The goal of this project is to predict the final outcome of the playoffs using the official statistics of all the games played by the NBA from January 2004 to March 2021. This report explains the implementation of the simulation program used. Briefly, the program works by taking as input the list of teams in the playoffs and simulates
The complete dataset available is made up of 5 .csv files: (games.csv, games details.csv, players.csv, ranking.csv, teams.csv). Only the data from games.csv were used in this project. The features available for each row are: (game_date, game_id, home_team, visitor_team, field_goal_percentage_home, field_goal_three_point_percentage_home, free_throws_made_home, rebound_home, assists_steals_home, blocks_home, field_goal_percentage_visitor, field_goal_three_point_percent_age visitor, free_throws_made_visitor, rebound_visitor, assists_steals_visitor, blocks_visitor, home_win, visitor_win). The variables represent the statistics of each match played from January 2004 until March 2021 for both the home team and the visitor. The binary variable we want to predict is home win {0,1}, 0 defeat, 1 win.
The proposed solution is to create a program that simulates the playoffs. This program consists of three main parts: a Machine Learning model capable of predicting the winner of a match, a stochastic features data generator on which to use the classifier, and finally the simulation program, which returns the final scoreboard. The figure below shows the flow diagram of the complete simulation model and the implementation for each part is explained in the sections below.
The goal is to create a function, which, taking two teams as input, returns the winner. To make the winner prediction, it was decided to test three different Machine Learning models, to compare their effectiveness and choose the best one. The Scikit-Learn library in Python was used for the implementation of the models. The games.csv dataset was divided by the 70% training set and 30% test set. To train the classifier it has been considered only the match statistics features, 10 in total. The output variable we want to predict is home win. Scikit-Learn allows us to use the models without hyperparameterization, so it was decided to make a comparison of the models with and without optimizations. The models have been optimized using the GridSearchCV function, which allows us to loop various fittings using combinations of hyperparameters chosen by us. Accuracy and AUC were chosen as metrics to evaluate the models.
To simulate the result of the playoffs it is also necessary to generate possible data on which to predict the future result of the games. To achieve this goal, a game sample function has been created, which, taken as input two teams, generates possible features of a match. First of all, to create these features, the original dataset was taken considering the most recent data, therefore to predict the 2022 season it was considered from 2020 onwards. For each team it was created a separate datasets with the match statistics results they obtained in each game of games.csv, therefore 5 features were taken as we do not distinguish the scores between home or visitor. Using the Fitter library available for Python it was possible to obtain the statistical distribution for each of the 5 features for each team. The game sample function in the simulator therefore generates a random sample of features from the distribution for each of the two input teams and by putting the features together it is possible to generate a possible result of a match between the two teams.
The simulator takes as input the list of teams admitted to the playoffs and is composed of three functions: (play ngames, play round, simulate). The play - ngames function takes two teams as input and generates a sample of features using game sample for each of the 7 games, the result for each game is predicted using the classifier. In the play_round function, play_ngames is repeated for each round, eliminating the eliminated teams from the list. Finally, to get a statistic and return a probability in the simulate function, N iterations of play round are made and for each round of the playoffs it is kept track of the times in which a team has managed to pass to the next round. Therefore the final result of the simulation is the probability that a team has to move on to the next round.
The figure below shows the output achieved after 5000 simulations of the play- offs of the current season 2021-2022, using Random Forest as a classifier. It is possible to see that with the exception of the Suns-Mavericks, Bucks-Celtics and Heat-Celtics matches, the rest was correctly predicted. The reason for the mistake can be admitted to the fact that Suns and Bucks played below expec-tations, Heat and Celtics was a very tight match and in general Celtics played beyond expectations, however the winner of the finals was correctly predicted. These predictions could be further improved by adding player statistics, injuries, or match strategies used in previous matches.
