The problem
A useful predictor needs to do more than output a winner. It must translate historical performance into calibrated probabilities, explain uncertainty and carry those probabilities through the tournament structure.
The approach
The pipeline combines team strength, form, head-to-head context, confederation and venue-related features. A multiclass model estimates win, draw and loss probabilities, then repeated simulations convert match-level probabilities into group and knockout outcomes.
The model and interface were designed together. The dashboard exposes probabilities and scenarios instead of hiding uncertainty behind one definitive prediction.
The results
On a 192-match retrospective benchmark, the blended model did not beat a plain Elo baseline on accuracy, and the difference was not statistically significant. The blend's pooled log-loss and Brier scores were numerically lower, but the margins were small, inconsistent across tournaments, and not statistically conclusive.
What I owned
- Data ingestion, validation and feature engineering.
- Model training, evaluation and probability outputs.
- Tournament simulation logic and live result tracking.
- Streamlit interface, deployment and technical documentation.
What comes next
Possible improvements include calibration monitoring, automated data checks, clearer scenario explanations and further model evaluation.