A Deeper Dive into ESPN's Win Projection
ESPN released their initial 2018-19 NBA regular season win projections today. The piece is worth reading in total; it's paywalled but ESPN Insider is a good deal for anyone who can afford it. The projections are based primarily on ESPN's Real Plus-Minus statistic and playing time estimates made by Kevin Pelton. The exact formula for both RPM and the projection system are secrets of the Worldwide Leader
The headline amongst Celtics fans will certainly be their 53-29 projection. That's going to raise the ire of some Bostonians as it doesn't seem to make sense for a team that won 55 games last season and functionally adds Gordon Hayward plus has a ton of young players on the up.
It's worth discussing why their projection is coming in "low," and if it really means anything.
Over-performing Point Differential
Stats like RPM that are on/off based are really working off a point differential and not a team record as their baseline. This is fine/good as team point differential has been a better predictor of future record than raw wins.
Using the formula that Daryl Morey developed years ago to predict a team's record from their raw points scored and allowed (generally referred to as "Pythagorean Wins") the Celtics played to the level of a 51 win team despite winning 55. That's the baseline that the projection system is working off of, so in one sense it's predicting a two win improvement, not a two loss hit.
A common reaction I see to this is "but that's because of Brad Stevens." While it's conceivable that Stevens steals an incremental win across 82 game with particularly brilliant late game sets, that's about all I can imagine even the best coach conjuring. The other coaches aren't morons; a great set might increase the odds of scoring by a few percent over a "normal" one and then there are only so many game opportunities for that to even matter.
The impact of a good coach would more likely be in things that would show up in point differential, not in wins above expected from that differential. The work that Brad has done with maximizing the impact of individual players, and with full game management, would be baked into that point differential, for example.
It's true that the Celtics have over-performed their differential in the last two season, but they under-performed it in the prior three and Stevens coached all those years, too. Again, it's conceivable that he's become a markedly better coach in some way that related to over-performing point differential, but not to the extent that we've seen between his third and fourth season.
If you accept that the team that Boston actually put on the floor last year was really more of a 50-52 win group and not a 55-win squad, it adjusts your reaction to the new projection somewhat.
Models Squeeze the Ends
The second thing to realize is that models like this always depress the top-end teams and lift the worst clubs up. The reality is that, other than some small number of outlier teams, the league basically sits between 18 and 62 wins.
If you devise a projection by running a whole bunch of simulations and taking an average, a bad simulation for a good team is simply going to have more of an impact than a good one can. If you think a team is a 58 win group, their "good" sims are going to show maybe a 65 win output while their "bad" sims, and a really rough injury season would be baked into the model's assumptions, could come in closer to 40 wins. The average of high and low when it gets progressively harder to add a win at the top end or loss at the low end is going to naturally pull individual teams towards the middle.
The best team in the model is the Warriors at 58 wins and they're three clear of the Raptors in second. The worst teams all have 25 wins. If you take the average top and bottom wins across the league in all the sims you would certainly get higher and lower numbers than that, but they would be spread across multiple teams.
Injury and Age
ESPN doesn't publish the exact formula that they use, but we can look at the 538 CARMELO model's projections for individual players to highlight some other factors that go into the Celtics' estimates.
In Pelton's write-up he notes that the system is conservative on Hayward and Horford. CARMELO is similarly bearish on Hayward because the history of players returning from major injuries is so spotty. It's possible that Hayward will take a year to get back to his old self, or even that he'll never be the same player again. It's also possible that he's immediately an All Star when he steps back out there. On average, you have to expect him to underperform his pre-injury self if you're projecting from historical comps.
If he has a bounce-back season and eliminates all the worst case scenarios, his projections for the years after will go way up. He has to prove it first, though.
With Horford it's a case of age. His RPM-measure performance improved from 2016-17 to 2017-18 which you wouldn't expect from someone over 30 years old. For that reason alone, most projections are going to pull him back down for the coming season. Considering that the team (smartly) seems to manage his regular season wear-and-tear to maximize his playoff impact, that's reasonable from both a stats and scouting perspective.
If Horford plays to the level he did last season and Hayward returns to near his pre-injury level in heavy minutes, the Celtics will likely beat their projection by a comfortable margin. Neither of those things are a given.
It's also worth remembering that development isn't always a straight line. Players like Jaylen Brown and Terry Rozier who made significant year-over-year strides may take a step back next season. Jayson Tatum, who had one of the greatest age-19 seasons in NBA history, may not show huge growth in year two because his baseline is already so high. If those three players stay healthy it's almost certain that their best years are ahead of them, and that their peak performance levels will be substantially better than last season, but there's no guarantee that they'll produce more next season on their way to their primes.
What Really Matters?
When looking at the output from a model like this, checking your team relative to the league is a more useful exercise than comparing to your raw win expectations.
The ESPN model projects that Celtics to have the fourth best record in the league, significantly behind the Warriors (fine), two wins behind the Raptors (who they were four actual wins behind last season), and 0.2 wins behind the Jazz (who had a 54 win differential last year).
Pegging Boston in a tier with Utah and Houston, trailing behind Golden State and a few wins short of Toronto seems completely reasonable. The fact that they're 5.5 games clear of their next closest competitor in the East, the 76ers (who only get projected for 50 games out of Embiid, and so have reasons of their own to look more deeply at the system), is a very good sign.
In total, the Celtics projection has them two wins ahead of last season by point differential. It has them picking up seven games on Toronto, from a 60-51 point differential gap to 55-53. It has them well clear of the 76ers and the rest of the chasing pack. What looks like a disappointing projection is actually quite good!