Evaluating the shooting in Steph Curry's 2015-16 MVP season

Introduction 

Steph Curry's second MVP season - the 2015-16 NBA season - is widely considered to be one of the greatest basketball seasons of all time:

The greatest MVP season in NBA history (Fox Sports)
Stephen Curry had the best NBA season ever (Time)
Steph Curry had the greatest season in NBA history (USA Today)

That year, his team - the Golden State Warriors - set a new NBA record for number of games won in the regular season, winning 73 games and losing only 9. Curry had won the MVP the previous year, in 2014-15, but had such a great leap forward in his play that some even suggested that he should win Most Improved Player (MIP) in addition to MVP.

Although Curry's clutch play, team leadership, and great passing all contributed to the incredible season he had, it was his shooting, particularly his three point shooting, that really elevated him above his competition.

For that reason, I decided to evaluate the shooting in Steph Curry's 2015-16 MVP season, using statistical methods available in the Python libraries: NumPy, Seaborn, Pandas, and Matplotlib.

Evaluation

For more detail on the methods used, see the methodology.

EDA

First I created a database containing every game Steph Curry had played in his career, as well as the top 100 shooting seasons of all time. This was done using the open source data available on Basketball Reference.

I then performed exploratory data analysis (EDA) on the datasets, previewing all the data and plotting histograms, boxplots, and empirical cumulative distribution functions (ECDF) for the key characteristics (3P%, 3PA, TS%) of the top 100 shooting season data. These included a comparison to the corresponding value in Curry's 2015-16 season. The graphs can be found here:

3P%

3PA

TS%

Permutation Hypothesis Testing

Whole Career

The EDA highlighted that the greatest outlier of Curry's season was the number of three point attempts he took (3PA). His 3P% and TS% were remarkable, but fell within what we would expect for an "all-time great" shooting season. To gauge just how extraordinary this 3PA statistic was, I performed a hypothesis test, with the null hypothesis: 'Steph's volume of three point shooting in 2015-16 is consistent with his volume across his career.' The test statistic was total number of three point shots made.

This was a hypothesis of exchangeability, so I generated 50,000 random permutations of 79 game seasons using every game Curry ever played as a base. Then I plotted an ECDF of all 5000 seasons to compare with Curry's 2015-16 season. The p-value shows the probability of having a season as extreme as Curry's 2015-16 season, assuming the null hypothesis is true.

We can see from the ECDF that not one of those 50,000 permuted seasons came close to the number of 3PA that Curry shot in 2015-16. The p-value of 0.00 allows us to reject the hypothesis, and conclude that Curry's 3PA in 2015-16 was significantly different from his 3PA volume across his career.

p-value: 0.00

2015-2018

After the 2015-16 season, the Golden State Warriors signed the basketball superstar Kevin Durant. Since then, they have been virtually unstoppable, winning two championships in a row, but Steph has not won any more MVP awards (despite winning them in back to back years before Durant's arrival).
 
To judge whether Steph has been shooting at the same level since then, I performed a second hypothesis test, with the null hypothesis: 'Steph's volume of three point shooting in 2015-16 is consistent with his volume since then.' This was performed using the same method as the first, but using the data of Curry's seasons from 2015-16 until 2017-18.

The ECDF shows that very few of the randomly generated seasons had as many 3PA as 2015-16. The p-value of 0.00768 is lower than standard significance levels of 0.05 and 0.01 - thus I reject the null hypothesis and can conclude that Curry's volume of 3PA in 2015-16 is significantly different from how he's been shooting since then.

NB. As the p-value is much closer to the significance level with the 2015-18 data, the conclusion is not as strongly asserted as the conclusion was with Curry's career data.

p-value: 0.00768

Hypothesis Testing of Correlation

Curry's three point shooting volume was an outlier but his three point shooting percentage was merely very good. How should we assess this? Does he deserve more credit for shooting at such a high percentage of success, given how many shots he took? Or was his breakthrough simply shooting as much as he did?

To roughly answer that question, the correlation between 3P% and 3PA needed to be calculated. To do this I performed a hypothesis test, with the null hypothesis: '3P% is not correlated with 3PA for all time great shooting seasons'. The test statistic was the Pearson correlation coefficient.

I generated 50,000 random permutations of 100 players worth of 3P% and 3PA statistics. This was to assume no correlation. I then calculated the Pearson correlation coefficient for each permutation of 100 players and plotted it an ECDF to compare to the correlation coefficient of the real data. The p-value shows the probability of having a correlation as extreme as the real data, assuming the null hypothesis is true.

The ECDF shows that there is likely no correlation between 3P% and 3PA among all time great shooting seasons. This is evidenced by the fact that the Pearson correlation coefficient of the real seasons lies within the interquartile range of the Pearson correlation coefficient of the 50,000 permuted top 100 seasons.

The p-value of 0.6859 is much higher than all practical significance levels; thus, we accept the null hypothesis and conclude that there is no correlation between 3P% and 3PA among all time great shooting seasons.

Pearson-r value of real data: -0.05065
Mean Pearson-r value of permutations: -0.00033
p-value: 0.6859

Conclusion

Steph Curry's MVP season was extraordinary for many reasons, but chief among them was his three point shooting volume (3PA). This was far higher than any great shooting season in history; and significantly higher than the rate Curry himself has shot during his career. Since then, he has shot at a significantly lower (but still elite!) rate.

Curry's three point accuracy (represented by 3P%) was elite during the season, but not as much of an outlier as his 3PA. The credit he deserves for keeping up such an accuracy while shooting such a high volume is limited: among the great shooting seasons, 3P% is not correlated with 3PA. The real breakthrough was the number of 3PA taken during that season.

All the data, graphs, and the Python code can be found on the GitHub repository for this project.