feat(sentiment): harden news layer -- market-flow filter, evidence floor, event-override off by default

Follow-up to the 2026-06-15 M27 sentiment IC diagnosis. The event_taxonomy
0.65 override was measured net-negative for IC across every variant tested
(event delta -0.0027 / -0.0146 / -0.0280), and noisy non-company headlines
drove individual-stock whipsaws (兆易 603986: 2 sector-IPO headlines, 0 company
news -> -71.3 sentiment).

Changes:
- event_taxonomy: market/fund-flow/list headlines are filtered before
  classification so they cannot misfire as hard corporate events (for example,
  "下滑" inside "市场下滑"). Adds is_market_flow / is_company_specific /
  company_specific_titles helpers.
- event_score: the 0.65 override blend is now gated by
  settings.sentiment_event_override_enabled (default OFF). Matched event types
  are still reported for observability; raw sentiment is kept. Re-enableable for
  a clean single-provider OOS re-test.
- sentiment.analyze_news: evidence floor returns neutral 0 and skips the LLM
  when a window has no company-specific headline. Mixed windows now send only
  company-specific titles into the LLM prompt/cache/event scoring.
- postmarket and backtest news-cache paths pass stock name+symbol aliases when
  available, so cross-company contamination filtering is not production-only.
- m27_alpha_diagnostic --event-ab forces the override ON so it keeps measuring
  the override's counterfactual IC effect independent of the production default.

Verification:
- PYTHONPATH=. pytest -q tests/test_news_sentiment_pack.py tests/test_m27_alpha_event_universe.py tests/test_m27_m28_integration.py tests/test_tavily_news.py tests/test_sentiment_cache.py tests/test_llm_runtime_provider.py
  -> 46 passed / 1 skipped
- make verify PYTHON=/Users/zeeechenn/mingcang/.venv/bin/python
  -> ruff clean, mypy clean, backend pytest 1224 passed / 6 skipped; local
     frontend step not runnable in this linked worktree because frontend
     node_modules is absent
- GitHub Actions on PR #20 after force-push -> backend lint/typecheck, backend
  tests, frontend test/build, and security/dependency audit all passed for both
  push and pull_request runs

This is a defensive change (removes a measured drag + adds robustness); it does
not manufacture new sentiment alpha. Production adoption for live test2 signals
should wait for a clean single-provider OOS confirmation (epoch reset).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: Zeeechenn

backend/analysis/event_taxonomy.py

@@ -27,12 +27,71 @@ class EventType:
 )
 EVENT_TAXONOMY = EVENT_TYPES
 
+# Weight of the hard-event polarity in the event_override blend (rest goes to raw
+# LLM sentiment). Module constant so it can be tuned/measured via m27 --event-ab.
+EVENT_OVERRIDE_WEIGHT = 0.65
+
+# Market/flow/list-roundup headlines that are NOT company-specific corporate events.
+# These must never trigger a strong event_override; they were empirically dragging
+# sentiment IC down (M27 diagnosis 2026-06-15: event delta -0.0146). Detected first
+# and excluded from hard-event classification so an incidental polarity keyword
+# (e.g. "下滑" inside "市场下滑") cannot misfire as an earnings_warning.
+MARKET_FLOW_KEYWORDS: tuple[str, ...] = (
+    "资金流", "主力净", "净流入", "净流出", "大宗交易", "杠杆资金", "两融",
+    "千股千评", "涨幅榜", "跌幅榜", "龙虎榜", "北向资金", "南向资金", "成交额",
+    "特大单", "主力动向", "基金浮亏", "基金浮盈", "板块异动", "概念涨", "概念跌",
+)
+# Roundup / list headlines tagged loosely to many symbols ("XX股一览", "7只个股…").
+_LIST_HEADLINE_KEYWORDS: tuple[str, ...] = (
+    "一览", "名单", "股大", "只个股", "多股", "股获", "股名单",
+)
+
+
+def is_market_flow(title: str) -> bool:
+    """True if a headline is market/fund-flow/list noise, not a company-specific event."""
+    text = str(title)
+    if any(k in text for k in MARKET_FLOW_KEYWORDS):
+        return True
+    return any(k in text for k in _LIST_HEADLINE_KEYWORDS)
+
+
+def is_company_specific(
+    title: str, company_aliases: list[str] | tuple[str, ...] | None = None
+) -> bool:
+    """True if a headline plausibly carries company-specific signal.
+
+    Always rejects market/flow/list noise. When ``company_aliases`` (e.g. the stock
+    name + code) are given, the headline must also mention one of them — this is what
+    catches cross-company contamination (a 长鑫-IPO headline loosely tagged to 兆易).
+    """
+    text = str(title)
+    if is_market_flow(text):
+        return False
+    aliases = [str(a) for a in (company_aliases or []) if a]
+    if aliases and not any(a in text for a in aliases):
+        return False
+    return True
+
+
+def company_specific_titles(
+    titles: list[str] | tuple[str, ...],
+    company_aliases: list[str] | tuple[str, ...] | None = None,
+) -> list[str]:
+    """Keep only headlines that are company-specific (see ``is_company_specific``)."""
+    return [str(t) for t in titles if is_company_specific(t, company_aliases)]
+
 
 def classify_events(events: list[str] | tuple[str, ...]) -> list[dict]:
-    """Classify event strings into the local A-share taxonomy."""
+    """Classify event strings into the local A-share taxonomy.
+
+    Market/flow/list noise is skipped up front so it cannot be misclassified as a
+    hard corporate event via an incidental keyword match.
+    """
     classified: list[dict] = []
     for event in events:
         text = str(event)
+        if is_market_flow(text):
+            continue
         for event_type in EVENT_TYPES:
             if any(keyword in text for keyword in event_type.keywords):
                 classified.append({
@@ -69,32 +128,58 @@ def build_event_extraction_prompt(symbol: str | None, titles: list[str]) -> str:
     )
 
 
-def event_score(sentiment: float, events: list[str] | tuple[str, ...]) -> dict:
-    """Return an event-aware score in [-1, 1], falling back to sentiment when no event matches."""
+def event_score(
+    sentiment: float,
+    events: list[str] | tuple[str, ...],
+    *,
+    enable_override: bool | None = None,
+) -> dict:
+    """Return an event-aware score in [-1, 1].
+
+    The event_override blend is gated by ``enable_override`` (defaults to
+    ``settings.sentiment_event_override_enabled``, which is OFF). When disabled, the
+    raw sentiment is kept (mode ``sentiment_fallback``) but matched event types are
+    still reported for observability. The override was measured net-negative for IC
+    (M27 diagnosis 2026-06-15) — keep it off unless a clean OOS re-test earns it back.
+    """
     clipped_sentiment = max(-1.0, min(1.0, float(sentiment or 0.0)))
     classified = classify_events(events)
-    if not classified:
+    if enable_override is None:
+        try:
+            from backend.config import settings
+
+            enable_override = bool(settings.sentiment_event_override_enabled)
+        except Exception:
+            enable_override = False
+    if not classified or not enable_override:
         return {
             "event_score": clipped_sentiment,
             "event_score_mode": "sentiment_fallback",
-            "event_types": [],
+            "event_types": classified,
         }
 
     polarity_avg = sum(item["polarity"] for item in classified) / len(classified)
-    score = 0.65 * polarity_avg + 0.35 * clipped_sentiment
+    score = EVENT_OVERRIDE_WEIGHT * polarity_avg + (1.0 - EVENT_OVERRIDE_WEIGHT) * clipped_sentiment
     return {
         "event_score": round(max(-1.0, min(1.0, score)), 4),
         "event_score_mode": "event_override",
         "event_types": classified,
     }
 
 
-def apply_event_score(sentiment_result: dict[str, Any], titles: list[str] | None = None) -> dict[str, Any]:
+def apply_event_score(
+    sentiment_result: dict[str, Any],
+    titles: list[str] | None = None,
+    *,
+    enable_override: bool | None = None,
+) -> dict[str, Any]:
     """Add event-aware scoring to an analyze_news-style result."""
     out = dict(sentiment_result)
     event_texts: list[str] = []
     if titles:
         event_texts.extend(titles)
     event_texts.extend(str(item) for item in out.get("key_events", []) or [])
-    out.update(event_score(float(out.get("sentiment") or 0.0), event_texts))
+    out.update(
+        event_score(float(out.get("sentiment") or 0.0), event_texts, enable_override=enable_override)
+    )
     return out

backend/analysis/sentiment.py

@@ -4,7 +4,7 @@
 from collections import OrderedDict
 from datetime import UTC, datetime
 
-from backend.analysis.event_taxonomy import apply_event_score
+from backend.analysis.event_taxonomy import apply_event_score, company_specific_titles
 from backend.config import settings
 from backend.llm import get_provider, has_runtime_llm_provider
 
@@ -57,6 +57,20 @@
     "event_score_mode": "sentiment_fallback",
     "event_types": [],
 }
+# Evidence floor: when a symbol's window has no company-specific headline (only
+# sector/fund-flow/list noise), neutralise sentiment instead of scoring noise.
+# Prevents the M27 whipsaw (e.g. 兆易 603986: 2 sector-IPO headlines, 0 company
+# news → -71.3 sentiment). Skips the LLM call entirely, so it is also cheaper.
+_NOISE_ONLY_FALLBACK = {
+    "sentiment": 0.0,
+    "summary": "仅板块/资金流新闻，无个股特定信号，情感置中",
+    "impact": "short",
+    "key_events": [],
+    "low_confidence": True,
+    "event_score": 0.0,
+    "event_score_mode": "evidence_floor",
+    "event_types": [],
+}
 
 
 def _titles_hash(titles: list[str]) -> str:
@@ -158,28 +172,37 @@ def _persistent_cache_set(key: str, titles_hash: str, symbol: str | None, value:
         return
 
 
-def analyze_news(titles: list[str], symbol: str | None = None) -> dict:
+def analyze_news(
+    titles: list[str],
+    symbol: str | None = None,
+    company_aliases: list[str] | None = None,
+) -> dict:
     """
     输入新闻标题列表，返回情感分析结果。
     {sentiment: float, summary: str, impact: str, key_events: list}
     相同股票+标题集合优先命中进程内和 SQLite 缓存，避免重复消耗 LLM 配额。
+    company_aliases（股票名/代码等别名）用于个股相关性过滤与证据地板；不传则只过滤行情/资金流噪声。
     """
     if not titles:
         return apply_event_score(_FALLBACK, [])
     if not has_runtime_llm_provider(settings):
         return apply_event_score(_DISABLED_FALLBACK, titles)
+    # Evidence floor: no company-specific headline (only noise / cross-company) → neutral, skip LLM.
+    analysis_titles = company_specific_titles(titles, company_aliases)
+    if not analysis_titles:
+        return dict(_NOISE_ONLY_FALLBACK)
 
-    cache_key, titles_hash = _cache_key(titles, symbol)
+    cache_key, titles_hash = _cache_key(analysis_titles, symbol)
     cached = _cache_get(cache_key)
     if cached is not None:
-        return apply_event_score(cached, titles)
+        return apply_event_score(cached, analysis_titles)
     cached = _persistent_cache_get(cache_key)
     if cached is not None:
         _cache_set(cache_key, cached)
-        return apply_event_score(cached, titles)
+        return apply_event_score(cached, analysis_titles)
 
     context = f"股票代码：{symbol}\n" if symbol else ""
-    prompt = context + "新闻标题：\n" + "\n".join(f"- {t}" for t in titles[:15])
+    prompt = context + "新闻标题：\n" + "\n".join(f"- {t}" for t in analysis_titles[:15])
 
     data = get_provider().complete_structured(
         prompt=prompt,
@@ -205,11 +228,11 @@ def analyze_news(titles: list[str], symbol: str | None = None) -> dict:
             "event_score": 0.0,
             "event_score_mode": "sentiment_fallback",
             "event_types": [],
-        }, titles)
+        }, analysis_titles)
 
     data["sentiment"] = max(-1.0, min(1.0, float(data.get("sentiment", 0))))
     data["key_events"] = data.get("key_events", [])[:3]
-    data = apply_event_score(data, titles)
+    data = apply_event_score(data, analysis_titles)
     _cache_set(cache_key, data)
     _persistent_cache_set(cache_key, titles_hash, symbol, data)
     return dict(data)

backend/backtest/news_cache.py

@@ -69,6 +69,16 @@ def _fetch_titles(symbol: str, date: str, db, lookback_days: int = 3) -> list[st
     return [r[0] for r in rows if r[0]]
 
 
+def _company_aliases_for_symbol(symbol: str, db) -> list[str] | None:
+    """Return name+symbol aliases when Stock metadata is available."""
+    from backend.data.database import Stock
+
+    stock = db.query(Stock).filter(Stock.symbol == symbol).first()
+    if stock is None or not stock.name:
+        return None
+    return [stock.name, symbol]
+
+
 def get_or_backfill(
     symbol: str,
     date: str,
@@ -105,7 +115,11 @@ def get_or_backfill(
         }
     else:
         from backend.analysis.sentiment import analyze_news
-        result = analyze_news(titles, symbol=symbol)
+        result = analyze_news(
+            titles,
+            symbol=symbol,
+            company_aliases=_company_aliases_for_symbol(symbol, db),
+        )
 
     # 写回缓存（即使是空结果也缓存，避免每次都重查 NewsItem）
     cache[key] = result
@@ -144,7 +158,11 @@ def backfill_all(
 
         try:
             from backend.analysis.sentiment import analyze_news
-            result = analyze_news(titles, symbol=s.symbol)
+            result = analyze_news(
+                titles,
+                symbol=s.symbol,
+                company_aliases=_company_aliases_for_symbol(s.symbol, db),
+            )
             if not result.get("key_events") and result.get("summary") == "解析失败":
                 stats["llm_fail"] += 1
             cache[key] = result

backend/config.py

@@ -106,6 +106,11 @@ class Settings(BaseSettings):
     weight_technical: float = 0.6
     weight_sentiment: float = 0.4
     new_framework_entry_threshold: float = 25.0
+    # Event-taxonomy override (0.65×事件极性 blend). Default OFF: the M27 sentiment IC
+    # diagnosis (2026-06-15) measured event delta consistently negative across every
+    # variant tested (-0.0027 / -0.0146 / -0.0280), i.e. the override only subtracts IC.
+    # Kept behind a flag so it can be re-enabled for a clean single-provider OOS re-test.
+    sentiment_event_override_enabled: bool = False
 
     # Signal profile: legacy Qlib framework or current new framework.
     paper_trading_profile: str = "auto"  # auto / test1_legacy_qlib / new_framework

backend/jobs/postmarket.py

@@ -148,7 +148,8 @@ def _postmarket_news_sentiment(stock, db) -> dict:
             logger.info("Tavily补充 %s: +%d条 (DB=%d条)",
                         stock.symbol, len(tavily_titles), len(titles) - len(tavily_titles))
 
-    sentiment_result = analyze_news(titles, symbol=stock.symbol)
+    company_aliases = [a for a in (stock.name, stock.symbol) if a]
+    sentiment_result = analyze_news(titles, symbol=stock.symbol, company_aliases=company_aliases)
     sentiment_result["news_audit"] = [
         {
             "title": audit.title,

backend/tools/m27_alpha_diagnostic.py

@@ -598,7 +598,9 @@ def _event_ab_row_frame(
         polarity_score = sentiment["polarity_score"]
         polarity_source = sentiment["polarity_source"]
         polarity_available = polarity_score is not None
-        adjusted = event_score(polarity_score or 0.0, titles)
+        # Force the override ON here: this A/B harness exists to measure the override's
+        # counterfactual IC effect, independent of the production default (which is OFF).
+        adjusted = event_score(polarity_score or 0.0, titles, enable_override=True)
         event_available = polarity_available or adjusted["event_score_mode"] == "event_override"
 
         ab_rows.append({

tests/test_m27_alpha_event_universe.py

@@ -32,8 +32,11 @@ def test_event_taxonomy_scores_events_and_falls_back_to_polarity():
     positive = apply_event_score(
         {"sentiment": -0.1, "key_events": []},
         ["公司公告获得重大合同并中标算力项目"],
+        enable_override=True,  # override is opt-in (default OFF per M27 IC diagnosis)
+    )
+    neutral = apply_event_score(
+        {"sentiment": 0.2, "key_events": []}, ["普通经营动态"], enable_override=True
     )
-    neutral = apply_event_score({"sentiment": 0.2, "key_events": []}, ["普通经营动态"])
 
     assert positive["event_score_mode"] == "event_override"
     assert positive["event_score"] > 0

tests/test_m27_m28_integration.py

@@ -30,6 +30,7 @@ def test_event_taxonomy_overrides_sentiment_when_event_matches():
     result = apply_event_score(
         {"sentiment": -0.2, "key_events": ["公司获得重大合同订单"]},
         [],
+        enable_override=True,  # override is opt-in (default OFF per M27 IC diagnosis)
     )
 
     assert result["event_score_mode"] == "event_override"