Data Analysis on NBA Success
The skills the author demoed here can be learned through taking Data Science with Machine Learning bootcamp with NYC Data Science Academy.
Motivation
Over the last decade the NBA always seems to be dominated by the same franchises. Most recently, the Golden State Warriors have gone to 5 straight championships, winning 3 of them. As a lifelong fan of a not so successful franchise (I’m from New York), I wanted to explore what makes teams successful. I sought to explore three questions for this data project:
- Can teams’ wins and losses be successfully predicted?
- What kind of players go onto successful teams?
- Given salary constraints, how can a competitive roster be more efficiently constructed?
Data
All data was scraped from NBA.com for individual player and team stats for the past 25 years. Given that the NBA website has dynamic tables I decided to use Selenium to scrape the data. Selenium can be used to drive through the web page and interact with the dynamic table.
Data Analysis
The three point shot has become the hallmark and standard if you want a competitive team in the NBA. One of the largest trends over the past decade has been the increase in average three point shots taken per game. The reason for this is simply that three points is worth more than two points, so players have abandoned the mid range game in favor of taking more threes. We can see from the graph below that average three point shots have more than doubled for the league as whole since the 2006 season.
However, teams have not gotten much better at shooting threes as we can see that their three point made percentage has remained relatively stable. The increase in threes has had an impact on effective field goal percentage (eFG%). eFG% is an important metric for players and teams as it takes into account the increased value of threes. The growth in threes taken has led to a several percentage point jump in eFG% over the past several years as can be seen in the second graph.
NBA 4 Factor Data Model
The first question I wanted to answer was what factors can help in determining team wins and losses. In order to do this I looked at the 4 factor model developed by Dean Oliver. As a quick introduction to the 4 factor model, the 4 factors determining success in the NBA are straight forward:
-
Scoring
This is measured by eFG% mentioned earlier. This gives a sense of points per field goal attempt by a team. Again, this metric adjusts for three point shots made. -
Rebounding percentage
This is somewhat intuitive, if you don’t score you want to grab rebounds. The higher percentage of available rebounds you get the better. -
Free throw rate
This metric gives a sense of how well a team does at getting free throw shots for each field goal attempted. In basketball, free point opportunities are a good thing. -
Turnover percentage
Don’t turn the ball over! You don’t want to lose precious scoring opportunities and give the other team more chances.
There is one more thing to mention for this, the 4 factor model is actually 8. It is actually 8 factors because teams play both offense and defense, so each of the 4 factors also has a defensive equivalent measure.
Plotting the correlation between the 8 factors and number of wins for the season gives us a sense that all of these factors are strongly correlated with wins. The factors are also in the direction that we would expect. For example, scoring is important and we can see eFG% and opponents eFG% have strong correlations (greater than 0.8) and eFG% has a positive relationship with wins and opponents eFG% has a negative relationship.
I wanted to use the 4 factor model to see how well these factors help predict teams wins and losses. In order to test this I ran a linear regression with wins between the independent variable we want to predict. The coefficients were significant and in the direction that we would expect. Additionally, the r squared was over 95%.
Predicted Wins vs Actual Wins
To get a sense of how accurately this model could predict wins for the season I plotted the predicted wins against the actual wins for each team.
The predicted wins were very close to the actual wins, and all predictions were within 1 standard deviation of the actual results. 2015 was an interesting year, it had both the largest difference between the expected wins (the Philadelphia 76ers) and actual wins for a team and also had the highest estimate wins for a team (Golden State Warriors).
This year the 76ers were the worst team in the league and were expected to be throwing the season in order to get the number 1 draft pick. Perhaps, this prediction could be used to search for teams that were intentional bad/ throwing the season in order to chase a high draft pick. 2015 was also the year the Golden State Warriors put up the best regular season record in NBA history. The expected wins captured this as it had the Warriors as the highest expected wins over the past 15 seasons. Overall, it seems that these factors do quite well at helping us predict expected wins.
Additional Work
As next steps, I would like to explore the factors and individual player contributions. This can give us a sense on how to construct a team that would improve each of the factors. Additionally, traditional positions are somewhat restrictive and outdated. Player skills have evolved and many players are multi positional. I would like to apply cluster analysis in order to group players into more natural position groups.