patrickjohnston / nba-rankings-2024

NBA-RANKINGS-2024

The model is a game-level classier that predicts the probability of either team winning the match-up based on historical box score stats. No assumptions are used to make predictions; the only judgment applied is the choices made for feature engineering and modeling framework.

What Does the Model Predict?

The model predicts the probability of either team in the match-up winning. It does not predict points scored.

How the Model Works

The model is based on a two-step procedure:

• Create PCAs using on box score statistics and player height. The PCAs capture player characteristics and are based on normalized features, such as three-point tendency, share of rebounds, true shooting percentage, etc.

• The PCAs are based on data leading up to the game. No data from the game itself is used, except for injury status.

• Train a shallow xgBoost model using the PCAs and a home court indicator. The model is trained on data since the 2014-2015 season. One-game-ahead in-sample AUC IS 0.73 and the one-game-ahead out-of-sample AUC is 0.71.

• Team performance is not included in the models. Predictions are purely based on roster composition and home-court advantage.

Benefits of this approach:

• Data engineering is simple since it relies on box score stats.
• This is a player level model — which means that it can adapt to changes to rosters, injuries, and line-ups. No need to wait and see how the team does post roster changes.
• We can also use it to simulate how roster changes can affect the outcome of a given match-up.

Downsides to this approach:

• It relies purely on box score stats. No advanced stats based on play-by-play data are included.
• The minute allocation algorithm is based on simple moving averages. In real life, coaches adapt to the opponent.

About the code

1 - read_and_transform_data: The step that reads the raw data from the hoopR package, creates the features, and organizes the data for training.

2 - train_model.R: Trains and validates the xgBoost model.

3 - plot.R Creates the scatter plot to analyze predicted versus actual team rankings.