CS2 Match Predictor
Enter five FACEIT nicknames per team. Names must be exact.
How accurate is it?
Tested on 1000 real FACEIT matches the model never saw during training — held out from the full 5000-match dataset (out-of-time: trained on older games, scored on newer ones).
- 69% Accuracy
- 0.753 AUC
- 0.032 Calibration (ECE, lower=better)
The probabilities are calibrated, not just ranked — a "70%" really wins about 70% of the time.
Walk-forward backtest
Train on the past, predict the next block, repeat — over 4440 out-of-sample matches, the held-out figures above hold up:
AUC 0.744 ECE 0.033 Accuracy 69%
How this works
- Read recent form. For each of the 10 players we pull their recent CS2 matches from FACEIT and compute four numbers: K/D ratio, ADR (average damage per round), headshot %, and win rate.
- Compare the two teams. We take the gap between the teams' average FACEIT Elo (the main skill signal) plus the difference in each form stat (Team A minus Team B).
- Predict. A machine-learning model, trained on 5000 real historical FACEIT matches, combines the Elo gap and the form stats into a calibrated win probability — the orange bar above.
Why you can trust the number. The model is trained leakage-safe: the features for any historical match use only each player's matches that finished before it, never the result it is trying to predict. It is then tested out-of-time — trained on older matches and scored on newer ones it has never seen, and its probabilities are calibrated (a "70%" really wins about 70% of the time). Accuracy is ~70% across typical matchups but closer to 60% when the two teams are evenly matched.
The blue bar is a model-free baseline: the standard Elo formula applied to the players' current FACEIT Elo. Showing it next to the model lets you see whether the model adds anything over the raw ratings.