How the model actually did
The headline number is the public record, every league match the model called, win or lose. Average confidence and log-loss measure something different: not whether the model was right, but how well its confidence matched reality.
Does 60% confidence mean a 60% win rate?
A well-calibrated model's confidence numbers should match its actual hit rate inside each bracket. If the bars below track close to their bracket range, the model isn't just right on average, it knows how sure it should be.
Calibration is computed across the full 70-match season.
Every match, in order
Match, pick, confidence, result. No deletions, no edits to picks after the fact.
| # | Date | Match | Model says | Confidence | Result |
|---|
What goes into a pick
Four inputs feed the win probability before every toss. Nothing here is exotic. The point is that it's consistent, timestamped, and applied the same way to every match.
- Last-minute team news and injuries announced after lineups lock
- Pitch reports released morning-of
- Weather and dew factor on the night
- A single player having the game of their life
- Momentum from a dressing-room incident or controversy
- Auction-fresh players with no IPL 2026 sample size yet
BBL is up next
Same model, same rules, same honesty, rebuilt for the Big Bash League using historical Cricsheet data and backtested before a single live pick goes out. Get notified when picks start.