APBRmetrics

Can someone recreate a statistical system that ranks the Nuggets, Blazers, and Heat ahead of the Cavaliers? it even makes less sense when you consider that ELO uses prior year data which should help the Cavaliers.

Elo Team
1770 Warriors
1661 Spurs
1636 Clippers
1617 Jazz
1602 Rockets
1600 Raptors
1587 Celtics
1587 Wizards
1569 Heat
1563 Trail Blazers
1552 Nuggets
1545 Cavaliers
1543 Thunder
1526 Pacers
1519 Bulls
1502 Bucks
1482 Pelicans
1482 Grizzlies
1479 Hawks
1463 Hornets
1463 Timberwolves
1441 Pistons
1420 Mavericks
1393 Kings
1374 Knicks
1372 Nets
1367 Lakers
1352 Magic
1340 Suns
1338 76ers

It's ELO, the reward/penalty associated with a game scales depending the ELO of the team you face. When you lose a bunch of games to people rated far below you, your score quickly drops.
If you zoom in on this season you can see how each win/loss affected their ELO. Their wins seem to give them an average of 5-15 points while their losses seem to cost them 15-25 points (just quickly eyeballing it).
https://projects.fivethirtyeight.com/co ... #cavaliers

it probably lines up with a 15-20 game rolling net rtg or rolling srs

the results are pretty clearly ridiculous though, and i'm guessing they'll make some revisions along with their win prediction methodology in general this offseason

sndesai1 wrote:it probably lines up with a 15-20 game rolling net rtg or rolling srs

the results are pretty clearly ridiculous though, and i'm guessing they'll make some revisions along with their win prediction methodology in general this offseason

I don't see how a 20 game rolling average is supposed to be more predictive than full season stats. I thought 538 Elo calling card is predictiveness.

It's recency based, but they do use the full season. Cleveland was so awful in the last chunk of the season that they collapsed.

https://fivethirtyeight.com/features/ar ... nderrated/

my bad - i didn't mean to imply it's only based on 20 games of data, just that it's biased towards recent games like atc said. it's a sophisticated power ranking in my book.

if somebody has tested the predictiveness of 538's elo rating, i'd like to read about it.

The main problem with their ELO rating is it's a team level rating. So if Cleveland rests their big 3 versus a bad team and loses big, their rating drops quite a bit, which isn't exactly fair since the ELO of the lineup they used for that game was much weaker than their usual lineup. The CARM-ELO rating is probably better because it apparently factors in which players are available for future games when calculating team strength.

One of the reasons that people like Elo ratings is that they're pretty straightforward to calculate.

I'm not sure it's all spelled out exactly, but they do go into a reasonable level of detail here:
https://fivethirtyeight.com/features/ho ... o-ratings/

The model explicitly does not control for resting players.

I ran a prediction accumulator for the NFL using the 538 model as one of the candidates a couple of years back. It did quite well.

Nate wrote:The model explicitly does not control for resting players.

Which means it has zero shot of beating Vegas. Its recency weighting seems too extreme as well, which is another reason why it's too low on the Cavs.

Nate talked about recently on a podcast. With their ELO NBA model, they are not trying to beat vegas, just have a decent prediction with a model that is easy to consume/understand by the public.

Well if you want to reference a good model in an indirect fashion, all you need to do is check the closing line at Pinnacle. Because the guys like Haralabob with the best models are the ones shaping the NBA market on a daily basis.

The betting market ratings at Inpredictable derive a rating for each team using a blend of the closing spread and the actual results. Those will probably be a better indicator of team strength than 538's ratings or any other public computer rating.

http://stats.inpredictable.com/rankings/nba.php