fix(ai_news): set weight=0 and add Spearman IC validation harness

검증 전 gradient 차단 + IC 측정 인프라.

- schema.py: DEFAULT_WEIGHTS["ai_news"] 0.8 → 0.0
  + 1회성 migration: 기존 운영 row 의 0.8 값 자동 reset
  (사용자가 명시 조정한 다른 값은 그대로 유지)
- ai_news/validation.py: compute_ic() — 일자별 score_raw × forward
  return Spearman 상관, ic_mean/ic_std/ic_per_day 반환, verdict 분류
  (skip/weak/strong)
- router.py: GET /api/stock/screener/ai-news/ic?days=30&horizon=1
- 단위 테스트 5개: empty DB, strong +IC, random ≈0 IC, min_news_count
  필터, horizon=5

배경: adversarial review 결과 — ai_news 가중치 0.8 이 검증 없이 출시됨.
4주+ 데이터 누적 후 IC > 0.05 확인 전까지 데이터 수집은 계속하되
가중합 영향만 차단. 운영 DB row 의 0.8 → 0.0 자동 reset 도 같은 의도.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

stock-lab/app/screener/ai_news/validation.py

@@ -0,0 +1,125 @@
+"""AI news sentiment validation — Spearman IC vs forward returns.
+
+핵심 metric: 일자별 score_raw 와 다음 N일 forward return 의 Spearman 상관.
+4주+ 누적 후 IC mean > 0.05 면 weight 활성화 가치 있음.
+"""
+
+from __future__ import annotations
+
+import datetime as dt
+import sqlite3
+from typing import Any, Dict, List, Optional
+
+import pandas as pd
+
+
+def _spearman(a: pd.Series, b: pd.Series) -> Optional[float]:
+    """Spearman rank correlation. None if insufficient/degenerate data."""
+    if len(a) < 5 or len(b) < 5:
+        return None
+    if a.std(ddof=0) == 0 or b.std(ddof=0) == 0:
+        return None
+    return float(a.rank().corr(b.rank()))
+
+
+def compute_ic(
+    conn: sqlite3.Connection,
+    *,
+    days: int = 30,
+    horizon: int = 1,
+    min_news_count: int = 1,
+    asof_today: Optional[dt.date] = None,
+) -> Dict[str, Any]:
+    """Compute daily Spearman IC of ai_news.score_raw vs forward return.
+
+    Returns:
+        {
+          "horizon_days": int,
+          "min_news_count": int,
+          "window_days": int,
+          "ic_count": int,           # 유효 일수
+          "ic_mean": float | None,
+          "ic_std":  float | None,
+          "ic_per_day": [{"date": "YYYY-MM-DD", "ic": float, "n": int}, ...],
+          "verdict": "skip" | "weak" | "strong",
+        }
+
+    verdict:
+      - skip: ic_count < 10
+      - weak: ic_mean in [-0.05, 0.05]
+      - strong: |ic_mean| > 0.05
+    """
+    asof_today = asof_today or dt.date.today()
+    cutoff = (asof_today - dt.timedelta(days=days)).isoformat()
+
+    sentiment = pd.read_sql_query(
+        "SELECT ticker, date, score_raw, news_count "
+        "FROM news_sentiment WHERE date >= ? AND news_count >= ? ORDER BY date",
+        conn, params=(cutoff, min_news_count),
+    )
+    if sentiment.empty:
+        return _empty_result(days, horizon, min_news_count)
+
+    # forward return 조회: 각 (ticker, date) 에 대해 close[date+horizon] / close[date] - 1
+    prices = pd.read_sql_query(
+        "SELECT ticker, date, close FROM krx_daily_prices "
+        "WHERE date >= ? ORDER BY ticker, date",
+        conn, params=(cutoff,),
+    )
+    if prices.empty:
+        return _empty_result(days, horizon, min_news_count)
+
+    prices = prices.sort_values(["ticker", "date"])
+    prices["fwd_close"] = prices.groupby("ticker", group_keys=False)["close"].shift(-horizon)
+    prices["fwd_ret"] = prices["fwd_close"] / prices["close"] - 1.0
+
+    merged = sentiment.merge(
+        prices[["ticker", "date", "fwd_ret"]], on=["ticker", "date"], how="inner"
+    )
+    merged = merged.dropna(subset=["fwd_ret"])
+    if merged.empty:
+        return _empty_result(days, horizon, min_news_count)
+
+    ic_rows: List[Dict[str, Any]] = []
+    for date, grp in merged.groupby("date"):
+        ic = _spearman(grp["score_raw"], grp["fwd_ret"])
+        if ic is not None:
+            ic_rows.append({"date": date, "ic": ic, "n": int(len(grp))})
+
+    if not ic_rows:
+        return _empty_result(days, horizon, min_news_count)
+
+    ic_series = pd.Series([r["ic"] for r in ic_rows], dtype=float)
+    ic_mean = float(ic_series.mean())
+    ic_std = float(ic_series.std(ddof=0)) if len(ic_series) > 1 else 0.0
+
+    if len(ic_rows) < 10:
+        verdict = "skip"
+    elif abs(ic_mean) > 0.05:
+        verdict = "strong"
+    else:
+        verdict = "weak"
+
+    return {
+        "horizon_days": horizon,
+        "min_news_count": min_news_count,
+        "window_days": days,
+        "ic_count": len(ic_rows),
+        "ic_mean": round(ic_mean, 4),
+        "ic_std": round(ic_std, 4),
+        "ic_per_day": ic_rows,
+        "verdict": verdict,
+    }
+
+
+def _empty_result(days: int, horizon: int, min_news_count: int) -> Dict[str, Any]:
+    return {
+        "horizon_days": horizon,
+        "min_news_count": min_news_count,
+        "window_days": days,
+        "ic_count": 0,
+        "ic_mean": None,
+        "ic_std": None,
+        "ic_per_day": [],
+        "verdict": "skip",
+    }

stock-lab/app/screener/router.py

@@ -280,6 +280,7 @@ def list_runs(limit: int = 30):
 
 from .ai_news import pipeline as _ai_pipeline
 from .ai_news import telegram as _ai_telegram
+from .ai_news import validation as _ai_validation
 
 
 @router.post("/snapshot/refresh-news-sentiment")
@@ -312,6 +313,23 @@ async def post_refresh_news_sentiment(asof: Optional[str] = None):
     return summary
 
 
+# ---------- /ai-news/ic ----------
+
+@router.get("/ai-news/ic")
+def get_ai_news_ic(days: int = 30, horizon: int = 1, min_news_count: int = 1):
+    """ai_news.score_raw 의 forward return IC (Spearman) 계산.
+
+    verdict:
+      - skip: ic_count < 10 (데이터 부족)
+      - weak: |ic_mean| <= 0.05
+      - strong: |ic_mean| > 0.05 (gradient 활성화 가치 있음)
+    """
+    with _conn() as c:
+        return _ai_validation.compute_ic(
+            c, days=days, horizon=horizon, min_news_count=min_news_count,
+        )
+
+
 @router.get("/runs/{run_id}")
 def get_run(run_id: int):
     with _conn() as c:

stock-lab/app/screener/schema.py

@@ -12,7 +12,9 @@ DEFAULT_WEIGHTS = {
     "rs_rating":    1.2,
     "ma_alignment": 1.0,
     "vcp_lite":     0.8,
-    "ai_news":      0.8,
+    # ai_news: 검증 전 gradient 차단 (4주 IC > 0.05 확인 후 활성화).
+    # 데이터 수집은 계속, 가중합 영향만 0.
+    "ai_news":      0.0,
 }
 DEFAULT_NODE_PARAMS = {
     "foreign_buy":  {"window_days": 5},
@@ -143,6 +145,11 @@ def ensure_screener_schema(conn: sqlite3.Connection) -> None:
         if "ai_news" not in w:
             w["ai_news"] = DEFAULT_WEIGHTS["ai_news"]
             changed = True
+        # One-time reset: ai_news default 0.8 → 0.0 (검증 전 gradient 차단).
+        # 사용자가 명시적으로 0.8 외 값을 설정했다면 영향 없음.
+        elif w.get("ai_news") == 0.8:
+            w["ai_news"] = 0.0
+            changed = True
         if "ai_news" not in p:
             p["ai_news"] = DEFAULT_NODE_PARAMS["ai_news"]
             changed = True