Skip to content

Latest commit

 

History

History
132 lines (100 loc) · 12.7 KB

README_AllStars.md

File metadata and controls

132 lines (100 loc) · 12.7 KB
Raw

Predicting the NBA All-Stars with Machine Learning

The goal of this analysis is to predict the NBA All-Stars for a specific year by applying machine learning algorithms on player performance data and All-Star selection data from other years. The analysis is based on the NumPy and Pandas data analysis packages and the Scikit-learn machine learning package for Python. The XGBoost algorithm and the Keras-TensorFlow deep learning libraries are tested as well. Matplotlib, Seaborn and Bokeh are used for visualization. NBA data are taken from basketball-reference.com. Data from 2000-2020 have been saved as csv-files in the data directory using the scraper functions in NBAanalysissetup.py, data from other years can be obtained by using Basketball_Reference_scraper.py.

Analysis

The analysis is presented as a Python Jupyter Notebook, and can be viewed online using Jupyter nbviewer (which has improved display rendering capabilities compared to Github). The outline of the analysis is summarized in the following:

1. Import prerequisites

2. User input

  • Choose the first and last year for which data has been scraped, and choose the year you want to predict. The years that are not selected are used for cross-validation and training of the ML algorithms.
  • Choose the minimum number of games a player has to have started per season to be included in the analysis.

3. Data processing

  • Data loading and preparation (feature selection, NaN handling, etc.):
    • Features included in this analysis: TW/TOT, G/TOT, GS/G, MP/G, 2P/48, 2P%, 3P/48, 3P%, FT/48, FT%, USG%, ORB%, DRB%, AST%, TOV%, STL%, BLK%, PF/48. (Definitions can be found here).
    • Target statistic is the players' All-Star selection status (AS), a binary variable.
  • Feature scaling as required by various ML algorithms.
  • Visualization of distributions of all features for All-Stars and non-All-Stars.

4. Unsupervised Learning: PCA & Clustering

5. Supervised Learning: Classification

6. Cross-validation

  • All classifiers are cross-validated by using training data and the LeaveOneGroupOut cross-validation scheme, where a group is defined as a single NBA season.
  • Validation curves are calculated and visualized.
  • Classification scores for all models are calculated and listed.
  • ROC and PR curves for all models are calculated and visualized.
  • Feature importances for Decision Tree ensemble models are calculated and listed for all CV groups.
  • Feature coefficients for linear models are calculated and listed for all CV groups.

7. Model training and predictions

  • Models are fitted using training data, fitted models are used to predict test data.
  • Confusion Matrices and classification scores for all models are calculated and visualized (if NBA All-Stars have been selected for test year).
  • Feature importances for Decision Tree ensemble models are calculated and listed.
  • Feature coefficients for linear models are calculated and listed.
  • For the Logistic Regression Classifier, the fitted logistic curves corresponding to all data features are visualized.
  • Decision function values / classification probability scores in 2-D feature space for all models are visualized.
  • NBA player predictions for all models are listed.

8. Ensemble model

  • An ensemble model of all selected classifiers with majority voting is created using the VotingClassifier class.
  • Ensemble model is cross-validated and classification scores are calculated and listed.
  • Ensemble model is fitted using training data, fitted ensemble is used to predict test data.
  • NBA player predictions and classification results per classifier in the ensemble model are listed.

9. Probability scores and final prediction

  • Classification probability scores and corresponding scoring ranks of test data are calculated for all models.
  • Predicted score distributions and probability calibration curves of training data are visualized for all models.
  • Calibrated classification probability scores and corresponding scoring ranks of test data are calculated for all models.
  • Predicted calibrated score distributions and calibrated probability calibration curves of test data are visualized for all models.
  • The NBA All-Stars per conference are listed in order of the median predicted calibrated score over all selected models.

NBA All-Star prediction 2020

At the time of writing (October 20, 2020) the NBA All-Stars for 2020 were already selected (NBA All-Star Game 2020). Therefore the analysis identifies three groups of players per conference:

  1. Deserved All-Stars: Players that were selected and are predicted as All-Stars.
  2. Questionable All-Stars: Players that were selected but are not predicted as All-Stars.
  3. Snubbed non-All-Stars: Players that are predicted but were not selected as All-Stars.

The NBA players in these groups for 2020 are listed below, based on 2010-2020 data and in order of the median predicted calibrated scoring rank over all selected models:

  • Western Conference:

    • Deserved All-Stars: Kawhi Leonard (LAC), Anthony Davis (LAL), James Harden (HOU), LeBron James (LAL), Luka Dončić (DAL), Russell Westbrook (HOU), Damian Lillard (POR), Devin Booker (PHO), Donovan Mitchell (UTA), Brandon Ingram (NOP)
    • Questionable All-Stars: Nikola Jokić (DEN), Chris Paul (OKC), Rudy Gobert (UTA)
    • Snubbed non-All-Stars: Paul George (LAC), Karl-Anthony Towns (MIN)

  • Eastern Conference:

    • Deserved All-Stars: Giannis Antetokounmpo (MIL), Joel Embiid (PHI), Jayson Tatum (BOS), Trae Young (ATL), Pascal Siakam (TOR), Jimmy Butler (MIA), Kyle Lowry (TOR), Kemba Walker (BOS), Bam Adebayo (MIA), Khris Middleton (MIL)
    • Questionable All-Stars: Ben Simmons (PHI), Domantas Sabonis (IND)
    • Snubbed non-All-Stars: Bradley Beal (WAS), Zach LaVine (CHI)

Discussion

There are several caveats to this analysis, for instance:

  • NBA player data used in the analysis are summary statistics over a complete season, while the NBA All-Star game is played after roughly 2/3 of the season. Therefore the analysis might actually be more suited to predict the selection of the All-NBA teams awarded at the end of the season.
  • All-Star level players can get injured early in the season and therefore are not selected for the All-Star game, but recover before the season ends and still pass the minimum number of games requirement to be included in the analysis. These will most likely end up as false positives in this analysis.
  • Similarly, players selected for the All-Star game who get injured just before the All-Star game are replaced by other players who otherwise would not have been selected. (In 2019 for instance, D'Angelo Russell is an injury replacement for Victor Oladipo) In this analysis a player's All-Star status is determined by who played in the actual All-Star game, therefore these injured players will most likely be classified as false positives. To limit these false positives, injured players who were unable to play in the All-Star games have been added to the scraped 2000-2020 csv-files by hand.
  • All-Star level players can transfer between conferences during a season, and therefore affect the All-Star predictions per conference. (In 2018 for instance, Blake Griffin played most of the season for the Los Angeles Clippers in the Western Conference, but transferred to the Detroit Pistons in the Eastern Conference just before the All-Star Game). No attempt has been made to correct for this effect.
  • All-Star selection is not only determined by a player's individual performance, but also by his team's performance before the All-Star break. Team performance is included in the analysis by the TW statistic (i.e. the number of team wins per season), but no attempt has been made to tune the weight of this statistic compared to other data features. Therefore it is possible these players will show up as false negatives in this analysis (In 2019 for instance, Klay Thompson played for the Golden State Warriors, the defending NBA champions).
  • Similarly, All-Star selection is (partly) based on fan voting, and therefore popular players can get selected even if they played poorly during the season (e.g. Kobe Bryant in 2014, 2015 and 2016). These players will most likely contribute to the false negatives in this analysis.
  • The 1998–99 and 2011-2012 NBA seasons were shortened to 50 and 66 regular season games per team respectively due to a lock-out, while the 2019-2020 NBA season was shortened and altered due to the Coronavirus pandemic. Therefore there was no All-Star game in 1998-99, while the analysis might be suboptimal for the 2011-2012 and 2019-2020 seasons.