Tennis In-Game Win Probability with R

An experiment in building an in-game win probability model for tennis matches. Uses XGBoost.

Sample match

• Point by point

• Point by point
• Video highlights

Accuracy

Women's model

Men's model

Feature Importance

Women's model

Men's model

Reference

Data is from the Match Charting Project.

Development notes

Run on the command line:

$ R --no-save < tennis-win-probability.R

Data cleanup tasks to do:

• Records are numbered by point Pts (approximately 100 per match)
• Set1 and Set2 are sets won by player 1 or two
• Same for Gm1 and Gm2
• The model uses points, games, and sets
• The identity of the player serving the ball is not currently included in the model
• Add estimated points (EPA) for potentially even greater accuracy