Self-Learning Calibration — LoL · CS · KBO

Each cron, every POT/POD pick logs its predicted probability. Once the match settles, the actual outcome (WIN/LOSS) is recorded. The self-learning loop then:

• Bucketizes predictions into 10% bins (0-10%, 10-20%, ..., 90-100%)
• Compares predicted hit rate vs actual within each bin (reliability diagram)
• Computes Brier score (sum of (pred - actual)²) — lower = sharper
• Computes ECE (Expected Calibration Error) — average |pred - actual| weighted by bin size
• If the model is systematically over- or under-confident (≥20 settled bets), recommends a shift

Once enough data accumulates (typically 30-50 settled bets per sport), the model will start applying corrections automatically. Until then, calibration metrics are surfaced as transparency.