hey neil. i had a question about your elo calculations. i am trying to recreate and am close but am off a little. here is the calculations for K which i have pulled from this link (including margin of victory) https://fivethirtyeight.com/features/introducing-nfl-elo-ratings/
K = (LN(ABS(PD)+1) * (2.2/((ELOW-ELOL)*.001+2.2))) * 20
Just wanted to follow up to let you know I’m 5-2 since I’m since incorporating your ELO and SRS rankings into my “system”. Should be 6-1 but my dumbass didn’t look at the injury reports and didn’t realize half the Saints offense was injured. Lost the other by .5 point lol. Thanks!
Sorry to bother you again, but I have another quick question. Is the win probability to win the game outright, or is that to cover the spread that’s shown next to it as well? Thanks!
It's just straight-up win probability, not vs the spread. (Putting rounding aside, odds to cover the Elo line shown should be, by definition, 50% for each team anyway, since Elo differential is what drives both the predicted spread and the win probability.)
Gotcha. Thanks a not for taking time to reply. I know you said Vegas uses much more sophisticated stuff with their lines and whatnot, but in theory since the Elo spread has Packers -9.5, and the actual line is -4.5, that could signal a strong play at -4.5, correct? I’m assuming if it’s a 50% probably offer the Elo spread of -9.5, it’s for to be reasonably higher to cover -4.5. Again, I’m still learning and trying to figure things out. Thank you!
Yeah, Elo would think they have a pretty good chance to cover a -4.5 if the Elo line is -9.5. And that's borne out by other ratings as well -- I have a totally separate SRS measure that is denominated in PPG relative to average under the "SRS Rankings " tab here:
It considers the Packers a +3.4 and the Cardinals a -1.6, so you would expect Green Bay to be favored by something like 7.5 at home (if we assume HFA is ~2.5 PPG). Another one I like is ESPN's FPI:
It has GB at +3.5 and Arizona +0.2, implying a margin of Packers -5.5 or -6 at home. So the numbers certainly seem to think Green Bay is being undervalued by Vegas in that one.
All good, haha. Yes that's exactly right -- if a metric is denominated in points per game versus league average, then you can predict the scoring margin by looking at the difference between the ratings and adding some home-field value (usually 2-3 points) to the team who's at home. And one cool thing is that the errors of these predictions tend to follow a normal distribution:
Hello! Quick question to see if I’m understanding this right. ESPNBet has the 49ers -3.5 but your system has them -1.5. Would it be smart to take Seattle +3.5?
In theory, yes. In practice, Vegas is using stuff WAY more sophisticated than Elo. So it's an interesting data point to use among many, and I maintain it mostly to give me interesting narratives to write off of, but I would be careful using it to actually bet money on.
Glad I didn’t bet it, lol. But I’m glad to have found your site and I’m beginning to understand what I’m looking at. Your Elo ranking and whatnot will have a place in my system if/when I ever get one. Thanks!
First of all, I want to say that your work has been outstanding. While experimenting with ELO probabilities, I came across your predictions, and I wanted to ask for some clarification.
How do you calculate win probabilities for each match? Is it based on ELO rankings? I tried to calculate it using the following simple formula, taking Cowboys vs. Steelers as an example:
Visitors (Cowboys) = 1 / (1 + 10^((Local ELO 1571 - Visitor ELO 1554) / 400)) = 47.56%
Local (Steelers) = 1 / (1 + 10^((Visitor ELO 1554 - Local ELO 1571) / 400)) = 52.44%
As you can see, there is a significant difference compared to your win probability prediction, which is:
Cowboys 39%
Steelers 61%
This is a big huge difference, as I’ve been using the ELO method to beat sportsbooks, and while it has been successful so far, it has also been quite slow.
Amazing and appreciated that you take the time to reply! Reading the article I noticed that calculating the Winning Probabilities are no enough to beat the Sportsbooks, but as hobby I'm on my way for do it, for instance using ELO as data, kelly as risk manager and cooking some montecarlo testing.
It looks like one of the biggest discrepancies is the Jets, whom you have as a bottom 10 team and a double-digit underdog week 1, vs. Vegas has them as a top 10-12 team and only a 4.5 underdog week 1. Are ELO ratings not accounting for Rodgers being back, and/or lower on the defense? Maybe worth mentioning a lot of people have said they believe the Jets defense stopped playing at its highest level last year after they were eliminated from playoff contention and had the worst offense in the league week in and week out.
Yeah, that makes sense -- these have no QB adjustment, so it doesn't know Rodgers is coming back (although I will point back to something I wrote last year, when he was a year younger and hadn't yet been injured, about how few QBs were viable at such an age: https://neilpaine.substack.com/i/136821390/aaron-rodgers-age-on-dec)
Cool makes sense. And yeah I’m lower on current Rodgers than consensus as well, just wanted to check whether ELO was even lower than I, or just didn’t have a QB adjustment haha.
I've been wanting to get that QB adjustment going again, but it's a lot to stay on top of. I used to maintain the QB depth charts for every team when I was at 538, but that was before I started running this Substack and all of the other stuff. It would also be interesting to see how to implement such a system in NFLseedR... I've just started playing around with the basics of it, and it's very cool, but idk if there's a mechanism to tell it that a certain QB will be available by a particular week and adjust the team's rating that way.
hey neil. i had a question about your elo calculations. i am trying to recreate and am close but am off a little. here is the calculations for K which i have pulled from this link (including margin of victory) https://fivethirtyeight.com/features/introducing-nfl-elo-ratings/
K = (LN(ABS(PD)+1) * (2.2/((ELOW-ELOL)*.001+2.2))) * 20
then i am plugging K into this formula which I found here https://stanislav-stankovic.medium.com/elo-rating-system-6196cc59941e
R’a = Ra + K*(Sa — Ea)
Ea = Qa /(Qa + Qb)
Qa = 10^(Ra/c)
Qb = 10^(Rb/c), 0 ≤ Ea ≤ 1
Wondering if I may be doing this wrong? Or any things that might be missing?
Just wanted to follow up to let you know I’m 5-2 since I’m since incorporating your ELO and SRS rankings into my “system”. Should be 6-1 but my dumbass didn’t look at the injury reports and didn’t realize half the Saints offense was injured. Lost the other by .5 point lol. Thanks!
Nice! Glad it's working out!
Sorry to bother you again, but I have another quick question. Is the win probability to win the game outright, or is that to cover the spread that’s shown next to it as well? Thanks!
It's just straight-up win probability, not vs the spread. (Putting rounding aside, odds to cover the Elo line shown should be, by definition, 50% for each team anyway, since Elo differential is what drives both the predicted spread and the win probability.)
Gotcha. Thanks a not for taking time to reply. I know you said Vegas uses much more sophisticated stuff with their lines and whatnot, but in theory since the Elo spread has Packers -9.5, and the actual line is -4.5, that could signal a strong play at -4.5, correct? I’m assuming if it’s a 50% probably offer the Elo spread of -9.5, it’s for to be reasonably higher to cover -4.5. Again, I’m still learning and trying to figure things out. Thank you!
Yeah, Elo would think they have a pretty good chance to cover a -4.5 if the Elo line is -9.5. And that's borne out by other ratings as well -- I have a totally separate SRS measure that is denominated in PPG relative to average under the "SRS Rankings " tab here:
https://neilpaine.substack.com/i/148020098/nfl-elo-rankings-win-projections-and-team-ratings
It considers the Packers a +3.4 and the Cardinals a -1.6, so you would expect Green Bay to be favored by something like 7.5 at home (if we assume HFA is ~2.5 PPG). Another one I like is ESPN's FPI:
https://www.espn.com/nfl/fpi
It has GB at +3.5 and Arizona +0.2, implying a margin of Packers -5.5 or -6 at home. So the numbers certainly seem to think Green Bay is being undervalued by Vegas in that one.
Ok so don’t kill me, but I’m a total noob at this but I THINK I’m understanding but I want to make sure.
Using your SRS, do I take the difference in the two totals, add 2.5ish to the home team, and that’s the SRS predicted line?
ESPN PPI is the same thing? Again, thanks a ton
All good, haha. Yes that's exactly right -- if a metric is denominated in points per game versus league average, then you can predict the scoring margin by looking at the difference between the ratings and adding some home-field value (usually 2-3 points) to the team who's at home. And one cool thing is that the errors of these predictions tend to follow a normal distribution:
https://fivethirtyeight.com/features/how-unexpected-was-clevelands-shellacking-of-cincinnati/
Hello! Quick question to see if I’m understanding this right. ESPNBet has the 49ers -3.5 but your system has them -1.5. Would it be smart to take Seattle +3.5?
In theory, yes. In practice, Vegas is using stuff WAY more sophisticated than Elo. So it's an interesting data point to use among many, and I maintain it mostly to give me interesting narratives to write off of, but I would be careful using it to actually bet money on.
Glad I didn’t bet it, lol. But I’m glad to have found your site and I’m beginning to understand what I’m looking at. Your Elo ranking and whatnot will have a place in my system if/when I ever get one. Thanks!
Hi Neil,
First of all, I want to say that your work has been outstanding. While experimenting with ELO probabilities, I came across your predictions, and I wanted to ask for some clarification.
How do you calculate win probabilities for each match? Is it based on ELO rankings? I tried to calculate it using the following simple formula, taking Cowboys vs. Steelers as an example:
Visitors (Cowboys) = 1 / (1 + 10^((Local ELO 1571 - Visitor ELO 1554) / 400)) = 47.56%
Local (Steelers) = 1 / (1 + 10^((Visitor ELO 1554 - Local ELO 1571) / 400)) = 52.44%
As you can see, there is a significant difference compared to your win probability prediction, which is:
Cowboys 39%
Steelers 61%
This is a big huge difference, as I’ve been using the ELO method to beat sportsbooks, and while it has been successful so far, it has also been quite slow.
Regards
Thank you! There's also a 65-point Elo bonus for the home team: https://fivethirtyeight.com/features/introducing-nfl-elo-ratings/
Thank You Neil!
Amazing and appreciated that you take the time to reply! Reading the article I noticed that calculating the Winning Probabilities are no enough to beat the Sportsbooks, but as hobby I'm on my way for do it, for instance using ELO as data, kelly as risk manager and cooking some montecarlo testing.
It looks like one of the biggest discrepancies is the Jets, whom you have as a bottom 10 team and a double-digit underdog week 1, vs. Vegas has them as a top 10-12 team and only a 4.5 underdog week 1. Are ELO ratings not accounting for Rodgers being back, and/or lower on the defense? Maybe worth mentioning a lot of people have said they believe the Jets defense stopped playing at its highest level last year after they were eliminated from playoff contention and had the worst offense in the league week in and week out.
Yeah, that makes sense -- these have no QB adjustment, so it doesn't know Rodgers is coming back (although I will point back to something I wrote last year, when he was a year younger and hadn't yet been injured, about how few QBs were viable at such an age: https://neilpaine.substack.com/i/136821390/aaron-rodgers-age-on-dec)
Cool makes sense. And yeah I’m lower on current Rodgers than consensus as well, just wanted to check whether ELO was even lower than I, or just didn’t have a QB adjustment haha.
I've been wanting to get that QB adjustment going again, but it's a lot to stay on top of. I used to maintain the QB depth charts for every team when I was at 538, but that was before I started running this Substack and all of the other stuff. It would also be interesting to see how to implement such a system in NFLseedR... I've just started playing around with the basics of it, and it's very cool, but idk if there's a mechanism to tell it that a certain QB will be available by a particular week and adjust the team's rating that way.