NBA Game Simulator

Team Chemistry Predictor

How do we test this?

Let's build a simulator!

Who would benefit?

• The NBA of course!
• Sports analysts
• Sports gamblers

Player Tracking

• Gives $(x, y)$ coordinates of players
• Gives $(x, y, z)$ coordinates of the ball
  ...at 25 frames per second!

Data Snippet

			game_id     game_date   period  game_clock  game_time_remain shot_clock  \
		  0021500568  2016-01-11     1       714.04       2874.04       18.27   
		  0021500568  2016-01-11     1       714.04       2874.04       18.27   
		  0021500568  2016-01-11     1       714.04       2874.04       18.27   
		  0021500568  2016-01-11     1       714.04       2874.04       18.27   

		  team_id      player_id    x_loc    y_loc     elevation     dist_to_ball  \
		  1610612744     203110     76.21    39.70       0.00           8.47
		  1610612748       2548     58.97     5.88       0.00          40.30
		  1610612748       2547     86.70    21.05       0.00          29.35
		  1610612748       2736     71.32    45.91       0.00           2.19

		  closest_to_ball    player_name     player_jersey  
		   False             Draymond Green       23  
		   False             Dwyane Wade           3  
		   False             Chris Bosh            1  
		   True              Luol Deng             9  

Visualize the data

Add Context to the Data

Combine NBA play-by-play with Position data


• Deduce 'shooting side', time and location of:
  shot (made or attempted), assist, block, free throw, rebound, steal or turnover
• Assume possession if player is closest and within a distance of 2ft from the ball, for a consecutive string of 15 frames (0.6 seconds)
• Deduce player passed if he had possession, and his resulting action was neither a shot or turnover

Steps to the simulator

Steps to the simulator

2. Determine Offensive Player Probabilities
• Movement:
  $ \mathbb{P}(\text{move to region B} | \text{in region A}$).
• Possession:
  $ \mathbb{P}(\text{has possession} | \text{his team on offense}$).
• Action:
  $ \mathbb{P}(\text{action} | \text{in region A}$),
  where action $=$ pass, shoot, or turnover.

Steps to the simulator

Steps to the simulator

• A player is considered a defender if he is the closest player on the opposing team to the player with possession.

Steps to the simulator

Combine NBA stats with total number of defended possessions
• $ \mathbb{P}$(action) = action per game / num poss defended,
where action = steals, blocks, offensive rebounds, defensive rebounds

Steps to the simulator

• Defenders matched to offense players in the following order: position played, height, random


• $ \mathbb{P}_o(\text{shot})$ = $\max(0, \mathbb{P}_o(\text{shot}) - \mathbb{P}_d(\text{block}))$


• $ \mathbb{P}_o(\text{turnover})$ = $ \min( \mathbb{P}_o(\text{turnover}) + \mathbb{P}_d(\text{steal}), 1)$

Steps to the simulator

• Each play is approx. 20 sec $\Rightarrow$ 144 plays per game
• 6 actions per play $\Rightarrow$ 848 actions per game
• Player has constant performance based on the minutes played.
(Results of each player are scaled to entire game)

Simulate!

	
	     PTS    FGA    FGM   FG3A   FG3M  OREB  DREB     STL   BLK    TO    PASS
	GSW	
	mean 84.48  74.79  32.97 12.76  6.18  5.89  40.66   0.53  2.13  17.82  249.08
	std  11.33   6.37   4.76  3.98  2.85  2.37   5.71   0.62  1.43   3.94   23.26
	rmse 31.67  15.16  10.47 13.10  3.31  2.98   6.70   4.91  1.43   8.37  169.27


	MIA
	mean 88.22  89.46  37.75 16.56  4.24 21.24  39.62   1.09  2.89  14.16  276.95
	std  10.62   6.16   4.86  3.97  2.02  4.89   5.15   0.96  1.51   3.60   22.84
	rmse 43.70  25.58  18.86  6.81  2.40  4.88  21.18   1.45  1.94   6.93  117.67
	        

Simulation Outcomes

Next steps

Improve the model by:
• Using more games
• Using more simulations
• Using players that play more minutes
• Validating passing and possession models

Thank you