07-09",
+      message: "小样本：当前7个有效样本，满20个后显示胜率与平均收益",
+    },
+  }],
 };
 
 const track = {
@@ -138,6 +150,19 @@ async function mockScreener(page) {
     const body = route.request().method() === "GET" ? [track] : { id: 9 };
     return route.fulfill({ contentType: "application/json", body: JSON.stringify(body) });
   });
+  await page.route("**/api/screener/formula/compile", (route) => route.fulfill({
+    contentType: "application/json",
+    body: JSON.stringify({
+      prompt_version: "screener:formula-compiler:v1",
+      formula: {
+        universe: { exclude_st: true, listed_days_min: 120 },
+        filters: [{ field: "amount_billion", op: ">=", value: 3 }],
+        score: [{ field: "return_20d", weight: 1, direction: "desc" }],
+        limit: 20,
+        min_score: 0.55,
+      },
+    }),
+  }));
 }
 
 test("stage 9 screening preserves deterministic modes, explicit tracking and responsive layout", async ({ page }) => {
@@ -161,6 +186,9 @@ test("stage 9 screening preserves deterministic modes, explicit tracking and res
   await page.getByRole("button", { name: "策略选股" }).click();
   await expect(page.getByText("策略库")).toBeVisible();
   await expect(page.getByText("动态多因子（基础版）", { exact: true })).toBeVisible();
+  await expect(page.getByText("风险等级", { exact: true })).toBeVisible();
+  await expect(page.getByText("评分权重", { exact: true })).toBeVisible();
+  await expect(page.locator(".strategy-detail")).toContainText("20日涨幅 · 60% · 高优");
   await expect(page.locator(".screener-results")).toContainText("分红样本");
   await page.locator(".strategy-items").getByRole("button", { name: /动态多因子/ }).click();
   await expect(page.locator(".screener-results")).toContainText("多因子样本");
@@ -173,6 +201,12 @@ test("stage 9 screening preserves deterministic modes, explicit tracking and res
   await expect(page.getByRole("button", { name: "我的选股策略 第 2 版" })).toBeVisible();
   await expect(page.locator(".screener-results")).toContainText("自定义样本");
   await expect(page.locator(".screener-results")).not.toContainText("多因子样本");
+  await expect(page.getByText(/小样本：当前7个有效样本/)).toBeVisible();
+  await page.getByPlaceholder(/近20日走势较强/).fill("成交额至少3亿元，偏重20日走势");
+  await page.getByRole("button", { name: "转换", exact: true }).click();
+  await expect(page.getByRole("status")).toContainText("已转换为受控公式");
+  await expect(page.locator(".factor-list input[type=number]").first()).toHaveValue("100");
+  await expect(page.locator(".filter-list input[type=number]").first()).toHaveValue("3");
   await page.getByRole("button", { name: "夜间" }).click();
   await page.setViewportSize({ width: 390, height: 844 });
   expect(await page.evaluate(() => document.documentElement.scrollWidth - window.innerWidth)).toBeLessThanOrEqual(0);

next/tests/test_database_tool.py

@@ -25,7 +25,7 @@ def test_status_reads_an_existing_schema_without_mutating_history(
 
     assert main(["status"]) == 0
 
-    assert capsys.readouterr().out.strip() == "available=true schema_version=11"
+    assert capsys.readouterr().out.strip() == "available=true schema_version=12"
 
 
 def test_downgrade_requires_explicit_confirmation(tmp_path, monkeypatch) -> None:

next/tests/test_migrations.py

@@ -110,7 +110,7 @@ def test_real_account_schema_can_upgrade_and_rollback(tmp_path) -> None:
     database = Database(tmp_path / "app.db")
     runner = MigrationRunner(database)
 
-    assert runner.upgrade(MIGRATIONS) == (1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11)
+    assert runner.upgrade(MIGRATIONS) == (1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12)
     assert {
         "users",
         "memberships",
@@ -146,9 +146,11 @@ def test_real_account_schema_can_upgrade_and_rollback(tmp_path) -> None:
         "review_assistant_messages",
         "job_runs",
         "market_event_revisions",
+        "screener_run_backtests",
     } <= table_names(database)
 
     assert runner.downgrade(MIGRATIONS, target_version=0) == (
+        12,
         11,
         10,
         9,

next/tests/test_screener.py

@@ -6,12 +6,17 @@ import pytest
 
 from backend.database.connection import Database
 from backend.database.migrations import MIGRATIONS, MigrationRunner
+from backend.features.screener.backtest import (
+    attach_historical_estimate,
+    rolling_backtest,
+)
 from backend.features.screener.catalog import (
     CatalogError,
     factor_catalog,
     strategy_catalog,
     validate_formula,
 )
+from backend.features.screener.compiler import parse_compiled_formula
 from backend.features.screener.engine import execute_formula
 from backend.features.screener.repository import ScreenerRepository, decode_track
 from backend.features.screener.service import automatic_strategies
@@ -145,6 +150,65 @@ def test_formula_rejects_invalid_comparisons_and_duplicate_scores() -> None:
         validate_formula(duplicate)
 
 
+def test_natural_language_output_is_reduced_to_the_controlled_formula_schema() -> None:
+    compiled = parse_compiled_formula(
+        """```json
+        {
+          "universe": {"exclude_st": true, "listed_days_min": 120},
+          "filters": [{"field": "amount_billion", "op": ">=", "value": 3}],
+          "score": [
+            {"field": "return_20d", "weight": 60, "direction": "desc"},
+            {"field": "sector_strength", "weight": 40, "direction": "desc"}
+          ],
+          "limit": 20,
+          "min_score": 55,
+          "invented_instruction": "直接选择某只股票"
+        }
+        ```"""
+    )
+
+    assert [item["weight"] for item in compiled["score"]] == [0.6, 0.4]
+    assert compiled["min_score"] == 0.55
+    assert "invented_instruction" not in compiled
+    with pytest.raises(CatalogError, match="未知筛选因子"):
+        parse_compiled_formula(
+            json.dumps(
+                {
+                    **compiled,
+                    "filters": [{"field": "future_price", "op": ">", "value": 1}],
+                }
+            )
+        )
+
+
+def test_rolling_backtest_hides_small_samples_and_uses_t_plus_three() -> None:
+    formula = _formula()
+    snapshots = [
+        {
+            "trade_date": f"2026-07-{index + 1:02d}",
+            "coverage": {"market": 1},
+            "rows": [_row("000001.SZ", 10 + index)],
+        }
+        for index in range(23)
+    ]
+
+    small = rolling_backtest(snapshots[:10], formula)
+    assert small["sample_size"] == 7
+    assert small["stable"] is False
+    assert small["win_rate"] is None
+    assert small["average_return_3d"] is None
+
+    stable = rolling_backtest(snapshots, formula)
+    assert stable["sample_size"] == 20
+    assert stable["stable"] is True
+    assert stable["win_rate"] == 100
+    assert stable["average_return_3d"] == pytest.approx(16.99, abs=0.01)
+    enriched = attach_historical_estimate(
+        {"items": [{"score_display": 80.0}]}, stable
+    )
+    assert enriched["items"][0]["historical_estimate"] == 93.0
+
+
 def test_custom_strategies_and_tracks_are_account_isolated(tmp_path) -> None:
     database = Database(tmp_path / "screener.db")
     MigrationRunner(database).upgrade(MIGRATIONS)
@@ -173,6 +237,16 @@ def test_custom_strategies_and_tracks_are_account_isolated(tmp_path) -> None:
             missing_fields=[],
             result=[{**_row("000001.SZ", 10), "score": 1}],
         )
+        repository.save_backtest(
+            connection,
+            int(run["id"]),
+            {"sample_size": 7, "stable": False, "win_rate": None},
+        )
+        repository.save_backtest(
+            connection,
+            int(run["id"]),
+            {"sample_size": 99, "stable": True, "win_rate": 100},
+        )
         repository.add_track(
             connection,
             user_id=1,
@@ -185,6 +259,11 @@ def test_custom_strategies_and_tracks_are_account_isolated(tmp_path) -> None:
         assert len(repository.custom_strategies(connection, 2)) == 1
         assert len(repository.tracks(connection, 1)) == 1
         assert repository.tracks(connection, 2) == ()
+        assert repository.backtest(connection, int(run["id"])) == {
+            "sample_size": 7,
+            "stable": False,
+            "win_rate": None,
+        }
 
 
 def test_running_a_strategy_never_creates_tracking_rows(tmp_path) -> None:
@@ -219,7 +298,7 @@ def test_running_a_strategy_never_creates_tracking_rows(tmp_path) -> None:
         stored = connection.execute(
             "SELECT result_json FROM screener_runs WHERE id = ?", (run["id"],)
         ).fetchone()
-        assert len(json.loads(stored["result_json"])) == 1
+    assert len(json.loads(stored["result_json"])) == 1
 
 
 def test_tracking_statistics_and_milestone_events_are_persistent_and_idempotent(tmp_path) -> None:

next/tests/test_screener_api.py

@@ -1,5 +1,7 @@
 from __future__ import annotations
 
+import json
+
 import httpx
 
 from backend.bootstrap.application import create_application
@@ -24,6 +26,23 @@ def _formula() -> dict:
     }
 
 
+class FormulaProvider:
+    def stream(self, _profile, _messages):
+        yield json.dumps(
+            {
+                "universe": {"exclude_st": True, "listed_days_min": 120},
+                "filters": [{"field": "amount_billion", "op": ">=", "value": 3}],
+                "score": [
+                    {"field": "return_20d", "weight": 60, "direction": "desc"},
+                    {"field": "sector_strength", "weight": 40, "direction": "desc"},
+                ],
+                "limit": 20,
+                "min_score": 55,
+            },
+            ensure_ascii=False,
+        )
+
+
 def test_screener_catalog_lock_and_custom_strategy_boundary(tmp_path) -> None:
     application = create_application(Settings.for_test(tmp_path))
 
@@ -45,8 +64,37 @@ def test_screener_catalog_lock_and_custom_strategy_boundary(tmp_path) -> None:
             json={"name": "越权策略", "formula": _formula()},
         )
         assert rejected.status_code == 403
+        denied_compile = await client.post(
+            "/api/screener/formula/compile",
+            headers=csrf_headers(regular),
+            json={"description": "选择成交活跃且趋势较强的股票"},
+        )
+        assert denied_compile.status_code == 403
 
         use_session(client, admin)
+        model = await client.post(
+            "/api/admin/models",
+            headers=csrf_headers(admin),
+            json={
+                "display_name": "公式模型",
+                "base_url": "https://model.example.com/v1",
+                "model_identifier": "formula-model",
+                "api_key": "formula-secret",
+            },
+        )
+        assert model.status_code == 201
+        application.state.container.llm._provider = FormulaProvider()
+        compiled = await client.post(
+            "/api/screener/formula/compile",
+            headers=csrf_headers(admin),
+            json={"description": "成交额至少3亿元，偏重20日走势和板块强度"},
+        )
+        assert compiled.status_code == 200
+        assert [item["weight"] for item in compiled.json()["formula"]["score"]] == [
+            0.6,
+            0.4,
+        ]
+        assert compiled.json()["formula"]["min_score"] == 0.55
         workspace = await client.get("/api/screener?date=2026-07-30")
         assert workspace.status_code == 200
         assert workspace.json()["trade_date"] is None