As @Neil_Paine talks about here, we also made a couple of important changes to our NBA projections this year. I'm going to simplify this a bit, but basically it comes down to PLAYER-BASED projections vs. TEAM-BASED projections.

LONG DORKY THREAD 1/

fivethirtyeight.com/features/what-…
In the past, we've really used two ways to project NBA games. One is TEAM-BASED solely using team performance, e.g. via Elo ratings.

Advantage: captures team "intangibles" (coaching, cohesion, recent form)

Big disadvantage: doesn't account for player personnel changes. 2/
Alternatively, PLAYER-BASED projections work by summing up the projections for individual players, e.g. by using our RAPTOR ratings. That is, team performance is assumed to be ~equal to the sum of the parts. This has the opposite set of strengths and weaknesses. 3/
If you had to pick one or the other, the PLAYER-BASED projections will generally be stronger. Even though basketball is a team sport, you just lose too much accuracy if your projections don't immediately account for, say, James Harden getting traded (if he gets traded) 4/
What's really best, though, *especially* for updating your ratings as the season goes along, is some sort of hybrid between TEAM-BASED and PLAYER-BASED. Whereas what we had last year was sort of on the extreme end of PLAYER-BASED with little/no team component. 5/
This gets a bit technical, but there are actually two sets of ratings we run for each player. One is called RAPTOR, which is what you see in the interactive here. But there's also a related system called PREDATOR or "predictive RAPTOR". 6/

projects.fivethirtyeight.com/2020-nba-playe…
PREDATOR was trained by looking at out-of-sample performance, especially for players who switch teams. So e.g. if Harden is traded to the Heat, PREDATOR would theoretically give you a better idea of his performance than RAPTOR. 7/
But what works best for projecting *player* performance may not be optimal for projecting *team* performance. Or at least, that's what we believe after looking at the problem this winter. What looks like "luck" to PREDATOR may really reflect a player's role with the team. 8/
Additionally, RAPTOR includes an important property that PREDATOR does not. Basically, it works backward by starting with the team's performance and forcing the player ratings to sum up to the team rating. In PREDATOR this assumption is relaxed. 9/
One issue with RAPTOR and other +/- systems is they can sometimes misallocate credit between teammates. So e.g. maybe they had Harden too high and Russell Westbrook too low. If you're forecasting how they'll do now that they're on separate teams, that could be a problem. 10/
But if you're forecasting how the Rockets are going to do when both Harden + Russ are still in the lineup, that's less of a problem. Under RAPTOR, the errors literally cancel out. Thus, RAPTOR itself is a TEAM-BASED rating in certain respects, whereas PREDATOR isn't. 11/
PREDATOR may nonetheless be better than RAPTOR for forecasting a player's performance in *future* seasons. But we were also using PREDATOR to account for a player's *in-season* performance and adjust his team's rating on that basis, when RAPTOR works better for that. 12/
So one change we've made is to now use RAPTOR for in-season adjustments. While our preseason projections did quite well last year, the in-season updating was pretty clunky, if I can be frank, leading to some... weirdness by the playoffs, and this should help with that. 13/
Another change is that a team's rating will now reflect a combination of its RAPTOR rating and its Elo rating. Elo, if you'll remember, is a purely TEAM-BASED measure, and it tends to put more emphasis on recent games. 14/
How do we weigh the RAPTOR component vs. the Elo component? It depends on how much continuity there has been in a team's lineup. If a team has high continuity, then team performance really does tend to be more robust so Elo may get 50% or more of the weight. 15/
Conversely, if there's a lot of turnover in the lineup because of injuries, trades, etc., looking at past team performance may not tell you very much. Instead, the projections will gravitate very heavily (in some cases 95%+) toward RAPTOR. 16/
Anyway! We're pretty happy with these changes, though we may take a longer, fuller look at RAPTOR this off-season when the system has 2 years under its belt and we have a full off-season to work with. In the meantime, we hope you'll enjoy our NBA coverage. Thanks, all. 17/17

• • •

Missing some Tweet in this thread? You can try to force a refresh
 

Keep Current with Nate Silver

Nate Silver Profile picture

Stay in touch and get notified when new unrolls are available from this author!

Read all threads

This Thread may be Removed Anytime!

PDF

Twitter may remove this content at anytime! Save it as PDF for later use!

Try unrolling a thread yourself!

how to unroll video
  1. Follow @ThreadReaderApp to mention us!

  2. From a Twitter thread mention us with a keyword "unroll"
@threadreaderapp unroll

Practice here first or read more on our help page!

More from @NateSilver538

18 Dec
At least with essential workers—if you can define the scope narrowly—there's an argument to be had about reaching herd immunity sooner. What will literally kill people is treating a broad set of preexisting conditions as being as important as age.
If you look at the research, virtually all of people at highest risk of dying from COVID are aged 70+. There are almost no preexisting conditions that matter remotely as much as age.

nature.com/articles/s4159… Image
But ACIP defines *more than 100m people* as having "high-risk medical conditions" that put them in the same priority tier as people age 65+ for vaccination. This is NOT following the science. States that want to save lives must give age higher priority. cdc.gov/vaccines/acip/… Image
Read 4 tweets
12 Dec
The chances that the election would be close enough that late-arriving ballots in PA would be enough to change the overall outcome was maybe something like 0.5% or 1%. (There were not many of these ballots.) So if that's what people were concerned about, it was overblown.
When I pointed out though that yes the Supreme Court could determine the result in a *very* close election but it would likely have to be very close indeed under certain specific scenarios, I got yelled at on this platform for not taking the chances seriously enough.
If you want to take the position that the election aftermath went even worse than expected in some ways, but better than expected in other ways—including the Supreme Court—that seems reasonable and prudent! It's imperative to take stock of what happened.
Read 4 tweets
10 Dec
This seems like a fairly realistic set of assumptions for what to expect on the COVID-19 front next year as vaccinations begin to roll out. Things start to get notably better by ~April but it takes until mid-to-late summer before we approach herd immunity.
covid19-projections.com/path-to-herd-i…
The one major downside scenario is if vaccines prevent disease but don't do much to curb *infection*. Otherwise though these seem like pretty middle-of-the-road assumptions. I think he may underplay the role of seasonality a bit, which could help in the summer.
The other thing that seems likely is there will be some pretty fierce debates in that interim period from April-July or so about faster or slower paths to reopening.
Read 5 tweets
9 Dec
I'd also note, in general, that there have been a lot of bad predictions from liberals when it comes to how the Supreme Court would behave. Every court term, including this most recent one, brings a major "surprise" ruling or two. It's been a blind spot for analysis.
Some of it may be that while it may be the Supreme Court has become more partisan, it's not nearly at the same hyperwarp speed at which Congress has become more partisan, so it provides an important constraint overall.
One perhaps-not-terrible heuristic is to think of the current SCOTUS as being.... Mitt Romney. It's certainly quite conservative and doesn't remotely endorse the liberal worldview. But it's not particularly partisan or Trumpist and it cares about its institutional legacy.
Read 4 tweets
9 Dec
There's a lot of "if it were closer or X and Y were different, SCOTUS would have stolen the election for Trump" in response to this, to which I have a few different responses:
a) Maybe! But this election was *pretty* close and the courts were *very* unsympathetic to Trump. What if it was Florida-in-2000 close? So close it wasn't clear who really won? Maybe that's different. But Florida was INCREDIBLY close, a once-in-several-lifetimes occurrence.
b) Here's a litmus test. Suppose on Nov. 2 I'd described the outcome of the election—Biden would win 4 key states by <=1 point, all of which have GOP legislatures and two of which have GOP governors, one of whom is Brian Kemp. Would liberals have been freaking out? (Yes.)
Read 4 tweets
7 Dec
Ossoff 48.7
Purdue 47.9

Warnock 49.2
Loeffler 47.0

projects.fivethirtyeight.com/georgia-senate…
These are just polling averages, FYI. We are not issuing probabilistic forecasts of the Georgia runoffs, not for any philosophical reason—they'll be back in 2022/24—but because our full-fledged Congressional model isn't really designed to handle one-off races like these.
To be honest that's probably for the best, because there are a lot of judgment calls in terms of what the "fundamentals" look like in this election. Is this equivalent to a midterm, in which you might expect a pretty red environment vs. Biden? Maybe not given how Trump is acting.
Read 5 tweets

Did Thread Reader help you today?

Support us! We are indie developers!


This site is made by just two indie developers on a laptop doing marketing, support and development! Read more about the story.

Become a Premium Member ($3/month or $30/year) and get exclusive features!

Become Premium

Too expensive? Make a small donation by buying us coffee ($5) or help with server cost ($10)

Donate via Paypal Become our Patreon

Thank you for your support!

Follow Us on Twitter!