Record news signals forward for a no-trade edge test

The trader labels crypto news published after 2026-10-16 15:00 with
the configured model (category and BTC direction) and, four hours
later, backfills BTC/ETH 1m opens at publish, +5m, +1h and +4h from
the klines table. A drawer on the trading page shows the cumulative
returns of trading the labelled direction, after fees. Nothing is
traded; this only collects an out-of-sample record.

backend/src/alphapulse/api/trading.py

@@ -17,11 +17,13 @@ from alphapulse.api.market import call_binance, client as market_client
 from alphapulse.core.auth import require_admin
 from alphapulse.core.db import get_session
 from alphapulse.core.redis import KEY_PREFIX, redis
-from alphapulse.models import AiProvider, TradingAccount, TradingAnalysis, TradingEvent, TradingFill, TradingOrder, TradingSymbol
+from alphapulse.models import AiProvider, NewsItem, NewsSignal, TradingAccount, TradingAnalysis, TradingEvent, TradingFill, TradingOrder, TradingSymbol
 from alphapulse.trading.client import OWN_CLIENT_ID, ExchangeError, public_get
 from alphapulse.trading.executor import MANUAL_ACTION_KEY
+from alphapulse.trading.news_signals import FORWARD_START, SYMBOLS as SIGNAL_SYMBOLS, news_text
 from alphapulse.trading.positions import build_positions
 from alphapulse.trading.plan import DAILY_ANALYSIS_LIMIT
+from alphapulse.trading.signal_stats import Signal, summarize
 from alphapulse.trading.snapshot import read_snapshot
 
 router = APIRouter(prefix="/api/trading", tags=["trading"], dependencies=[Depends(require_admin)])
@@ -179,6 +181,37 @@ class PositionHistoryPage(BaseModel):
     items: list[PositionHistoryItem]
 
 
+class NewsSignalStat(BaseModel):
+    category: str
+    symbol: str
+    horizon: str
+    n: int
+    mean_bp: float | None
+    t: float | None
+    # 扣除来回手续费后的平均收益
+    net_bp: float | None
+    win_rate: float | None
+
+
+class NewsSignalItem(BaseModel):
+    news_id: int
+    published_at: datetime
+    source_code: str
+    title: str
+    category: str
+    direction: str
+    # 进场后的原始涨跌（bp，未按方向取正负）：{symbol: {"1h": ..., "4h": ...}}；未满 4 小时为空
+    returns: dict[str, dict[str, float]] | None
+
+
+class NewsSignalsOut(BaseModel):
+    start: datetime
+    labeled: int
+    filled: int
+    stats: list[NewsSignalStat]
+    recent: list[NewsSignalItem]
+
+
 async def get_account(session: AsyncSession) -> TradingAccount:
     return await session.get(TradingAccount, 1) or TradingAccount(id=1)
 
@@ -454,6 +487,32 @@ async def list_position_history_symbols(session: Annotated[AsyncSession, Depends
     return list(await session.scalars(select(TradingFill.symbol).distinct().order_by(TradingFill.symbol)))
 
 
+@router.get("/news-signals")
+async def list_news_signals(session: Annotated[AsyncSession, Depends(get_session)]) -> NewsSignalsOut:
+    """新闻信号前向记录：按标注方向交易的累计统计和最近 50 条"""
+    rows = (await session.execute(
+        select(NewsSignal, NewsItem.published_at, NewsItem.source_code, news_text())
+        .join(NewsItem, NewsItem.id == NewsSignal.news_id).order_by(NewsItem.published_at)
+    )).all()
+    filled = [Signal(published_at, s.category, s.direction, s.prices) for s, published_at, _, _ in rows if s.prices]
+    horizons = ("1h", "4h")
+
+    def returns(prices: dict | None) -> dict[str, dict[str, float]] | None:
+        if prices is None:
+            return None
+        return {sym: {h: (float(p[h]) / float(p["entry"]) - 1) * 1e4 for h in horizons} for sym, p in prices.items()}
+
+    return NewsSignalsOut(
+        start=FORWARD_START, labeled=len(rows), filled=len(filled),
+        stats=[NewsSignalStat(**r) for r in summarize(filled, SIGNAL_SYMBOLS, horizons)],
+        recent=[
+            NewsSignalItem(news_id=s.news_id, published_at=published_at, source_code=source, title=title,
+                           category=s.category, direction=s.direction, returns=returns(s.prices))
+            for s, published_at, source, title in reversed(rows[-50:])
+        ],
+    )
+
+
 @router.post("/symbols/{symbol}/analyze", status_code=status.HTTP_202_ACCEPTED)
 async def request_analysis(symbol: str, session: Annotated[AsyncSession, Depends(get_session)]) -> None:
     """排队一次手动分析：该币停止交易时为试运行，计划只记录、不下单"""

backend/src/alphapulse/models/__init__.py

@@ -2,7 +2,7 @@ from alphapulse.core.db import Base
 from alphapulse.models.ai import AiProvider, ChatMessage, ChatSession, ContextVersion, SkillNode, SkillVersion
 from alphapulse.models.market import Kline, KlineSyncState, MarketSymbol
 from alphapulse.models.news import CrawlLog, CrawlRun, NewsItem, NewsRevision, RawBody, RawResponse, Source
-from alphapulse.models.trading import TradingAccount, TradingAnalysis, TradingEvent, TradingFill, TradingOrder, TradingSymbol
+from alphapulse.models.trading import NewsSignal, TradingAccount, TradingAnalysis, TradingEvent, TradingFill, TradingOrder, TradingSymbol
 
 __all__ = [
     "AiProvider",
@@ -16,6 +16,7 @@ __all__ = [
     "KlineSyncState",
     "MarketSymbol",
     "NewsItem",
+    "NewsSignal",
     "NewsRevision",
     "RawBody",
     "RawResponse",

backend/src/alphapulse/models/trading.py

@@ -145,3 +145,19 @@ class TradingFill(Base):
     commission: Mapped[Decimal] = mapped_column(Numeric)
     commission_asset: Mapped[str] = mapped_column(String(16))
     time: Mapped[datetime] = mapped_column(DateTime(timezone=True))
+
+
+class NewsSignal(Base):
+    """新闻信号前向记录：加密相关新闻的模型标注，以及发布后 BTC/ETH 的价格，检验按标注方向交易是否有优势。只记录，不下单"""
+
+    __tablename__ = "news_signals"
+
+    news_id: Mapped[int] = mapped_column(BigInteger, ForeignKey("news_items.id"), primary_key=True)
+    # price_report（行情播报）/ event（事件）/ irrelevant（与币价无关）
+    category: Mapped[str] = mapped_column(String(16))
+    # 对 BTC 短期（几小时）的影响：bullish / bearish / neutral
+    direction: Mapped[str] = mapped_column(String(8))
+    model: Mapped[str] = mapped_column(String(64))
+    labeled_at: Mapped[datetime] = mapped_column(DateTime(timezone=True), server_default=func.now())
+    # 发布 4 小时后回填的 1m 开盘价：{symbol: {"open": 发布所在分钟, "entry": +5 分钟, "1h": +65 分钟, "4h": +245 分钟}}；为空表示尚未回填
+    prices: Mapped[dict | None] = mapped_column(JSONB)

backend/src/alphapulse/trading/engine.py

@@ -1,4 +1,4 @@
-"""模拟盘交易进程：同步账户快照、执行器对账、调度分析。
+"""模拟盘交易进程：同步账户快照、执行器对账、调度分析、记录新闻信号。
 
 用法：python -m alphapulse.trading.engine
 """
@@ -13,6 +13,7 @@ from alphapulse.core.db import SessionLocal
 from alphapulse.models import TradingAnalysis, TradingSymbol
 from alphapulse.trading.analyst import run_analysis
 from alphapulse.trading.executor import executor_loop
+from alphapulse.trading.news_signals import news_signal_loop
 from alphapulse.trading.schedule import routine_requests
 from alphapulse.trading.snapshot import snapshot_loop
 
@@ -74,6 +75,7 @@ async def main() -> None:
         tg.create_task(snapshot_loop())
         tg.create_task(executor_loop())
         tg.create_task(analysis_loop())
+        tg.create_task(news_signal_loop())
 
 
 if __name__ == "__main__":

backend/src/alphapulse/trading/news_signals.py

@@ -0,0 +1,144 @@
+"""新闻信号前向记录：给新进来的加密相关新闻打标签（类别、对 BTC 的方向），发布 4 小时后回填 BTC/ETH 的价格。
+
+研究（2026-10-16，见 CONTEXT.md）：事件新闻按模型标注方向持有 4 小时，样本内约 +15bp 但不显著，且模型可能记得那段行情。
+前向记录只看标注时还没发生的行情，用来确认或否定这条线索。只记录，不下单。
+"""
+
+import asyncio
+import logging
+import re
+from datetime import UTC, datetime, timedelta
+
+from sqlalchemy import func, select
+from sqlalchemy.dialects.postgresql import insert
+
+from alphapulse.ai import llm
+from alphapulse.ai.tools import LOCAL_TZ
+from alphapulse.core.db import SessionLocal
+from alphapulse.models import AiProvider, Kline, NewsItem, NewsSignal, TradingAccount
+
+logger = logging.getLogger("alphapulse.trading.news_signals")
+
+# 研究样本截止到 2026-10-16 14:15，之后发布的新闻才计入，两份样本互不重叠
+FORWARD_START = datetime(2026, 10, 16, 15, 0, tzinfo=LOCAL_TZ)
+CRYPTO_PATTERN = "比特币|加密货币|以太坊|加密市场|数字货币|稳定币|Coinbase|微策略|币安|BTC"
+SYMBOLS = ("BTCUSDT", "ETHUSDT")
+# 发布所在分钟之后第几分钟的 1m 开盘价：发布后第 5 分钟进场（抓取加分析约 1–3 分钟），持有 1h、4h
+OFFSETS = {"open": 0, "entry": 5, "1h": 65, "4h": 245}
+LOOP_SECONDS = 60
+BATCH = 40
+CATEGORIES = {"price_report", "event", "irrelevant"}
+DIRECTIONS = {"bullish", "bearish", "neutral"}
+
+SYSTEM = """你是加密货币新闻分类员。只根据给出的新闻文字判断，不要使用你对之后行情的任何记忆。
+对每条新闻给出：
+- category：price_report（只是在报道价格已经发生的涨跌，如「比特币跌破 78000 美元」「比特币日内涨超 4%」，包括概念股跟随币价的涨跌）；
+  event（会影响币价的事件或信息，如机构增持/减持、ETF 资金流、监管政策、宏观政策对加密市场的影响、黑客攻击、交易所动态）；
+  irrelevant（与 BTC/ETH 价格基本无关，如个别公司财报细节、分析师调整目标价、技术合作）。
+- direction：对 BTC 价格短期（几小时内）的影响，bullish / bearish / neutral。price_report 按它报道的涨跌方向填。
+用 label 工具一次提交全部结果，id 与输入一一对应。"""
+TOOL = llm.ToolSpec("label", "提交分类结果", {
+    "type": "object",
+    "properties": {"items": {"type": "array", "items": {
+        "type": "object",
+        "properties": {
+            "id": {"type": "integer"},
+            "category": {"type": "string", "enum": sorted(CATEGORIES)},
+            "direction": {"type": "string", "enum": sorted(DIRECTIONS)},
+        },
+        "required": ["id", "category", "direction"],
+    }}},
+    "required": ["items"],
+})
+
+
+def news_text():
+    """快讯常常没有标题，取正文开头"""
+    return func.coalesce(func.nullif(NewsItem.title, ""), func.left(NewsItem.content_text, 120))
+
+
+def clean(text: str) -> str:
+    # 去掉「财联社10月7日电，」之类的日期前缀，减少模型联想到具体日期
+    return re.sub(r"^(财联社)?\d{1,2}月\d{1,2}日电[，,]?", "", text).strip()[:200]
+
+
+async def label_new() -> int:
+    """给还没打标签的加密新闻打标签，返回本轮标注条数。模型漏掉的条目下一轮重试"""
+    async with SessionLocal() as session:
+        rows = (await session.execute(
+            select(NewsItem.id, news_text()).outerjoin(NewsSignal, NewsSignal.news_id == NewsItem.id)
+            .where(NewsSignal.news_id.is_(None), NewsItem.published_at >= FORWARD_START, news_text().op("~")(CRYPTO_PATTERN))
+            .order_by(NewsItem.published_at).limit(BATCH)
+        )).all()
+        if not rows:
+            return 0
+        account = await session.get(TradingAccount, 1)
+        provider = await session.get(AiProvider, account.provider_id) if account and account.provider_id else await session.scalar(
+            select(AiProvider).where(AiProvider.is_default)
+        )
+        if provider is None:
+            raise RuntimeError("没有可用的模型配置，无法给新闻打标签")
+        session.expunge(provider)
+
+    completion = None
+    text = "\n".join(f"{news_id}\t{clean(t)}" for news_id, t in rows)
+    async for item in llm.stream(provider, SYSTEM, [{"role": "user", "text": text}], [TOOL], "low"):
+        if isinstance(item, llm.Completion):
+            completion = item
+    call = next((c for c in completion.tool_calls if c["name"] == "label"), None)
+    if call is None:
+        raise RuntimeError("模型没有调用 label 工具")
+    ids = {news_id for news_id, _ in rows}
+    values = [
+        {"news_id": x["id"], "category": x["category"], "direction": x["direction"], "model": provider.model}
+        for x in call["arguments"].get("items", [])
+        if x.get("id") in ids and x.get("category") in CATEGORIES and x.get("direction") in DIRECTIONS
+    ]
+    if len(values) < len(rows):
+        logger.warning("新闻标签：%d 条中模型只返回了 %d 条有效结果，其余下一轮重试", len(rows), len(values))
+    if values:
+        async with SessionLocal() as session:
+            await session.execute(insert(NewsSignal).values(values).on_conflict_do_nothing())
+            await session.commit()
+    return len(values)
+
+
+async def fill_prices() -> int:
+    """发布满 4 小时的信号回填价格，返回回填条数。1m K 线有缺口时留到下一轮"""
+    cutoff = datetime.now(UTC) - timedelta(minutes=OFFSETS["4h"] + 2)
+    filled = 0
+    async with SessionLocal() as session:
+        pending = (await session.execute(
+            select(NewsSignal, NewsItem.published_at).join(NewsItem, NewsItem.id == NewsSignal.news_id)
+            .where(NewsSignal.prices.is_(None), NewsItem.published_at <= cutoff)
+        )).all()
+        for signal, published_at in pending:
+            minute = int(published_at.timestamp()) // 60 * 60_000
+            times = {key: minute + offset * 60_000 for key, offset in OFFSETS.items()}
+            prices = {}
+            for symbol in SYMBOLS:
+                opens = dict((await session.execute(
+                    select(Kline.open_time, Kline.open).where(
+                        Kline.market == "um", Kline.symbol == symbol, Kline.interval == "1m", Kline.open_time.in_(times.values()),
+                    )
+                )).tuples().all())
+                if len(opens) < len(times):
+                    break
+                prices[symbol] = {key: str(opens[t]) for key, t in times.items()}
+            else:
+                signal.prices = prices
+                filled += 1
+        await session.commit()
+    return filled
+
+
+async def news_signal_loop() -> None:
+    while True:
+        try:
+            labeled = await label_new()
+            filled = await fill_prices()
+            if labeled or filled:
+                logger.info("新闻信号：标注 %d 条，回填价格 %d 条", labeled, filled)
+        except Exception:
+            logger.exception("新闻信号记录出错，稍后重试")
+        await asyncio.sleep(LOOP_SECONDS)

backend/src/alphapulse/trading/signal_stats.py

@@ -0,0 +1,54 @@
+"""新闻信号的统计（纯函数，便于测试）：按模型标注方向交易的收益，同一件事的多条报道只算一次"""
+
+import math
+from dataclasses import dataclass
+from datetime import datetime, timedelta
+
+# 多个来源报道同一件事：同一类别 10 分钟内只算第一条
+DEDUPE = timedelta(minutes=10)
+# 吃单进出各 0.04%
+FEE_BP = 8.0
+DIRECTION_SIGN = {"bullish": 1, "bearish": -1}
+
+
+@dataclass
+class Signal:
+    published_at: datetime
+    category: str
+    direction: str
+    # {symbol: {"entry": ..., "1h": ..., "4h": ...}}，价格为十进制字符串
+    prices: dict
+
+
+def dedupe(signals: list[Signal]) -> list[Signal]:
+    """signals 按发布时间排序"""
+    last: dict[str, datetime] = {}
+    kept = []
+    for s in signals:
+        if s.category not in last or s.published_at - last[s.category] > DEDUPE:
+            kept.append(s)
+            last[s.category] = s.published_at
+    return kept
+
+
+def summarize(signals: list[Signal], symbols: tuple[str, ...], horizons: tuple[str, ...]) -> list[dict]:
+    """每个类别、币、持有时长：利多做多、利空做空的收益（bp），中性不交易"""
+    rows = []
+    kept = dedupe(signals)
+    for category in sorted({s.category for s in kept}):
+        for symbol in symbols:
+            for horizon in horizons:
+                pnl = [
+                    DIRECTION_SIGN[s.direction] * (float(s.prices[symbol][horizon]) / float(s.prices[symbol]["entry"]) - 1) * 1e4
+                    for s in kept if s.category == category and s.direction in DIRECTION_SIGN
+                ]
+                n = len(pnl)
+                mean = sum(pnl) / n if n else None
+                sd = math.sqrt(sum((x - mean) ** 2 for x in pnl) / (n - 1)) if n > 1 else None
+                rows.append({
+                    "category": category, "symbol": symbol, "horizon": horizon, "n": n,
+                    "mean_bp": mean, "t": mean / (sd / math.sqrt(n)) if sd else None,
+                    "net_bp": mean - FEE_BP if mean is not None else None,
+                    "win_rate": sum(x > 0 for x in pnl) / n if n else None,
+                })
+    return rows