2. 15 players on a team
5 players on the court
The coach rotates players
between the court and the court
Data-driven approach:
real-time performance prediction
Performance predictions
3. Human behaviour is non-Markovian
Positive: shots, assists, steals… Negative: turnovers, fouls…
Plays come in streaks
4. Human behaviour is non-Markovian
3
3
4
5
4 3 5
Plays come in streaks
Positive: shots, assists, steals… Negative: turnovers, fouls…
5. Play Predict :: algorithm
Basketballvalue.com
API
Bulk download Scrape
Pandas
DataFrames
MySQL
6. Play Predict :: algorithm
Basketballvalue.com
API
Bulk download Scrape
Probability
density
functions
Performance
predictions
Pandas
DataFrames
MySQL
Python Flask,
Bootstrap,
Javascript,
AWS
~150 players
~2000 events/player
Unique model
for each player
7. Play Predict :: validation
Training set
Testset
0.90.2
0.2
0.9
Validation set:
Played over 75% of
games (89 players).
Testing predictions
to continue streak
from 2 to 3.
Pearson R = 0.7
9. Play Predict :: performance metrics
Streaks
1. Chris Paul
2. David Lee
3. Beno Udrih
4. Jose Calderon
5. Pau Gasol
Performance per game
1. Chris Paul
2. Lebron James
3. Dwayne Wade
4. Deron Williams
5. Steve Nash
Performance per play
1. Jose Calderon
2. Jason Kidd
3. Steve Nash
4. Chris Duhon
5. Sean Singletary
(Accepted) Value over replacement player:
1. LeBron James
2. Chris Paul
3. Dwayne Wade
4. Brandon Roy
5. Pau Gasol
Most VP: LeBron James