The 2017-18 NBA playoffs season has begun, and I figured it might be fitting to write a post related to NBA playoffs championships.
I was looking at NBA teams’ regular-season rankings yesterday, provided via NBA Stats, and I was disturbed: There were too many rankings for too many metrics:
- Ranking by points scored
- Ranking by blocks/assists/steals/rebounds
- Ranking by field goal percentage
- Ranking by 3-point field goal percentage
- Ranking by field goal attempts
- Etc.
All in all, there were over 20 different rankings available that ranked teams based on the teams’ different regular season performances.
Lost in the breadth of these various rankings, I wanted to know which ones were relevant in predicting the playoffs champions. I asked myself, “Which one of these rankings can best tell me who’s going to win the playoffs championship?”
Using Python, I pulled the data from the 1996-97 season through the 2016-17 season. The nba_py library made this process incredibly easy and painless, for which I will be forever grateful.
Then, for each ranking method, I counted the number of times in which #1 ranked teams won the playoffs championships.
This is the result:
| Ranking Method | Number of Instances Where #1 Ranked Team Won Championship |
|---|---|
| PLUS_MINUS_RANK | 5 |
| W_PCT_RANK | 4 |
| PTS_RANK | 3 |
| FGM_RANK | 3 |
| FG_PCT_RANK | 3 |
| AST_RANK | 3 |
| PFD_RANK | 2 |
| BLKA_RANK | 2 |
| OREB_RANK | 2 |
| TOV_RANK | 2 |
| STL_RANK | 2 |
| REB_RANK | 1 |
| BLK_RANK | 1 |
| FG3_PCT_RANK | 1 |
| PF_RANK | 1 |
| FT_PCT_RANK | 1 |
| FGA_RANK | 0 |
| FTA_RANK | 0 |
| FTM_RANK | 0 |
| FG3A_RANK | 0 |
| FG3M_RANK | 0 |
| DREB_RANK | 0 |
In this small sample size (of 20 years), the plus-minus ranking method recorded the highest number of correct predictions of the playoffs winning teams.
Whereas, team rankings based on regular season’s performance of defensive rebounds, as important as they may be, didn’t seem to be a good predictor.
It’s definitely something to keep in mind, as we watch the playoff games.
P.S. The code that I used to perform this analysis can be found here.
