How We’re Improving Our NBA Forecast For 2022-23
It all comes down to who plays and how much.
Ever since we introduced a depth charts-based method for keeping track of NBA rosters in our NBA forecast model, one of its biggest recurring criticisms — from those outside the ranks of Boston Celtics haters, that is — has been that the playing-time projections are just off for certain teams and players. That’s fair: Though we’ve done our best to apportion the 240 individual minutes available on each team per regulation game, the results have not always been completely precise. And since predicted minutes play such a heavy role in determining a team’s performance rating for each game, these differences in playing time affected our game-to-game projections.
But like any NBA player trying to get better, we spent the summer locked in the (computer) lab improving our game. Specifically, we’re making a tweak this season to how we project minutes played, at least for games in the near term. For games within the next two weeks of the current day, we’ll be blending our existing playing-time projections with what we’re calling a “history-based” minutes projection.
So what exactly does that mean, and what has changed?
Well, at first, it’ll seem like nothing is different. But once the 2022-23 season really gets going, we’ll start integrating the new history-based minutes projections into our overall playing-time forecasts. The history-based projections consist of a rolling average of the actual minutes played in recent games by each player, multiplied by their projected availability for today’s game.1 For a game being played today, that rolling average will get 60 percent weight when projecting a player’s minutes, while our classic depth charts-based projection will only get 40 percent weight.
When researching this, we calculated a rolling average of players’ actual minutes played over the past five games. We used data from the last five games that a team played within the past 15 days, during which the player played at least 1 minute.2 Ideally, we would use a rolling average of each player’s five previous games, but if, say, the player played in only four games, we would use that data anyway. Basically, we used as much previous game data as possible (up to five games ago) to calculate rolling averages for each player. And if a player hadn’t played in any of his team’s five most recent games within the last 15 days, then we stuck with our tried-and-true algorithmic, depth chart-based projections. We then multiplied each player’s rolling average by their projected availability.
Each player will get a fresh start on their history-based minutes projections at the beginning of each season and the playoffs,3 so it will take a little while to see the new projections in action after the season starts or moves into a new phase. And in the long term — beyond a couple of weeks into the future — we found that the old depth chart-based system does a better job than the new history-based system. (That’s why we gradually phase out the history-based projections when forecasting future games, eventually dropping their weight to 0 percent — and boosting the depth charts-based projections to 100 percent — for games 15 days in the future and beyond.)
But when it comes to games in that short-term sweet spot, this new method should make for improved forecasts — hopefully, decidedly so. Based on our backtesting, incorporating those rolling averages helps improve the accuracy of our projections by a surprising amount, especially when blended with our original playing-time forecasts. If you imagine a spectrum spanning from relying purely on depth charts to having perfect information about how much each player would play in each game, our new method is situated about halfway in-between.
So where does this all leave us for 2022-23? Who are the winners and losers of this adjustment to our forecast model?
Well, we won’t know until after the season starts. But for now, just remember that if our model seems off on some particular team, faulty playing-time projections won’t be to blame nearly as often anymore.
Jay Boice contributed research.