Revisiting Preseason Highly Plausible Projections

Mandatory Credit: Jayne Kamin-Oncea-USA TODAY Sports

For the second year in a row my preseason win predictions narrowly won the APBR[1. Association for Professional Basketball Research] contest for projecting the records of every team in the NBA.[2. Very very very narrowly over Nylon’s own Nathan Walker.] Like most of the systems analysts use in the contest, I use a combination of the expected playing time and metrics to estimate player production, in my case a combination of my Player Tracking Plus Minus (PT-PM) and an regularized adjusted plus minus (RAPM). One of the sources of error comes in simply from changes from the expected playing time by players, from injuries, player movement, or evolution in the coach’s rotations.

Now that the season is over, I can estimate the net amount of error from having inaccurate playing time estimates by running the preseason player efficiency projection through the actual minutes played by every player on the roster in what is known as a retrodiction, an exercise I did last year as well. In addition to providing a measure of the error introduced simply by missing on the minutes projections, the exercise also gives an estimate of which teams added the most talent weighted by playing time or lost the most expected production.

As I found last year, the minutes projections overall did not appear to be a major source of error on net for my preseason predictions. One partial explanation is that players performing expectations are likely to lose minutes to players with improved performance, as there is in season correlation to PT-PM’s estimate of performance and playing time.

Overall, there was a modest reduction in average absolute error[3. Error whether positive or negative] and slightly more improvement in root mean squared error[4. A measure that puts larger penalties on big misses] (RMSE):

• Absolute error fell 4% from 5.7 to 5.5 wins.
• RMSE fell 7% from 7 to 6.5 wins.

The team by team results were fairly interesting as well:

• My predictions for thirteen teams were one win away, either closer or further, from the final results using actual minutes played in the retrodiction compared to preseason estimate. While two predictions were no different
• The biggest improvement, both estimated strength of the rotation and in accuracy of the prediction was for Portland with a seven game improvement in both the accuracy of prediction and in the estimated strength of the team fielded. It is not too surprising that preseason minutes  were off for the Trail Blazers had one of the highest roster turnovers in the league. Dallas also improved in the estimate of team production by seven games with Wes Matthews, Zaza Pachulia and Dirk Notwiski playing more than my projections
• On the other end, the Chicago Bulls narrowly beat out the New Orleans Pelicans for the most “lost” production over the course of the year. For the Bulls the issue was a combination of injuries to key players like Joakim Noah and more minutes for players like Doug McDermott, not projected to be good at all after his struggles in his rookie year.

The rest of the results are below with the retrodiction estimate, pre-season estimate, actual wins in 2016, error in both predictions, the difference in the net error, and the estimated gained or lost production:

So, like last year, the minutes projections did not appear to be the major source of error in my pre-season projections. I will have to look to improve my player metrics and fine tune some of the adjustments for special situations like returning rookies, players returning from injuries and players changing teams[5. Who tend to have less stable output than returning players]. And I still have no idea what happened to the Houston Rockets.