Shadow cohort 적중률
v1.8 (production) vs v2.1-B-shadow 가중치 일별 Brier + 적중률 delta. 동일 경기 동일 input 으로 quant 재계산 (debate LLM 호출 X, 비용 0). n=150 도달 후 production 적용 결정.
내부 evidence cohort 입니다. 실제 예측은 예측 기록 에서 확인하세요.
누적 표본 (n)
52
v1.8 평균 Brier
0.2432
v2.1-B-shadow 평균 Brier
0.2406
delta -0.0026
SHADOW_WEIGHTS (v2.1-B + shadow factor 활성)
sp_fip
16%
sp_xfip
5%
lineup_woba
17%
bullpen_fip
11%
recent_form
12%
war
9%
head_to_head
2%
park_factor
4%
elo
9%
sfr
0%
park_weather
3%
umpire_sz
2%
shadow factor: park_weather 3% · umpire_sz 2% — production weight 0 양립.
Calibration plot (10-bucket)
예측 확률 vs 실제 적중률. 대각선 위 = 잘 맞은 bucket, 대각선 아래 = 과신 (over-confident). v1.8 (production) vs v2.1-B-shadow (shadow) 두 모델 calibration 차이 시각화.
표본 수집 중 — n=52 / 100 (n=100 도달 시 신뢰도 ↑)
일별 cohort (최대 14일)
| 날짜 | n | v1.8 Brier | v2.1-B-shadow Brier | delta | v1.8 적중 | v2.1-B-shadow 적중 |
|---|---|---|---|---|---|---|
| 2026-06-03 | 5 | 0.2879 | 0.2748 | -0.0132 | 0/5 (0.0%) | 2/5 (40.0%) |
| 2026-06-02 | 5 | 0.2065 | 0.2163 | +0.0098 | 5/5 (100.0%) | 4/5 (80.0%) |
| 2026-05-31 | 5 | 0.2202 | 0.2203 | +0.0001 | 3/5 (60.0%) | 4/5 (80.0%) |
| 2026-05-30 | 5 | 0.2180 | 0.2300 | +0.0120 | 4/5 (80.0%) | 4/5 (80.0%) |
| 2026-05-29 | 5 | 0.2406 | 0.2381 | -0.0024 | 2/5 (40.0%) | 2/5 (40.0%) |
| 2026-05-28 | 2 | 0.2562 | 0.2322 | -0.0240 | 1/2 (50.0%) | 1/2 (50.0%) |
| 2026-05-27 | 4 | 0.2492 | 0.2473 | -0.0018 | 1/4 (25.0%) | 2/4 (50.0%) |
| 2026-05-26 | 1 | 0.2809 | 0.2450 | -0.0359 | 0/1 (0.0%) | 1/1 (100.0%) |
| 2026-05-24 | 3 | 0.2567 | 0.2318 | -0.0249 | 2/3 (66.7%) | 2/3 (66.7%) |
| 2026-05-23 | 1 | 0.2601 | 0.2362 | -0.0239 | 0/1 (0.0%) | 1/1 (100.0%) |
| 2026-05-22 | 1 | 0.2500 | 0.2144 | -0.0356 | 1/1 (100.0%) | 1/1 (100.0%) |
| 2026-05-21 | 2 | 0.2182 | 0.2454 | +0.0272 | 1/2 (50.0%) | 1/2 (50.0%) |
| 2026-05-17 | 1 | 0.2500 | 0.2798 | +0.0298 | 0/1 (0.0%) | 0/1 (0.0%) |
| 2026-05-16 | 2 | 0.2168 | 0.2400 | +0.0231 | 2/2 (100.0%) | 1/2 (50.0%) |
본 페이지는 shadow cohort 누적 진단용입니다. v1.8 production 가중치 변경 X.