This project aims to predict NBA player longevity in the league

The goal of this project is predict if a player is likely to stay in the NBA league using his first year statistics (points, rebonds etc...)

• Define the target (how many seasons played can be considered as high longevity in the NBA?)
• Get NBA DATA

The CRoss-Industry Standard Process for Data Mining - (CRISP-DM) will be the method I will be using to process the data.

The datasets were collected online from different sources. The datasets contains each NBA player main statistics for each seasons from the 1996/1997 until the 2016/2017 season.

To build my model and make predictions I followed the following steps:

• Cleaning the data
• Build the target
• Splitting data into train and test sets
• Build a baseline model using Logistic Regression with a 10 folds Cross Validation (using KFold from scikit learn in order to compare the same folds across different models)
• Creating interactions features
• Creating polynomials features
• Building pipelines for the following classifiers
• K-Nearest Neighbors
• Random Forests
• XGBoost
• Then digged deeper using GridSearchCV on our different models
• Chose a threshold independent score (ROC AUC) to chose the best model
• Plot the Confusion Matrix for our corresponding threshold

Gridsearch XGBoost Roc Curve

Gridsearch XGBoost Metrics

Gridsearch XGBoost Tree

By default the classifier will have a recall of 95% when predicting short longevity when it actually is a short longevity player while having around 15% of players misclassified as short longevity. We can always adjust the threshold later depending on our needs.

I can now deploy this model in the following situations:

• When the NBA pre-season starts.
• When the NBA season starts.
• When teams have a lot of injuries and give playing time to players on the end of their bench.

Having this model while extrapolating season stats based on stats/per minutes played can be very useful especially in pre-season where the coaches could use a similar model to chose who they are going to give playing time to.

You can find our slides presentation here => Slides presentation Or in the pdf presentation.pdf in this repository.