Methodology, measured

Model Calibration Lab

Catatan pick memberi tahu apa yang menang. Kalibrasi memberi tahu apakah probabilitas itu berarti: saat kami bilang 60%, apakah terjadi 60% dari waktu? Di bawah ini, setiap prediksi yang settled, dinilai dua cara — sharp market (devigged) dan model kami sendiri — dengan perbandingan yang tidak nyaman terlihat jelas.

Prediksi settled 4,538
Brier score 0.598

Model kami (Elo + form): 0.617

Log-loss 1.001

Model kami (Elo + form): 1.028

Calibration error 0.8pp

Model kami (Elo + form): 0.4pp

Probabilitas prediksi vs frekuensi observed

Setiap titik adalah band probabilitas 10 poin (min. 30 observasi). Semakin dekat ke diagonal, semakin baik kalibrasinya. Setiap match menyumbang tiga hasilnya.

Sharp market (devigged) Model kami (Elo + form) kalibrasi sempurna
0% 0% 25% 25% 50% 50% 75% 75% 100% 100% Model kami (Elo + form) · 0–10%: 7.5% → 7.1% (n=112) Model kami (Elo + form) · 10–20%: 16.4% → 16.8% (n=1056) Model kami (Elo + form) · 20–30%: 26% → 25.9% (n=6205) Model kami (Elo + form) · 30–40%: 34.7% → 34.6% (n=2646) Model kami (Elo + form) · 40–50%: 44.8% → 43.9% (n=2047) Model kami (Elo + form) · 50–60%: 54.3% → 56% (n=1024) Model kami (Elo + form) · 60–70%: 63.9% → 64.3% (n=381) Model kami (Elo + form) · 70–80%: 74.1% → 75.2% (n=105) Model kami (Elo + form) · 80–90%: 83% → 88.2% (n=34) Sharp market (devigged) · 0–10%: 7.4% → 5.6% (n=270) Sharp market (devigged) · 10–20%: 16.2% → 17.5% (n=1657) Sharp market (devigged) · 20–30%: 25.3% → 25.1% (n=5261) Sharp market (devigged) · 30–40%: 34.3% → 33.6% (n=2748) Sharp market (devigged) · 40–50%: 44.8% → 43.9% (n=1680) Sharp market (devigged) · 50–60%: 54.6% → 56.2% (n=1077) Sharp market (devigged) · 60–70%: 64.4% → 63.9% (n=598) Sharp market (devigged) · 70–80%: 74.6% → 79.2% (n=240) Sharp market (devigged) · 80–90%: 83.7% → 86.1% (n=72)
Band Sharp market (devigged) — Predicted (avg) Observed Obs. Model kami (Elo + form) — Predicted (avg) Observed Obs.
0–10% 7.4% 5.6% 270 7.5% 7.1% 112
10–20% 16.2% 17.5% 1657 16.4% 16.8% 1056
20–30% 25.3% 25.1% 5261 26% 25.9% 6205
30–40% 34.3% 33.6% 2748 34.7% 34.6% 2646
40–50% 44.8% 43.9% 1680 44.8% 43.9% 2047
50–60% 54.6% 56.2% 1077 54.3% 56% 1024
60–70% 64.4% 63.9% 598 63.9% 64.3% 381
70–80% 74.6% 79.2% 240 74.1% 75.2% 105
80–90% 83.7% 86.1% 72 83% 88.2% 34

Brier dan log-loss lebih rendah lebih baik; calibration error adalah rata-rata gap berbobot observasi antara prediksi dan observed. Inti yang jujur: devigged sharp market lebih terkalibrasi daripada model kami — dan itulah alasan situs ini mem-price semuanya terhadap market, bukan menjual prediksi model. Metodologi: probabilitas sesuai tampilan saat logging, tanpa retro-fitting; hasil dari hasil resmi.

Metodologi lengkap →
Berguna? Kirim ke teman yang suka taruhan.