The Minnesota Timberwolves franchise isn’t renowned for its defense. Jimmy Butler is Minnesota’s most recent player to receive DPOY votes. Kevin Garnett won the award in 2008 with the Boston Celtics. However, people knew he was a good defensive player with the Wolves. Besides Butler and Garnett, Trenton Hassell is the only other Minnesota player to receive votes.
But that’s about to change.
The debate is somewhat contentious, but Rudy Gobert has a real shot to win DPOY over Victor Wembanyama. He’s the current odds-on favorite according to most sportsbooks, despite Shaquille O’Neal’s disapproval.
Gobert’s resurgence to defensive greatness has fueled Minnesota’s historic season. The team’s 107.7 defensive rating leads over the second-place Celtics by 2.8 points. If it weren’t for Boston’s absurd season (they currently have the third-best net rating of all time), the Wolves would have the league’s best net rating. Their 6.8 net rating is second in the NBA — not bad after FanDuel made them the ninth most likely team to win the Finals this year.
Wembanyama is a game-changing defender, but the San Antonio Spurs have been abysmal this season. They are currently 19-58 and the 15th seed in the West.
Team wins and seeding haven’t been the most important factor for voters since the award began in 1983, or at least that’s what I determined in the analysis below.
Using data from Basketball Reference, I looked at every player that has received DPOY votes, their stats, and their final share of the voting points to help predict Gobert’s chances of winning for the fourth time in his career.
Not every stat is worth keeping when predicting a DPOY winner. However, it’s worth noting that FG% correlates somewhat strongly, relative to other stats, with DPOY voting share (which I’ll refer to as Share). Naturally, this is because winners have often been near-the-basket centers rather than three-point shooters.
Ultimately, I narrowed down 11 stats to be features for prediction models:
- Defensive Win Shares (DWS)
- Blocks per game
- Defensive Box Plus/Minus (DBPM)
- Win Shares (WS)
- WS per 48 minutes
- Total rebounds per game
- Team wins
- Team conference seed
- Steals per game
- Defensive Rating (DRtg)
- And games played
I did this mostly through the correlation matrix below:
You can see correlations with Share, the ninth item I listed on each axis, through the shade and hue of each intersecting box. The dark green box in the bottom left means that DRtg correlates positively with Rank, which was the player’s finish in the final voting.
DWS, DBPM, blocks, and total rebounds correlated the strongest with high a Share. Team wins were somewhat significant, and a player’s overall impact on winning beyond defense, shown via WS, was also highly important.
The actual model building isn’t particularly interesting. The dataset of players and stats needs to be split into two: a training set and a test set. Predictions will be made based on the stats and Shares in the training set. The test set is used to validate the model’s performance since the model is not built with data from the test set.
There are many different types of models to use. Many people use Linear Regression, but I also used Ridge Regression.
Our Linear Regression model looks at how players with high Shares performed statistically. Did players who racked up blocks perform better in Shares than those who swiped loads of steals? Did players with high Shares have more Defensive Win Share rebounds than those with lower Shares? The model will find out and be able to predict the Shares of a new player outside of the dataset, like Gobert and Wembanyama, based on their stats for 2024.
Ridge Regression works much the same way being a “branch” of Linear Regression, though it’s better for datasets full of correlated features.
Models aren’t all created equally. Some are more reliable than others. The R-squared value of a model describes how well data fits a model or how much the variance in the independent variable accounts for change in the dependent variable. Both models, Linear and Ridge, had R-squared values of about 0.16, which is poor, with Ridge Regression having a slight edge.
Models that predict human activity, such as DPOY voting where voters may be biased, don’t need to have an R-squared higher than 0.5. A value of 0.16 is not ideal. However, I will note we are not predicting a single value. Instead, we’re looking at how predicted values compare to each other and determining the 2024 DPOY “winner” based on which value is greater. The actual predicted Share for each player might be untrustworthy, but we’ll at least have a ranking.
I input Gobert, Wembanyma, Anthony Davis, Bam Adebayo, and Jarrett Allen into the model. I prorated each player’s team wins, and games played to complete the data. Their stats are as follows:
The results are encouraging for Minnesota fans. The Linear Regression model determined that a player with Gobert’s features would have a Share of about 22.6% (this doesn’t really matter). Davis was predicted to be 16.6%, Wembanyama at 15.1%, Adebayo at 10.5%, and Allen at 10.2%. Gobert is the clear winner via Linear Regression.
Ridge Regression outputted almost exactly the same results. Gobert was at 22.1%, Davis 16.2%, Wembanyama 15%, Adebayo 10.5%, and Allen 9.6%.
Whether consciously or not, voters will likely favor Gobert’s league-leading DWS, excellent overall WS and WS/48, and solid rebounding and block totals. Minnesota’s success this season certainly helps.
Wembanyama has had a remarkable rookie season and will almost certainly pick up a DPOY award someday. But based on former DPOY voting, Gobert should win in 2024.
