Evaluating the 2017–18 NBA Win Total Projections
Finally. It’s now April and we finally get playoff basketball. We have had phenomenal regular season story lines, but at last we get too see games with higher stakes, harder fouls, and the tactical beauty of the playoffs.
But, before we revel in postseason play, it is important to evaluate the regular season before it is only a fleeting memory. Back in October, I hurriedly projected the win totals for every NBA team. Though an important, fun, and useful exercise, predictive modeling of this kind is only valuable if you can evaluate the model’s performance. So, let’s take some time to see how this model performed.
I warned in October that this model was far more risky than others of its kind, predicting the top teams in the league to be truly dominant and the bottom teams to be historically bad. Thus, I knew the model would struggle, but I was surprised at how well it actually did. The model contained a Root-Mean-Squared-Error (RMSE) of 10.13 wins. This simply means that on average, the model was off on a given teams total by 10.13 wins.
Yes, this sounds atrocious. And it is in a lot of ways. I tracked several other “expert” win total forecasts and they performed just slightly better, yet none were as successful as Vegas (surprise!) with an RMSE of 7.11 [1]. However, the model predicted 13 of the 16 playoff teams correctly (7 in the West and 6 in the East).
While the error in my model was high, there were a number of teams with unique circumstances that contributed to the poor performance. Philadelphia outperformed its projected win totals by 25.5 wins, while San Antonio lost 18.6 more games than expected with the injury to Kawhi Leonard. Injuries and trades create a substantial amount of variability in predicting something as long and volatile as an NBA season. Who would have thought Kawhi would have played less than a dozen games or that DeMarcus Cousins would be sidelined half the year or that the Oladipo trade would work out so well for Indiana?
The errors in the model do show a normal bell-shaped distribution which does speak to the model’s construction. There is much to be learned from this modeling process, and most of my concerns from October have been realized since. Namely, we must dive further into the minutia of the model and reevaluate player projections and player minute totals.
Overall, though not completely satisfied, I am happy with the first iteration of the NBA win totals model. It would have lost you money in Vegas — the model would have won 50% of its bets, but therefore lost money due to the house take — but provided a good illustration of what could be expected from the 2017–18 NBA season. And boy was it fun.
Now, enough of the regular season. Let’s enjoy the playoffs!
For further comments, questions, or critiques please email me at aepowell95@gmail.com or find me on twitter @aptigers12.
[1] The Vegas lines were from Westgate Sportsbook and adjusted to sum to 1230 wins for all teams in the league.
