Your work is really good, especially considering that 538 is no longer around. Do you have any links to ELO ratings from previous weeks, just as a way of seeing how teams have improved or regressed? I know that on 538 they had it so that you could look at rankings from previous weeks/years?
ok cool, I do essentially the same to standardize DVOA against FPI, GPF, etc. I wonder why you use the /25 math to scale Elo though? By the same logic you use for the dolphin predictive values, if you scaled Elo against the average distribution of various PPG ratings you get largely different values. for instance, Ari adjusts to -5.9 (using the /25 math) vs roughly -7.6 when scaled against other power ratings scales. Baltimore (my team) would be 8.9 vs 11.4. Could you explain your rationale? Thanks.
Late comment, but dividing by 25, means that team's that are favored by 400/25 = 16, should win 10/11 times (91%). It makes the elo scale match up with how Vegas does point spreads.
Typically in Vegas odds you'd see that,
+2 = 55%
+3 = 60%
+4 = 65%
+6 = 70%
+7 = 75%
+9 = 80%
+12.5 = 85%
+14 = 90%
The reason the elo ratings are more "compact" is because 1. they are predictive rather than explanatory 2. probably have less information about injuries, starting QB, etc. so are necessarily more conservative estimates that are "regressed to the mean"
Honestly it was just out of convenience... That "divide by 25" rule of thumb is what I've always used to convert between NFL Elo and pointspreads. I haven't really looked into how accurate it is in quite a while, so it's possible it needs updating.
My pleasure. Someday, somewhere I'll revive the QB adjustments too. Just too much else going on and not enough time to get it up and running by the start of the season.
Your work is really good, especially considering that 538 is no longer around. Do you have any links to ELO ratings from previous weeks, just as a way of seeing how teams have improved or regressed? I know that on 538 they had it so that you could look at rankings from previous weeks/years?
The whole Elo archive is here: https://github.com/Neil-Paine-1/NFL-elo-ratings
How are you deriving your PPG Ratings for Elo and the Andy Dolphin Ratings? Thanks (Sorry if I overlooked that somewhere).
Elo is easy, you just take the difference in rating from league average and divide by 25 (as per our handy explainer: https://fivethirtyeight.com/features/introducing-nfl-elo-ratings/) ... For Dolphin, I had to rescale his "predictive" values from here (http://www.dolphinsim.com/ratings/nfl/) into point ratings by converting them to a common scale with the other point-based ratings using Z-scores.
ok cool, I do essentially the same to standardize DVOA against FPI, GPF, etc. I wonder why you use the /25 math to scale Elo though? By the same logic you use for the dolphin predictive values, if you scaled Elo against the average distribution of various PPG ratings you get largely different values. for instance, Ari adjusts to -5.9 (using the /25 math) vs roughly -7.6 when scaled against other power ratings scales. Baltimore (my team) would be 8.9 vs 11.4. Could you explain your rationale? Thanks.
Late comment, but dividing by 25, means that team's that are favored by 400/25 = 16, should win 10/11 times (91%). It makes the elo scale match up with how Vegas does point spreads.
Typically in Vegas odds you'd see that,
+2 = 55%
+3 = 60%
+4 = 65%
+6 = 70%
+7 = 75%
+9 = 80%
+12.5 = 85%
+14 = 90%
The reason the elo ratings are more "compact" is because 1. they are predictive rather than explanatory 2. probably have less information about injuries, starting QB, etc. so are necessarily more conservative estimates that are "regressed to the mean"
Honestly it was just out of convenience... That "divide by 25" rule of thumb is what I've always used to convert between NFL Elo and pointspreads. I haven't really looked into how accurate it is in quite a while, so it's possible it needs updating.
I make my picks based on your composite rankings - any chance for an update this week?
Yep! Thanks for the reminder.
Fantastic! Thanks for all your work!
Do you have the playoff rules coded so you could add a column for the odds to make the playoffs?
Sadly, I did not have time to do that before the season started. (The tiebreakers are always a pain in the behind to program!) For that, I think the NY Times' Playoff Picture interactive is great: https://www.nytimes.com/interactive/2023/upshot/nfl-playoff-picture.html
In fact, they have that whole thing set up as well or better than we had back at 538. They even use Elo as part of the ratings: https://www.nytimes.com/2023/10/13/upshot/nfl-playoff-odds-simulator.html They say it's blended with betting market ratings, so I have to assume it's mixed with something like this: https://stats.inpredictable.com/rankings/nfl.php And honestly that's exactly how I would do it.
How do u calculate the win probabilities?
Winprob = 1 / (10^(-ELODIFF/400) + 1)
https://fivethirtyeight.com/features/introducing-nfl-elo-ratings/
Question, since QBs aren't factored in should we be factoring in the difference based on the lines provided?
Yes, I would mentally account for QB absences on top of these lines using something like this: https://ftw.usatoday.com/lists/nfl-quarterbacks-against-the-spread-worth-ranked-backups
Wow bro, thank you so much! This is super helpful!
I lived on 538 last season looking at all these data points.
I am so glad I found them again! Thanks for the time and effort. My time will be spent here a lot!
Wow, big swing on expected margin for the Jets-Broncos game based on the Bears Thursday night win.
That shouldn’t have changed (other results don’t affect Elo until a team plays), lemme see what’s going on there.
OK, I checked and it was saying DEN -1.5 before today's update too. Maybe you were looking at a different game?
Yep, my mistake.
No worries! Thanks for reading
Neil, thank you for keeping these going!
My pleasure. Someday, somewhere I'll revive the QB adjustments too. Just too much else going on and not enough time to get it up and running by the start of the season.
Does win probability mean win against the spread or just an out right win?
Just a win straight-up.
Neil, you have the Raiders-Steelers game as OAK -.5. The Raiders are in Las Vegas! Should be SIN CITY -.5.
Oh man, I need to update my team IDs! Lol. Thanks for the catch!