The purpose of this project was to:
- build model(s) to predict winners of matches based on in-game player statistics
- practice ensemble modeling
- practice with the SHAP library
- practice working with parquet files
The data used for this analysis was taken from this Kaggle page. It includes only mens matches from 1968 through 2022. Match statistics were not available until 1991.
- Since there were no in-game statistics collected (# aces, % of 1st serve in, etc.) before 1991, I decided to filter those matches out from the model. Prior to this decision, I made an attempt at creating a fully observed dataset by imputing pre-1991 values based on post 1991 values. However, any model that I produced had a lousy training/test accuracy or significantly overfit. The best results came from breaking the data out by surface and fitting individual models per surface. Yet, the highest accuracy I was able to achieve was still only approximately 65%. Once I removed the pre-1991 values, the average accuracy rose to approximately 75%.