Model Accuracy Tracker

How CSDB tracks prediction model accuracy

A scheduled job prices every upcoming match with the same model that runs on the predictions pageand stores the probability, the confidence and the model version at that moment. When the match finishes, only the outcome columns are filled in — the probability itself is never rewritten, and shipping a new model version cannot improve the old one's record. Each match is scored once, using the last snapshot taken before kickoff.

What the metrics mean

• Overall accuracy— percentage of matches where the model's favourite won. A 60%+ accuracy on CS2 BO3s is meaningful, given how high-variance the game is.
• Current streak — consecutive correct (positive) or incorrect (negative) predictions, useful for spotting hot or cold periods.
• Calibration— when the model says "70% confidence," does the favourite actually win ~70% of the time? Calibration measures whether the model's confidence scores are honest.
• Brier score — the squared error between the stated probability and what happened, averaged over every scored match. Lower is better; 0.25 is the score you get by calling everything 50/50, so anything meaningfully below that is the model earning its keep. It only appears here because the probabilities were stored in advance.

What we do and don't include

• ✅ Every BO1, BO3, and BO5 played by HLTV-eligible teams
• ✅ Both upset wins and chalk results
• ❌ Matches where the model didn't have enough data on either team (typically tier-3 newcomers)
• ❌ Forfeits and walkovers (no actual play)

Why transparency matters

Any model can look good on hand-picked matches. By publishing the full track record, CSDB lets you decide whether the model's edge is real. If you're using the hot picks page or the predictions tool for actual decisions, this page is where you check whether to trust them.