LIVE NBA Model: Sparse Impacts Model
The Data Jocks’ NBA Model estimates assigns each team an overall rating. It does this by rating each team’s best player, their second best player, and wrapping up the other 13 into a “remainder” rating. The overall rating is the sum of these three parts. The table below summarizes the Sparse Impacts Model results as of February 27 2024 Below the table is some information about model accuracy.
Click here to read about the intuition behind the sparse impacts model
Click here to read about testing the sparse impacts model.
Sparse Impacts Model
| Team | RTG | Player1 | RTG1 | Player2 | RTG2 | Remainder |
|---|---|---|---|---|---|---|
| BOS | 11.0 | Derrick White | 4.5 | Jayson Tatum | 1.1 | 5.4 |
| PHI | 7.0 | Joel Embiid | 10.0 | Tyrese Maxey | 4.0 | -7.0 |
| OKC | 6.5 | Shai Gilgeous-Alexander | 7.4 | Chet Holmgren | 1.9 | -2.9 |
| MIN | 5.9 | Anthony Edwards | 3.8 | Karl-Anthony Towns | 1.0 | 1.2 |
| DEN | 4.6 | Nikola Jokić | 6.6 | Jamal Murray | 0.7 | -2.7 |
| NYK | 4.5 | Jalen Brunson | 6.7 | Donte DiVincenzo | 2.4 | -4.6 |
| NOP | 4.3 | Jonas Valančiūnas | 1.0 | Zion Williamson | -0.2 | 3.5 |
| CLE | 3.7 | Donovan Mitchell | 7.5 | Evan Mobley | -2.7 | -1.1 |
| LAC | 3.6 | Kawhi Leonard | 6.1 | James Harden | 0.3 | -2.8 |
| MIL | 3.5 | Damian Lillard | 6.3 | Giannis Antetokounmpo | 3.3 | -6.1 |
| DAL | 2.7 | Luka Dončić | 6.0 | Kyrie Irving | 3.8 | -7.2 |
| PHO | 2.2 | Devin Booker | 1.9 | Kevin Durant | 1.1 | -0.9 |
| GSW | 2.2 | Stephen Curry | 5.5 | Kevon Looney | 3.3 | -6.7 |
| SAC | 1.6 | Domantas Sabonis | 2.7 | De’Aaron Fox | 2.4 | -3.5 |
| IND | 1.5 | Tyrese Haliburton | 4.8 | Isaiah Jackson | 1.1 | -4.4 |
| HOU | 1.0 | Fred VanVleet | 3.5 | Alperen Şengün | -1.6 | -0.9 |
| LAL | 0.9 | Anthony Davis | 5.2 | LeBron James | 1.8 | -6.1 |
| ORL | 0.8 | Cole Anthony | 1.2 | Goga Bitadze | -0.6 | 0.2 |
| MIA | -0.5 | Kevin Love | 1.7 | Jimmy Butler | -1.7 | -0.5 |
| ATL | -2.4 | Bogdan Bogdanović | 2.2 | Trae Young | -0.2 | -4.3 |
| UTA | -2.6 | Kelly Olynyk | 3.5 | Lauri Markkanen | 3.1 | -9.2 |
| CHI | -2.6 | DeMar DeRozan | 1.0 | Alex Caruso | -1.2 | -2.4 |
| BRK | -3.1 | Mikal Bridges | -0.2 | Nic Claxton | -1.0 | -1.9 |
| TOR | -3.3 | Scottie Barnes | 8.3 | Pascal Siakam | 3.4 | -15.0 |
| MEM | -5.7 | Desmond Bane | 2.9 | Jaren Jackson Jr. | 1.1 | -9.7 |
| SAS | -6.9 | Devin Vassell | 3.3 | Keldon Johnson | 1.9 | -12.1 |
| POR | -8.4 | Matisse Thybulle | -0.1 | Jerami Grant | -0.5 | -7.8 |
| WAS | -9.2 | Daniel Gafford | 1.4 | Tyus Jones | 0.9 | -11.4 |
| DET | -11.1 | Marcus Sasser | -2.1 | Bojan Bogdanović | -2.8 | -6.3 |
| CHO | -11.6 | LaMelo Ball | -0.2 | Terry Rozier | -0.2 | -11.2 |
Sparse Impacts Model Performance
Year to date, the sparse impacts model has accurately predicted 66.1% of games. Compare this to the other results for popular models found here.
In predicting the line for games, our model has an absolute error of 11.0 points. This means that the difference between the actual game score and the predicted margin of victory was 11.0 points. The root mean squared error (RMS error) in predicting margin of victory is 14.1 points.
Sparse Impacts Model Hyper Parameters
As discussed in our introduction article and test and evaluation article about the sparse impacts NBA model, there are a few hyper parameters to be tuned. For transparency, we include these choices on this page.
Instead of rating the top N players in the league, we decided it was better to rate the top 2 players on each team. This decision was made because the optimal number of players to rate was somewhere between 30 and 60. Choosing to fix the number of players per team simply results in easier-to-digest data and model results.
Our model uses a home court advantage of 3 points. This parameter was not learned like player ratings, it was tuned as a hyper parameter in increments of 0.5 points.
Our model uses value over replacement player as a prior to bootstrap the small sample size and reduce impacts of the multicollinearity problem. The weight of the prior is equal to 20 games of data in rating a player. This means that after a player has been in 20 games, the prior and the observed data are equal weight. At the beginning of the season, the prior is more impactful. Later in the season, observed game data dominates.