Week 10 Ratings
Julio
As usual, a 100 point difference means the higher rated
teams wins 62.5% of the time; 200 pts: 75%; 300 pts: 85%;
400 pts: 95%.
Rankings for week 10 2010-11 Season
Average rating is 1598.126
Rank Team Rating
1 Jets 2025
2 Atlanta 1938
3 New England 1842
4 Indianapolis 1792
5 Baltimore 1777
6 Pittsburgh 1740
7 Philadelphia 1735
8 Tampa Bay 1697
9 Miami 1693
10 Chicago 1684
11 Green Bay 1679
12 New Orleans 1673
13 Cleveland 1656
14 Tennessee 1631
15 Oakland 1630
16 Seattle 1606
17 Giants 1580
18 Jacksonville 1569
19 San Diego 1566
20 Denver 1565
21 Kansas City 1515
22 Washington 1501
23 Minnesota 1499
24 San Francisco 1462
25 Houston 1456
26 Cincinatti 1454
27 St. Louis 1446
28 Arizona 1434
29 Dallas 1419
30 Detroit 1321
31 Carolina 1295
32 Buffalo 1260
eric
I don't buy it. The Jets don't beat Green Bay 90% of the time. Heck
they go into OT vs. Detroit.
Dano
"eric" <wart...@gmail.com> wrote in message
news:e217ed0f-0550-4fe6...@r29g2000yqj.googlegroups.com...
Yeah. So the Pats actually LOST to Cleveland right? And the Falcons barely
eked out a win over the 49'ers.
Any given Sunday...
eric
> "eric" <warth...@gmail.com> wrote in message
Exactly my point. The 90% for 400 points is way too high.
Grinch
> teams wins 62.5% of the time; 200 pts: 75%; 300 pts: 85%;
> 400 pts: 95%.
Those numbers look like the Elo Chess rating system. It's now used to
produce the rankings for golf pros, tennis pros, all kinds of
competitors, even universities competing for students, so that's OK.
But Sagarin produces an Elo ranking for the NFL and it has the Jets
5th.
And the near 800-pt gap from top to bottom is **way** too large, the
bottom teams have zero chance against the top ones. (Literally, they
have the same chance that I would've had playing chess against Bobby
Fischer at his best). That just ain't the NFL.
E.g ., these rankings give the Jets a 90% chance against *above*
average teams like New Orleans and Green Bay. That's not credible.
(What happened against Green Bay?)
Elo ratings become accurate only after about 30 games, IIRC. Maybe
that's the problem, small sample size. That's a problem for all
football rating systems based on W-L as Elo is, because there are so
few games in the season.
That's why most football rating systems are based on plays (yards per
play, success rate) of which there are thousands in a season, or
Pythagorean pts for-against ratios (Sagarin's "Pure Points").
Anyhow, you should do something to adjust your top-bottom spread to
reasonable proportions.
Julio
Grinch <oldna...@mindspring.com> wrote:
>
> On Nov 16, 9:46=A0pm, Julio <hoolio3sanc...@yahoo.com> wrote:
> > As usual, a 100 point difference means the higher rated
> > teams wins 62.5% of the time; 200 pts: 75%; 300 pts: 85%;
> > 400 pts: 95%.
>
> Those numbers look like the Elo Chess rating system. It's now used to
> produce the rankings for golf pros, tennis pros, all kinds of
> competitors, even universities competing for students, so that's OK.
>
> But Sagarin produces an Elo ranking for the NFL and it has the Jets
> 5th.
I've pointed out in earlier seasons that this rating system is the
Elo system. If Sagarin is using an Elo system, it's not based on
wins and losses, because then his rankings would be the same as mine.
>
> And the near 800-pt gap from top to bottom is **way** too large, the
> bottom teams have zero chance against the top ones. (Literally, they
> have the same chance that I would've had playing chess against Bobby
> Fischer at his best). That just ain't the NFL.
I feel confident that if the Jets tooks on Detroit for all the marbles
on a neutral field *today*, they would smoke Detroit.
>
> E.g ., these rankings give the Jets a 90% chance against *above*
> average teams like New Orleans and Green Bay. That's not credible.
> (What happened against Green Bay?)
No rating system is perfect. The Jets haven't lost
to any bad teams, until they do, their rating may be artificially
high. The rating system can't tell if a team is getting
better, whether because they are just getting more practice
or because they are adjusting to injuries (like GB), or getting
worse because of injuries, or suddenly turning it on because of
a new coach, like this week with Dallas. For example, no rating
system can predict when a good team is going to have a bad week,
just because they say collectively "Ah f--- it" and they almost
all do, that's why nobody goes goes 16-0 (well, almost nobody....),
but that bad week doesn't really have anything to do with
their overall ability or desire to win.
The point of this rating system is so you can see at a glance who
might be underrated or overrated based on their record alone.
The chance of winning is an estimation that gets better with
more games played, it's not a guarantee.
This past week, the week 9 ratings picked the winner correctly
8 times out of 14, and of the 6 incorrect (numbers are rating
differential from week 9 ratings):
miami-tennesee 49
buff-detroit 278
denver kc 64
sf-stL 130
dal giants 358
ne-pitt 16
only two had rating differentials large enough to make the odds
greater than about 50/50. One of those was Dallas, with the new
coach, and the other was Buffalo, which was obviously a better
team than its record would suggest and bound to win against
someone sometime despite its lowest rating of anyone up to that
point.
J.
Grinch
> In article <4b24e66f-5140-42d4-b5d6-ea0aa7183...@a37g2000yqi.googlegroups.com>
>
> Grinch <oldna...@mindspring.com> wrote:
>
> > On Nov 16, 9:46=A0pm, Julio <hoolio3sanc...@yahoo.com> wrote:
> > > As usual, a 100 point difference means the higher rated
> > > teams wins 62.5% of the time; 200 pts: 75%; 300 pts: 85%;
> > > 400 pts: 95%.
>
> > Those numbers look like the Elo Chess rating system. It's now used to
> > produce the rankings for golf pros, tennis pros, all kinds of
> > competitors, even universities competing for students, so that's OK.
>
> > But Sagarin produces an Elo ranking for the NFL and it has the Jets
> > 5th.
>
> I've pointed out in earlier seasons that this rating system is the
> Elo system.
But Dude, you're not following the rules of the Elo system.
It uses binary W/L outcomes so it needs a significant number of them
to come up with a meaningful rating.
E.g, the USCF requires 26 games for an Elo rating. With 2 or 4 or 9
the margin of error is *immense*, which one can see in the huge
distances you've got between teams here.
You've got top teams with a 90% chance against the middle teams, which
is not plausible. Your top-bottom gap also exceeds 100% winning
probability (736 points Elo) which is not plausible either.
> If Sagarin is using an Elo system, it's not based on
> wins and losses, because then his rankings would be the same as mine.
>
> > And the near 800-pt gap from top to bottom is **way** too large, the
> > bottom teams have zero chance against the top ones. (Literally, they
> > have the same chance that I would've had playing chess against Bobby
> > Fischer at his best). That just ain't the NFL.
>
> I feel confident that if the Jets tooks on Detroit for all the marbles
> on a neutral field *today*, they would smoke Detroit.
I'd hope you'd think that, since a 634 pt rating difference is a 99%
winning probability. But do you really believe any NFL team would
beat any other 99 out of 100?
And are you equally confident that the Jets would have a 79% chance
against the Colts, 89% against the Packers (who already beat the
Jets), and 89% against the Saints?
> > E.g ., these rankings give the Jets a 90% chance against *above*
> > average teams like New Orleans and Green Bay. That's not credible.
> > (What happened against Green Bay?)
>
> No rating system is perfect. The Jets haven't lost
> to any bad teams, until they do, their rating may be artificially
> high.
Exactly so. In other words, "not enough games rated".
Which is why the 26 requirement.
> The rating system can't tell if a team is getting
> better, whether because they are just getting more practice
> or because they are adjusting to injuries (like GB), or getting
> worse because of injuries, or suddenly turning it on because of
> a new coach, like this week with Dallas. For example, no rating
> system can predict when a good team is going to have a bad week,
> just because they say collectively "Ah f--- it" and they almost
> all do, that's why nobody goes goes 16-0 (well, almost nobody....),
> but that bad week doesn't really have anything to do with
> their overall ability or desire to win.
All of which reduces to: "not enough games rated for a reliable
result".
> The point of this rating system is so you can see at a glance who
> might be underrated or overrated based on their record alone.
And by *how much* -- which are the numbers that are not plausible. No
NFL team is anyhting like a 90+% favorite against a *middle* of the
league team.
The fact that the Jets were shut out at home by the Packers doesn't
necessarily mean they shouldn't be favored in a rematch -- but it does
probably means they shouldn't be a 9 to 1 favorite! :-)
> The chance of winning is an estimation that gets better with
> more games played, it's not a guarantee.
All such estimates have defined margins of error. Calculate yours at
9 games, it should be included in the table.
> This past week, the week 9 ratings picked the winner correctly
> 8 times out of 14, and of the 6 incorrect (numbers are rating
> differential from week 9 ratings):
>
> miami-tennesee 49
> buff-detroit 278
> denver kc 64
> sf-stL 130
> dal giants 358
> ne-pitt 16
>
> only two had rating differentials large enough to make the odds
> greater than about 50/50.
More precisely, the favorites by your week 9 rankings had an average
71% chance of winning, which would have been 10-4.
They actually won 57%, 8-6, one over .500.
I'm not trying to beat on you, I've used the Elo system for several
things myself since I was a chess player figuring out my own ratings
under it, more years ago than I want to remember.
But every rating system has rules, and under Elo rules nine game
results just isn't enough, in fact 16 isn't enough. Which is why, as
I said before, football team rating systems generally use play results
(which exist in the thousands) or Pythagorean points for-against.
> One of those was Dallas, with the new
> coach, and the other was Buffalo, which was obviously a better
> team than its record would suggest and bound to win against
> someone sometime despite its lowest rating of anyone up to that
> point.
In other words: Buffalo hadn't played enough games for its rating to
be accurate. Yes, exactly so.
>
> J.
Julio
Grinch <oldna...@mindspring.com> wrote:
>
> On Nov 18, 11:24=A0pm, Julio <hoolio3sanc...@yahoo.com> wrote:
> > In article <4b24e66f-5140-42d4-b5d6-ea0aa7183...@a37g2000yqi.googlegroups=
> .com>
> >
> > Grinch <oldna...@mindspring.com> wrote:
> >
> > > On Nov 16, 9:46=3DA0pm, Julio <hoolio3sanc...@yahoo.com> wrote:
> > > > As usual, a 100 point difference means the higher rated
> > > > teams wins 62.5% of the time; 200 pts: 75%; 300 pts: 85%;
> > > > 400 pts: 95%.
> >
> o
> > > produce the rankings for golf pros, tennis pros, all kinds of
> > > competitors, even universities competing for students, so that's OK.
> >
> > > But Sagarin produces an Elo ranking for the NFL and it has the Jets
> > > 5th.
> >
> > I've pointed out in earlier seasons that this rating system is the
> > Elo system.
>
> But Dude, you're not following the rules of the Elo system.
>
> It uses binary W/L outcomes so it needs a significant number of them
> to come up with a meaningful rating.
>
> E.g, the USCF requires 26 games for an Elo rating. With 2 or 4 or 9
> the margin of error is *immense*, which one can see in the huge
> distances you've got between teams here.
That's not true. You get a provisional rating below 100 games. Nobody
starts out playing 26 games at once. These days 4 game tournaments are strung
out over a month and you get a rating for each game you play.
No rating system can be perfectly accurate only over 16 games. Rating
systems that rate plays can't predict a given Sunday either. This rating
system does what I want it to do, which is show at a glance a relative
rating of the team. Whether a difference of 200 points means that the
higher rated team wins *exactly* 85 percent of the time is irrelevant,
it's just an approximation, as are all rating systems. I don't believe
that Detroit would loss 99 games out of 100 to the Jets, but I don't
believe a 2200 rated chess master would lose 99 times out of a hundred
to a 2700 grandmaster either. That is why Las Vegas likes the NFL!
J.
Julio
A few months ago Magnus Carlsen played in the Chess Olympiad.
At the time he was the highest rated player in the world at 2826.
But his performance rating for the tournament was 2664, and he
lost 3 games to players rated 2710, 2728 and 2627. The chance
that a player of that rating (with over 1000 rated games!) would
lose to 3 players in a single tournament is too low to calculate. Yet
it happened. Then later at the the Nanjing tournament he beat
Topalov, rated 2803, twice, and had a 2901 performance rating
for that tournament. Any given day happens in chess too, even
with many rated games.
J.
Grinch
> Just to give you some perspective:
> A few months ago Magnus Carlsen played in the Chess Olympiad.
> At the time he was the highest rated player in the world at 2826.
> But his performance rating for the tournament was 2664, and he
> lost 3 games to players rated 2710, 2728 and 2627. The chance
> that a player of that rating (with over 1000 rated games!) would
> lose to 3 players in a single tournament is too low to calculate.
Really? Ya' think?
The chance that he would lose to just these three players in a row, as
per the official FIDE rating tables:
2826 vs
2710 ... 34% x
2728 ... 37% x
2627 ... 24% = 3%
That's too low for you to calculate?
Of course that's *3 of 3 in a row*. As for 3 out of the typical 15 "in
a single tournament", that's very different.
Say the opposition had an average rating of the three guys above,
2688. The 2862 player would be expected to lose 31%, or 4.5 games.
Or in 8 games played at the Olympiad against 2688 average competition
he'd be expected to lose 2.5, not so very far from 3!
What's "too low to calculate" for you: 31%, 4.5, or 2.5?
I'm beginning to sense your problem: You are innumerate, and wish to
flaunt it to the world as proof of your mastery of things numerical.
Which makes you very foolish too.
I hope you are a young kid so you can learn from this.
>Yet it happened. Then later at the the Nanjing tournament he beat
> Topalov, rated 2803, twice,
He beat a guy who's not as good as him! Twice!!
Holy Cow. Anything can happen in this world.
> and had a 2901 performance rating for that tournament.
Wow. He had a tourney performance rating all of 76 points higher than
the rating he came in with.
What a miracle! Or ... maybe not?
What are the odds of such a thing happening? The precise odds.
You have no idea, right?
Grinch
> In article <836fe80d-694f-422c-9d39-6ff3f21c3...@j25g2000yqa.googlegroups.com>
>
> Grinch <oldna...@mindspring.com> wrote:
>
> > On Nov 18, 11:24=A0pm, Julio <hoolio3sanc...@yahoo.com> wrote:
> > > In article <4b24e66f-5140-42d4-b5d6-ea0aa7183...@a37g2000yqi.googlegroups=
> > .com>
>
> > > Grinch <oldna...@mindspring.com> wrote:
>
> > > > On Nov 16, 9:46=3DA0pm, Julio <hoolio3sanc...@yahoo.com> wrote:
> > > > > As usual, a 100 point difference means the higher rated
> > > > > teams wins 62.5% of the time; 200 pts: 75%; 300 pts: 85%;
> > > > > 400 pts: 95%.
>
> > > > Those numbers look like the Elo Chess rating system. =A0It's now used t=
> > o
> > > > produce the rankings for golf pros, tennis pros, all kinds of
> > > > competitors, even universities competing for students, so that's OK.
>
> > > > But Sagarin produces an Elo ranking for the NFL and it has the Jets
> > > > 5th.
>
> > > I've pointed out in earlier seasons that this rating system is the
> > > Elo system.
>
> > But Dude, you're not following the rules of the Elo system.
>
> > It uses binary W/L outcomes so it needs a significant number of them
> > to come up with a meaningful rating.
>
> > E.g, the USCF requires 26 games for an Elo rating. With 2 or 4 or 9
> > the margin of error is *immense*, which one can see in the huge
> > distances you've got between teams here.
>
> That's not true. You get a provisional rating below 100 games.
100 games? Say "26 games".
And do you think you are helping your case by saying the USCF requires
100 games for an Elo rating but you can do it in 9? :-)
A "provisional rating" you can get with *zero* games. It's just an
estimate they slap on you. The error margin is, you know, *huge*.
That's why players with them aren't eligible for the many valuable
prizes and goodies of the chess world.
> Nobody
> starts out playing 26 games at once. These days 4 game tournaments are strung
> out over a month and you get a rating for each game you play.
> No rating system can be perfectly accurate only over 16 games.
Some, like the Elo, don't pretend to be anywhere near accurate under
26 -- certainly not at 9!
> Rating
> systems that rate plays can't predict a given Sunday either. This rating
> system does what I want it to do, which is show at a glance a relative
> rating of the team. Whether a difference of 200 points means that the
> higher rated team wins *exactly* 85 percent of the time is irrelevant,
> it's just an approximation, as are all rating systems.
We are really defining *exactly* and "approximation" down, it seems.
When your system pegs the Jets as 9-1 favorites over a 6-3 Green Bay
team that already shut them out at home, you are being neither "exact"
nor "approximate", eh? And we go down the list seeing implausibly huge
gaps like this between one team and another all the way down.
A rating system has two purposes: To rank teams ordinally, 1,2,3,...
and to give a realistic estimate of the strength gap between them, the
probability of one defeating the other.
When the error margin is so *huge* on the strength gaps, obviously
that's failure one -- and error gaps that huge overlap so the ordinal
ranking is wrong too. Failure 2 of 2.
I asked you to calculate the error margin for Elo at 9 games. You
haven't, and from your inability to even calculate simple win
probabilities off the Elo table it seems you can't. Here's a hint:
Two competitors of exactly equal strength play two games against each
other. The odds are 50/50 that one or the other will win both, and
50/50 that the two games will split.
You Elo rate the games. In the cases where one won both you come out
with an 800-point difference, over a 100% chance of winning in the
future. If you rate all four cases, the "average error" is 400 points
after 2 games -- a 92% chance of winning/losing against an exactly
equal opponent. Neither exact nor approximate!
This average error reduces only gradually with more games -- thus the
need for 26. More than 2, and more than 9 (but not 100).
But let me not be a pure naysayer, I strive to be helpful:
There is a better way to achieve your proclaimed objective, using only
W-L and with no problem of not enough game: Figure the stength of
schedule adjusted winning percentage of each team.
There's a formula for determining the win % strength of a team against
an average schedule from its % against an unbalanced one. Around .500
you just subtract the difference from it -- e.g., a .600 team playing
opponents with a combined .450 record has an adjusted .550 record. But
as scores get away from .500 you have to use the formula. Figure for
each team then iterate until the numbers stabilize as doing an Elo.
The result is exactly what you want. It ranks the teams on W-L by
strength of schedule as you are trying to use Elo to do ... but it is
fully accurate on the scale of strength gaps and so is ordinally much
more accurate ... and the numbers are expressed in terms everyone
understands, W-L % v adjW-L %, people can see the differences with
their eyeballs -- nobody has to try to figure out what 1794 vs 1467
means.
And I've never seen anyone do it for the NFL, so if you look up the
formula you can claim the system as all your own. My gift to you.
Every rating system has its purpose. The Elo system was designed for a
situation very different than the NFL -- rating players who compete
against huge numbers of strangers *without* W-L %es that can be
adjusted, as in chess, golf, bowling, tennis, video games, Scrabble
uses it, etc.
If the chess world consisted of 32 players who played nobody but each
other, Aprad would never have bothered.
>I don't believe that Detroit would loss 99 games out of 100 to the Jets,
You're right -- nothing like it.
>but I don't
> believe a 2200 rated chess master would lose 99 times out of a hundred
> to a 2700 grandmaster either.
Right again -- 96 times.
>That is why Las Vegas likes the NFL!
> J.- Hide quoted text -
>
> - Show quoted text -
Julio
Grinch <oldn...@mindspring.com> wrote:
> You Elo rate the games. In the cases where one won both you come out
> with an 800-point difference, over a 100% chance of winning in the
> future. If you rate all four cases, the "average error" is 400 points
> after 2 games -- a 92% chance of winning/losing against an exactly
> equal opponent. Neither exact nor approximate!
No idea where you get that. There is no 800 point difference on
winning two games, each game gets a rating and those ratings are
factored into the average rating of the winning team. You beat
someone rated 1500, your game rating is 1900, you do it again, your
game rating is 1900 and your overall rating is 1900 if you've only
played two games, some other number if you've played more. If your
rating after 5 games was 1700, and you beat two 1500 teams, then
your rating would be ((5*1700) + (2*1900))/7.
No rating system can accurately rate teams with only 16 games and
where no team plays every other team. Of course my rating system is
an approximation. So is the football outsiders rating system, which
couldn't predict GB beating the Jets either. And I don't think any rating
system in the world could have predicted they would be shut out at
home after a bye (a bizarre result in itself regardless of the relative
rating of the two teams).
I've already stated in previous arguments in previous years that you
could get a similar ranking by multiplying the teams winning percentage
by the winning percentage of its opponents. But my system, particularly
later in the season, is good at predicting winners. The week 10 rankings
picked 12, maybe 13 of this weeks 16 games correctly, pending Monday night's
game. There is no reason at all for you to pay attention to it, it's just
there for anyone who is interested.
J.