How good is CrickenZen?
Transparent model performance so you can judge the engine before you trust the live numbers. Every metric here is sourced from completed matches only.
Proof data is not available yet. This can happen when no completed match data exists for this league, or the snapshot has not been generated. Run the snapshot build script to create it.
Headline proof
Not ready
Probability error. Lower is better.
Not ready
Calibration gap. Lower is better.
Not ready
Pre-match call hit rate.
Not ready
Evaluated ball states from completed matches.
Window: None
How to read this
Brier Score
Brier score measures the average squared error between predicted probabilities and actual outcomes. A score of 0 means perfect predictions; 1 is the worst. Lower is better.
ECE
Expected Calibration Error (ECE) compares predicted confidence against actual accuracy. If the model says 60% and it wins 60% of the time, calibration is perfect (ECE = 0). Lower is better.
Accuracy
Accuracy tracks discrete pre-match prediction calls. It counts how often the model-favored team actually won. This is not the same as probability calibration — a model can have good accuracy on clear favourites but poor calibration on close matches.
Calibration ≠ Accuracy
Brier and ECE measure probability quality (calibration). Accuracy measures discrete call hit rate. They are different metrics and should not be added together or averaged into a single trust score.
All metrics are computed from completed matches only. Live ball-state data is excluded. Proof ledger rows are derived from the model's first-ball probability compared against the actual winner. Sample sizes and dates are shown so you can judge how current this evidence is.
Segment metrics will appear once proof data is available.
Recent proof ledger
Proof ledger rows will appear once completed match data is available.
See current win probability and projections.
Public prediction cards with score state and insight.
Monte Carlo, ODM, full timeline, and alerts.