감성평가 정확성 강화 (A+B+C+D+E)

score-engine/main.py: A) 시간감쇠: exp 3일 반감기 (3d→0.5, 7d→0.2) A) 유사뉴스 중복 보정: sqrt(similar_count) cap 2.5x C) catalyst 패턴 매칭: "영업이익 급증" 등 세분 라벨을 실적/수주/배당/리스크/ M&A/신제품/규제/정책/모멘텀 9개 그룹으로 통합 B) calc_sentiment_momentum: 최근 3일 vs 이전 4일 변화율 (-50~+50) D) calc_market_sentiment_baseline + per-stock sentiment_alpha (-100~+100, 시장 평균 대비 고유 edge) E) calc_news_surge_and_attention: 7d/28d 비율 + log 관심도(0~100) 종합점수 통합: - news_score (0.18)는 A+C로 직접 정확도 향상 - sentiment_momentum * 0.06 + sentiment_alpha * 0.03 → max ±5 보너스 - surge≥3 AND news_score>10 → +2 (강한 호재 attention 가산) stock_scores 스키마: 4컬럼 추가 - sentiment_momentum, sentiment_alpha, attention_score, news_surge_ratio dashboard-api: - /api/formulas/matrix: 5컬럼(news_score + 4신규) 추가 노출 - index.html: 10공식 매트릭스에 뉴스/M/α/관심/× 5컬럼 추가 검증: 1464종목 재산출 완료. attention >0 종목 1294(88%), 모멘텀 변동 259종목. 추천 분포 안정 (강력매수 19 / 매수관심 104 / 관망 1249 / 매도관심 73 / 강력매도 22). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-20 22:54:17 +09:00
parent 324c7a4b95
commit 47e7a5fb66
3 changed files with 223 additions and 26 deletions
@@ -1730,6 +1730,11 @@ async function renderFormulas(){
        <th title="Mohanram G-Score (vs 섹터 중앙값)">G</th>
        <th title="Amihud 비유동성">Amh</th>
        <th title="저베타 알파 (Frazzini-Pedersen BAB)">β</th>
+        <th title="catalyst×감쇠×중복 가중 뉴스점수 (-100~100)">뉴스</th>
+        <th title="감정 모멘텀: 최근 3일 vs 이전 4일 변화율">M</th>
+        <th title="시장 평균 대비 sentiment alpha">α</th>
+        <th title="최근 7일 뉴스건수 log 스케일 관심도 (0~100)">관심</th>
+        <th title="평소 대비 뉴스 폭증 배율 (x)">×</th>
      </tr></thead><tbody>`;
    rows.forEach(r=>{
      let sig = r.signals || {};
@@ -1757,6 +1762,36 @@ async function renderFormulas(){
        ${_cell('G', r.g_score!=null?(+r.g_score).toFixed(0):'-', hit('gscore'), 'Mohanram G (vs 섹터)')}
        ${_cell('Amh', r.amihud_illiq!=null?(+r.amihud_illiq).toFixed(0):'-', hit('amihud'), 'Amihud 비유동성')}
        ${_cell('β', r.market_beta!=null?(+r.market_beta).toFixed(2):'-', hit('beta'), '60일 KOSPI 회귀 베타')}
+        ${(()=>{ // 뉴스점수
+          const v = r.news_score;
+          if(v==null) return '<td style="text-align:center;color:#37474F">-</td>';
+          const c = v>=10?'#69F0AE':v<=-10?'#FF8A80':'#90A4AE';
+          return `<td style="text-align:center;font-family:'JetBrains Mono',monospace;color:${c};font-weight:700">${(+v).toFixed(0)}</td>`;
+        })()}
+        ${(()=>{ // 모멘텀
+          const v = r.sentiment_momentum;
+          if(v==null) return '<td style="text-align:center;color:#37474F">-</td>';
+          const c = v>=5?'#69F0AE':v<=-5?'#FF8A80':'#90A4AE';
+          return `<td style="text-align:center;font-family:'JetBrains Mono',monospace;color:${c}">${(+v).toFixed(1)}</td>`;
+        })()}
+        ${(()=>{ // alpha
+          const v = r.sentiment_alpha;
+          if(v==null) return '<td style="text-align:center;color:#37474F">-</td>';
+          const c = v>=5?'#69F0AE':v<=-5?'#FF8A80':'#90A4AE';
+          return `<td style="text-align:center;font-family:'JetBrains Mono',monospace;color:${c}">${v>0?'+':''}${(+v).toFixed(0)}</td>`;
+        })()}
+        ${(()=>{ // attention
+          const v = r.attention_score;
+          if(v==null) return '<td style="text-align:center;color:#37474F">-</td>';
+          const c = v>=70?'#FFD740':v>=40?'#69F0AE':'#90A4AE';
+          return `<td style="text-align:center;font-family:'JetBrains Mono',monospace;color:${c}">${(+v).toFixed(0)}</td>`;
+        })()}
+        ${(()=>{ // surge
+          const v = r.news_surge_ratio;
+          if(v==null) return '<td style="text-align:center;color:#37474F">-</td>';
+          const c = v>=3?'#FFD740':v>=1.5?'#69F0AE':'#546E7A';
+          return `<td style="text-align:center;font-family:'JetBrains Mono',monospace;color:${c}">${(+v).toFixed(1)}×</td>`;
+        })()}
      </tr>`;
    });
    h += `</tbody></table></div>`;
@@ -2017,7 +2017,9 @@ async def formulas_matrix(limit: int = Query(default=50, ge=5, le=200)):
                   s.total_score, s.recommendation, s.buy_votes, s.sell_votes,
                   s.magic_score, s.f_score, s.altman_z, s.peg, s.momentum_pct,
                   s.beneish_score, s.gpa_pct, s.g_score, s.amihud_illiq, s.market_beta,
-                   s.signals, s.sector
+                   s.signals, s.sector,
+                   s.news_score, s.sentiment_momentum, s.sentiment_alpha,
+                   s.attention_score, s.news_surge_ratio
            FROM stock_scores s
            LEFT JOIN dart_corps d ON d.stock_code = s.stock_code
            WHERE s.score_date = (SELECT MAX(score_date) FROM stock_scores)
@@ -9,7 +9,7 @@
 - 매일 장 마감 후 자동 집계 + 텔레그램 알림
 """
 import asyncio, json, os
-from datetime import datetime, date, timedelta
+from datetime import datetime, date, timedelta, timezone
 from typing import Optional
 import asyncpg, httpx, redis.asyncio as aioredis, structlog
 from apscheduler.schedulers.asyncio import AsyncIOScheduler
@@ -994,33 +994,166 @@ async def calc_position_size(conn, stock_code: str, total_score: float) -> tuple
    return round(max(1.0, min(15.0, size)), 2), round(vol, 2)


-# ── M4: catalyst 가중치 ───────────────────────────────────
+# ── 감성 평가 강화 모듈 (M4 + A/B/C/D/E) ──────────────────
+import math as _math
+
+# catalyst 가중치 (정규화된 라벨 기준)
 CATALYST_WEIGHTS = {
-    "실적": 1.5, "수주": 1.3, "배당": 1.2, "리스크": 1.4, "기타": 1.0, "모멘텀": 0.8,
+    "실적": 1.5, "수주": 1.3, "배당": 1.2, "리스크": 1.4,
+    "M&A": 1.3, "신제품": 1.2, "규제": 1.3, "정책": 1.2,
+    "기타": 1.0, "모멘텀": 0.8,
 }

-async def calc_news_score_weighted(conn, stock_code: str, week_ago: date) -> tuple[float, dict]:
-    """catalyst별 가중치 적용된 뉴스 점수"""
+# 세분 catalyst → 정규화 그룹 (EXAONE이 자유 형식으로 뱉어내는 라벨을 통합)
+_CATALYST_PATTERNS = [
+    ("실적",   ["실적", "영업이익", "매출", "순이익", "어닝", "분기실적", "흑자", "적자", "감익", "증익"]),
+    ("수주",   ["수주", "계약", "공급", "납품", "선정"]),
+    ("배당",   ["배당", "환원", "자사주"]),
+    ("리스크", ["리스크", "악재", "소송", "징계", "리콜", "조사", "조작", "회계", "감리", "제재"]),
+    ("M&A",    ["인수", "합병", "분할", "지분", "스왑"]),
+    ("신제품", ["신제품", "출시", "런칭", "공개", "발표"]),
+    ("규제",   ["규제", "법안", "허가", "인증", "승인", "심사"]),
+    ("정책",   ["정책", "정부", "지원금", "보조금", "세제", "예산"]),
+    ("모멘텀", ["모멘텀", "기대감", "전망", "관심"]),
+]
+def _map_catalyst(raw: str | None) -> str:
+    if not raw: return "기타"
+    s = str(raw).strip()
+    if s in CATALYST_WEIGHTS: return s
+    for group, keys in _CATALYST_PATTERNS:
+        if any(k in s for k in keys):
+            return group
+    return "기타"
+
+def _time_weight(age_days: float, halflife_days: float = 3.0) -> float:
+    """exp 시간감쇠. 3일 반감기 (0d→1.0, 3d→0.5, 7d→0.20)."""
+    return _math.exp(-age_days / max(halflife_days, 0.5) * _math.log(2))
+
+def _similar_weight(similar_count: int | None) -> float:
+    """동일 사건 다수 매체 보도 보정 — sqrt 스케일링, cap 2.5x."""
+    n = max(1, int(similar_count or 1))
+    return min(2.5, _math.sqrt(n))
+
+
+async def calc_news_score_weighted(
+    conn, stock_code: str, week_ago: date, now: datetime | None = None
+) -> tuple[float, dict]:
+    """
+    catalyst×intensity×시간감쇠×중복가중 적용된 뉴스 점수 (-100~+100).
+    동시에 attention/neutral 카운트도 함께 반환 (호재·악재만 점수에 반영).
+    """
+    now = now or datetime.now(timezone.utc)
+    week_ago_dt = datetime.combine(week_ago, datetime.min.time(), tzinfo=timezone.utc)
    rows = await conn.fetch("""
-        SELECT sentiment, intensity, COALESCE(catalyst, '기타') AS catalyst
+        SELECT sentiment, intensity, COALESCE(catalyst, '기타') AS catalyst,
+               analyzed_at, COALESCE(similar_count, 1) AS sim
+        FROM news_analysis
+        WHERE primary_stock=$1 AND analyzed_at >= $2
+    """, stock_code, week_ago_dt)
+    if not rows:
+        return 0.0, {"pos": 0, "neg": 0, "neutral": 0, "total": 0}
+    score = 0.0
+    pos = neg = neutral = 0
+    for r in rows:
+        sent = r["sentiment"]
+        if sent == "중립":
+            neutral += 1
+            continue
+        if sent not in ("호재", "악재"):
+            continue
+        cat = _map_catalyst(r["catalyst"])
+        cw  = CATALYST_WEIGHTS.get(cat, 1.0)
+        intensity = float(r["intensity"] or 1)
+        # 시간감쇠 (3일 반감기)
+        try:
+            age_days = (now - r["analyzed_at"]).total_seconds() / 86400.0
+        except Exception:
+            age_days = 3.0
+        tw = _time_weight(max(0.0, age_days))
+        sw = _similar_weight(r["sim"])
+        delta = intensity * 5.0 * cw * tw * sw
+        if sent == "호재":
+            score += delta; pos += 1
+        else:
+            score -= delta; neg += 1
+    return max(-100.0, min(100.0, score)), {
+        "pos": pos, "neg": neg, "neutral": neutral, "total": len(rows)
+    }
+
+
+async def calc_sentiment_momentum(
+    conn, stock_code: str, now: datetime | None = None
+) -> float:
+    """
+    최근 3일 가중 sentiment 합 - 그 이전 4일 가중 sentiment 합 → 모멘텀 (-50~+50).
+    """
+    now = now or datetime.now(timezone.utc)
+    cutoff_recent = now - timedelta(days=3)
+    cutoff_old    = now - timedelta(days=7)
+    rows = await conn.fetch("""
+        SELECT sentiment, intensity, analyzed_at, COALESCE(similar_count, 1) AS sim,
+               COALESCE(catalyst, '기타') AS catalyst
        FROM news_analysis
        WHERE primary_stock=$1 AND analyzed_at >= $2
          AND sentiment IN ('호재','악재')
-    """, stock_code, datetime.combine(week_ago, datetime.min.time()))
-    if not rows:
-        return 0.0, {"pos": 0, "neg": 0, "total": 0}
-    score = 0.0
-    pos = neg = 0
+    """, stock_code, cutoff_old)
+    recent = old = 0.0
    for r in rows:
-        w = CATALYST_WEIGHTS.get(r["catalyst"], 1.0)
-        intensity = float(r["intensity"] or 1)
-        if r["sentiment"] == "호재":
-            score += intensity * 5 * w
-            pos += 1
+        cw = CATALYST_WEIGHTS.get(_map_catalyst(r["catalyst"]), 1.0)
+        sw = _similar_weight(r["sim"])
+        val = float(r["intensity"] or 1) * cw * sw
+        if r["sentiment"] == "악재": val = -val
+        if r["analyzed_at"] >= cutoff_recent:
+            recent += val
        else:
-            score -= intensity * 5 * w
-            neg += 1
-    return max(-100.0, min(100.0, score)), {"pos": pos, "neg": neg, "total": len(rows)}
+            old += val
+    # 일평균으로 정규화 (3일 vs 4일) 후 차이 → 약간 증폭
+    momentum = (recent / 3.0) - (old / 4.0)
+    return max(-50.0, min(50.0, momentum * 2.0))
+
+
+async def calc_news_surge_and_attention(
+    conn, stock_code: str, now: datetime | None = None
+) -> tuple[float, float]:
+    """
+    (surge_ratio, attention_score) 반환.
+    surge_ratio = 최근 7일 일평균 뉴스 / 이전 28일 일평균 뉴스 (>1 = 평소보다 폭증).
+    attention_score = 최근 7일 전체 뉴스 건수 (중립 포함) — 50건 이상이면 100 cap, log 스케일.
+    """
+    now = now or datetime.now(timezone.utc)
+    row = await conn.fetchrow("""
+        SELECT
+          COUNT(*) FILTER (WHERE analyzed_at >= $2)                    AS recent7,
+          COUNT(*) FILTER (WHERE analyzed_at >= $3 AND analyzed_at < $2) AS prev28
+        FROM news_analysis WHERE primary_stock=$1
+    """, stock_code, now - timedelta(days=7), now - timedelta(days=35))
+    recent7 = int(row["recent7"] or 0)
+    prev28  = int(row["prev28"]  or 0)
+    rate_recent = recent7 / 7.0
+    rate_prev   = max(prev28 / 28.0, 0.05)
+    surge = rate_recent / rate_prev
+    surge = max(0.0, min(10.0, surge))
+    attention = min(100.0, _math.log1p(recent7) * 25.0)  # 0건→0, 7건→55, 30건→85, 50건→100
+    return float(surge), float(attention)
+
+
+async def calc_market_sentiment_baseline(conn, week_ago: date) -> float:
+    """전체 시장 종목당 평균 가중 sentiment (sentiment_alpha 산출용 baseline)."""
+    week_ago_dt = datetime.combine(week_ago, datetime.min.time(), tzinfo=timezone.utc)
+    row = await conn.fetchrow("""
+        SELECT AVG(per_stock)::float AS mean FROM (
+          SELECT primary_stock,
+                 SUM(CASE
+                   WHEN sentiment='호재' THEN intensity * 5.0
+                   WHEN sentiment='악재' THEN -intensity * 5.0
+                   ELSE 0 END) AS per_stock
+          FROM news_analysis
+          WHERE analyzed_at >= $1 AND primary_stock IS NOT NULL
+            AND sentiment IN ('호재','악재')
+          GROUP BY primary_stock
+        ) t
+    """, week_ago_dt)
+    return float((row and row["mean"]) or 0.0)


 # ── H3: KOSPI 200일 데이터 수집 (네이버 finance) ──────────
@@ -1391,6 +1524,9 @@ async def calculate_daily_scores():
                logger.warning("weights.load_err", error=str(e))

    async with pg_pool.acquire() as conn:
+        # D: 시장 sentiment baseline 1회 계산 (전 종목 sentiment_alpha 산출용)
+        market_sentiment_baseline = await calc_market_sentiment_baseline(conn, week_ago)
+        logger.info("sentiment.market_baseline", value=round(market_sentiment_baseline, 2))
        # H3: 시장 레짐 1회 계산 (전 종목 동일 적용)
        regime_label, regime_adj = await calc_market_regime(conn)
        await conn.execute("""
@@ -1740,9 +1876,17 @@ async def calculate_daily_scores():
            if ensemble_summary:
                fin_reasons.append(f"공식보팅 [{ensemble_summary}]")

-            # M4: catalyst 가중 뉴스점수로 교체 (위에서 계산한 raw_news 대체)
+            # M4 + A/C: catalyst 가중 + 시간감쇠 + similar_count 적용된 뉴스 점수
            news_score_w, news_stats = await calc_news_score_weighted(conn, stock, week_ago)
            news_score = news_score_w
+            # B: 감정 모멘텀 (3일 vs 4일 변화율)
+            sentiment_momentum = await calc_sentiment_momentum(conn, stock)
+            # E: 뉴스 surge + attention(중립 포함 총 관심도)
+            news_surge_ratio, attention_score = await calc_news_surge_and_attention(conn, stock)
+            # D: 시장 평균 대비 sentiment alpha (per-stock raw sum - 시장 평균)
+            sentiment_raw_sum = float(news_stats.get("pos", 0)) * 5.0 - float(news_stats.get("neg", 0)) * 5.0
+            sentiment_alpha = max(-100.0, min(100.0,
+                sentiment_raw_sum - market_sentiment_baseline))

            # 펀더멘털 통합: 기존 + 추세 + 이익품질 + 매직포뮬러 + F-Score (DCF는 종합점수에 별도 가중)
            fundamental_combined = max(-100.0, min(100.0,
@@ -1754,6 +1898,18 @@ async def calculate_daily_scores():
                     + technical_score * 0.15 + dart_score * 0.10
                     + foreign_score * 0.14 + short_score * 0.06
                     + price_score * 0.03 + mos_score * 0.10)
+            # B+D: 감정 모멘텀 + 시장 alpha 보너스 (max ±5)
+            sentiment_bonus = max(-5.0, min(5.0,
+                sentiment_momentum * 0.06 + sentiment_alpha * 0.03))
+            total += sentiment_bonus
+            if abs(sentiment_bonus) >= 1.5:
+                fin_reasons.append(
+                    f"감정 {('+' if sentiment_bonus>0 else '')}{sentiment_bonus:.1f}"
+                    f"(모멘텀 {sentiment_momentum:+.1f} · alpha {sentiment_alpha:+.0f})")
+            # E: news surge ≥3.0 + 뉴스점수 양수 → 강한 attention 신호로 +2 (악재 surge는 차감 안 함)
+            if news_surge_ratio >= 3.0 and news_score > 10:
+                total += 2.0
+                fin_reasons.append(f"뉴스 surge ×{news_surge_ratio:.1f}")
            # 앙상블 보팅 가산점: 학습 가중치 적용 (max ±18, 균등 시 6공식 합 = 18)
            ensemble_bonus = 0.0
            for fname, fsig in sig_map.items():
@@ -1832,10 +1988,11 @@ async def calculate_daily_scores():
                    magic_score, f_score, roc_pct, earnings_yield_pct,
                    altman_z, peg, momentum_pct, beneish_score,
                    signals, buy_votes, sell_votes,
-                    gpa_pct, g_score, amihud_illiq, market_beta)
+                    gpa_pct, g_score, amihud_illiq, market_beta,
+                    sentiment_momentum, sentiment_alpha, attention_score, news_surge_ratio)
                VALUES ($1,$2,$3,$4,$5,$6,$7,$8,$9,$10,$11,$12,$13,$14,$15,$16,$17,$18,$19,
                        $20,$21,$22,$23,$24,$25,$26,$27,$28,$29,$30,$31,$32,$33,$34,
-                        $35,$36,$37,$38,$39,$40,$41,$42,$43,$44,$45)
+                        $35,$36,$37,$38,$39,$40,$41,$42,$43,$44,$45,$46,$47,$48,$49)
                ON CONFLICT (stock_code, score_date) DO UPDATE SET
                    news_score=$9, dart_score=$12, price_score=$14,
                    technical_score=$15, foreign_score=$16, short_score=$17,
@@ -1847,9 +2004,11 @@ async def calculate_daily_scores():
                    magic_score=$31, f_score=$32, roc_pct=$33, earnings_yield_pct=$34,
                    altman_z=$35, peg=$36, momentum_pct=$37, beneish_score=$38,
                    signals=$39, buy_votes=$40, sell_votes=$41,
-                    gpa_pct=$42, g_score=$43, amihud_illiq=$44, market_beta=$45
+                    gpa_pct=$42, g_score=$43, amihud_illiq=$44, market_beta=$45,
+                    sentiment_momentum=$46, sentiment_alpha=$47,
+                    attention_score=$48, news_surge_ratio=$49
            """, stock, name, today,
-                news_stats["pos"], news_stats["neg"], 0, news_stats["total"],
+                news_stats["pos"], news_stats["neg"], news_stats.get("neutral", 0), news_stats["total"],
                avg_int, news_score,
                dart_pos, dart_neg, dart_score, price_change, price_score,
                technical_score, foreign_score, short_score,
@@ -1861,7 +2020,8 @@ async def calculate_daily_scores():
                magic_score, f_score, roc_pct, ey_pct,
                altman_z, peg_val, mom_val, beneish_val,
                json.dumps(sig_map, ensure_ascii=False), vote_counts["매수"], vote_counts["매도"],
-                gpa_val, g_val, amihud_val, beta_val)
+                gpa_val, g_val, amihud_val, beta_val,
+                sentiment_momentum, sentiment_alpha, attention_score, news_surge_ratio)

            # 미국증시 overnight 보정값 별도 컬럼 저장
            if us_info["adj"] or us_info["top"]: