Nate Silver: Pats 2003-2014 Best 12 Yr Stretch in NFL History

[Zeus]

I have no idea how this works or what it means yet but I expect to have some fun trying to figure it out.

Pats rated #1 with a 57% win probability vs. Buffalo this week. (see NFL predictions link below)

http://projects.fivethirtyeight.com/complete-history-of-the-nfl/#ne

http://projects.fivethirtyeight.com/2015-nfl-predictions/

http://fivethirtyeight.com/datalab/patriots-nfl-dynasty-greatest-of-all-time/

"After the New England Patriots’ Super Bowl win last season, we tried to calculate whether they had become the greatest NFL dynasty ever. The answer, as always, was that it depends on how you count. From 2003 to 2014, the Patriots’ average season-ending Elo rating was 1711, the best 12-season stretch in league history. But the best 13-season stretch — and 14-season and 15-season and on and on, to at least 25 seasons — belongs to the San Francisco 49ers."​

 

[Ice_Ice_Brady]

CHEATERS!!! 1849 is the highest score ever...does this account for...CHEATING???

 

[Rob0729]

 

[BradyFTW!]

and soon enough, 2001-2015 will be the best 15 year stretch in history.

 

[venecol]

538 shows NE on their week to week with a high probability of defeating everyone they face except Denver which they have at 50%/50%. But their simulations show an average 12-4 season. I should have stayed awake in that statistics class.

 

[FourierSeries]

As long as we're discussing prediction algorithms, Cortana (Windows virtual assistant) outperformed Nate Silver's predictions last week, going 13-3. It correctly predicted NE over Pittsburgh and the Rams upset over Seattle. Last season it had a 67% accuracy and predicted NE would win the SB back in December.

Cortana is picking Buffalo this week.

 

[Bruins29]

Did they recalculate the 49ers with the "remove the stickum" factor?

 

[PP2]

As long as we're discussing prediction algorithms, Cortana (Windows virtual assistant) outperformed Nate Silver's predictions last week, going 13-3. It correctly predicted NE over Pittsburgh and the Rams upset over Seattle. Last season it had a 67% accuracy and predicted NE would win the SB back in December.

Cortana is picking Buffalo this week.

So much for Cortana.

 

[FourierSeries]

So much for Cortana.

To be fair, it's giving Buffalo a 53% chance of winning. That's basically a "pick em." Vegas has the game as a pick em as well, I believe.

It has a pretty good track record, but you can't expect it to be right all the time.

 

[Concussion Protocol]

The Patriots* obviously got to 1849 intentionally just to mock the 49ers. Wouldn't expect anything less from those classless cheaters

 

[PP2]

To be fair, it's giving Buffalo a 53% chance of winning. That's basically a "pick em." Vegas has the game as a pick em as well, I believe.

It has a pretty good track record, but you can't expect it to be right all the time.

It probably is as good an algorithm program as they come, but there are some things an algorithm cannot measure or value, such as how angry Brady is, as well as the effect that this is having on the team right now:

 

[FourierSeries]

True. This will probably be one of the times it gets one wrong.

 

[Soul_Survivor88]

I have no idea how this works or what it means yet but I expect to have some fun trying to figure it out.

Pats rated #1 with a 57% win probability vs. Buffalo this week. (see NFL predictions link below)

http://projects.fivethirtyeight.com/complete-history-of-the-nfl/#ne

http://projects.fivethirtyeight.com/2015-nfl-predictions/

http://fivethirtyeight.com/datalab/patriots-nfl-dynasty-greatest-of-all-time/

"After the New England Patriots’ Super Bowl win last season, we tried to calculate whether they had become the greatest NFL dynasty ever. The answer, as always, was that it depends on how you count. From 2003 to 2014, the Patriots’ average season-ending Elo rating was 1711, the best 12-season stretch in league history. But the best 13-season stretch — and 14-season and 15-season and on and on, to at least 25 seasons — belongs to the San Francisco 49ers."​

I'm glad Nate Silver mentioned the longevity of San Francisco, because this is another example of the kind of double standard that's been imposed on New England There are so many all-time great teams and players who've been proven to have cheated. People ought to remember that the 49ers cheated the salary cap during the 90s, and were punished for doing so. But despite these infractions, no one has ever tried to discredit what Steve Young or the Niners have achieved. Brady and Belichick ought to be treated the same way.

 

[PatsFanSince74]

538 shows NE on their week to week with a high probability of defeating everyone they face except Denver which they have at 50%/50%. But their simulations show an average 12-4 season. I should have stayed awake in that statistics class.

It probably has something to do with what someone smarter than me once called the "Flaw of Averages." In this case, just because your model says that, "On average, you win every week," it doesn't mean that "You win every week."

Nate is pretty transparent with his methodology, which is why he (and Sam Wang at Princeton) pretty much revolutionized the traditional method of predicting elections and exposed the weaknesses of standard polling (Nate has done a better job of marketing himself, but Sam might be smarter and better). So, I'm sure Nate explains his methodology somewhere and I just don't have the time to track it down this morning.

But, if I had to venture a guess in this case, it would be that the Standard Deviations around the mean outcomes of games more than two or three weeks out in the future are pretty large. The result would be that a few hundred thousand simulations of potential season W/L records for, for example, the Patriots most likely yield a wide range of possible records, the mean of which has 11.6 wins after week 1. The result will change week by week.

 

[zydecochris]

538 shows NE on their week to week with a high probability of defeating everyone they face except Denver which they have at 50%/50%. But their simulations show an average 12-4 season. I should have stayed awake in that statistics class.

It probably has something to do with what someone smarter than me once called the "Flaw of Averages." In this case, just because your model says that, "On average, you win every week," it doesn't mean that "You win every week."

Nate is pretty transparent with his methodology, which is why he (and Sam Wang at Princeton) pretty much revolutionized the traditional method of predicting elections and exposed the weaknesses of standard polling (Nate has done a better job of marketing himself, but Sam might be smarter and better). So, I'm sure Nate explains his methodology somewhere and I just don't have the time to track it down this morning.

But, if I had to venture a guess in this case, it would be that the Standard Deviations around the mean outcomes of games more than two or three weeks out in the future are pretty large. The result would be that a few hundred thousand simulations of potential season W/L records for, for example, the Patriots most likely yield a wide range of possible records, the mean of which has 11.6 wins after week 1. The result will change week by week.

Well said, PatsFanSince74. As an example, just to simplify things a bit for illustration, suppose that the Patriots had a 75% win percentage every single week (starting from the beginning of the season). It seems like "75%, wow, we will win every game!" But that just means that over a large sample size, the Pats would win 3 out of every 4 games. Thus, the most likely result would be a 12-4 season.

However, one shouldn't conclude: "Great! 12-4! I will drive to Las Vegas and wager all of my life savings on the Pats finishing with a 12-4 record!". Why not? For one thing, the prediction model has to be far from perfect, it can't possibly take into account all relevant factors (especially injuries). And, as you pointed out, one must have a statistically significant sample size, and 16 games definitely doesn't qualify. Thus, 12-4 is the most likely record (with a 75% win probability for each game), but it could easily be some other result. The small sample size (16 games) is why statistical analysis is arguably somewhat less useful with football than with with baseball (162 games per regular season, a much larger sample size), where they keep all sorts of sophisticated stats. Although basketball has only 82 games, its sample size is still five times that of football, and it has recently become much more stats driven also.

 

[brownj12]

538 shows NE on their week to week with a high probability of defeating everyone they face except Denver which they have at 50%/50%. But their simulations show an average 12-4 season. I should have stayed awake in that statistics class.

It probably has something to do with what someone smarter than me once called the "Flaw of Averages." In this case, just because your model says that, "On average, you win every week," it doesn't mean that "You win every week."
.

Nothing complicated about it at all. Imagine you have a bag with 3 balls in it, 2 blue balls and 1 red ball. Every week for 16 weeks you pick out one ball. Every week the odds say you are most likely going to pick out a blue ball, and each week it would be a good bet that you will pick out a blue ball. However you actual odds of picking 16 blue balls in 16 weeks is about 1 in 656. If you pick balls 16 times over and over again, on average you will get 10.67 blue balls per set. Occaisionally you will get 16, even more rarely you will get 0. usually you will end up with 10 or 11. But over enough sets you will end up with an average of 10.67.

 

[PatsFanSince74]

Well said, PatsFanSince74. As an example, just to simplify things a bit for illustration, suppose that the Patriots had a 75% win percentage every single week (starting from the beginning of the season). It seems like "75%, wow, we will win every game!" But that just means that over a large sample size, the Pats would win 3 out of every 4 games. Thus, the most likely result would be a 12-4 season.

However, one shouldn't conclude: "Great! 12-4! I will drive to Las Vegas and wager all of my life savings on the Pats finishing with a 12-4 record!". Why not? For one thing, the prediction model has to be far from perfect, it can't possibly take into account all relevant factors (especially injuries). And, as you pointed out, one must have a statistically significant sample size, and 16 games definitely doesn't qualify. Thus, 12-4 is the most likely record (with a 75% win probability for each game), but it could easily be some other result. The small sample size (16 games) is why statistical analysis is arguably somewhat less useful with football than with with baseball (162 games per regular season, a much larger sample size), where they keep all sorts of sophisticated stats. Although basketball has only 82 games, it has recently become much more stats driven also.

Absolutely. The sample size is so ridiculously small that forecasting is a crap shoot.

My guess is that Nate's model shows a distribution of possible W/L outcomes for the Pats season that ranges from 7--9 or 8--8 to 15--1 or 16--0 out to the fourth or fifth Standard Deviation. The 11.6 wins is the mean of the distribution with a standard deviation of some sort, probably 1.5--2.0 games.

So, in the second week, if the Standard Deviation is 1.5 games, Nate's model would be saying that there is ~67% confidence that the pats will win between 10.1 (10) and 13.1 (13) games and ~99% confidence that the pats will win between 8.6 (9) and 14.6 (15) games, etc. In other words, "Interesting" but not "Helpful."

 

[PatsFanSince74]

Nothing complicated about it at all. Imagine you have a bag with 3 balls in it, 2 blue balls and 1 red ball. Every week for 16 weeks you pick out one ball. Every week the odds say you are most likely going to pick out a blue ball, and each week it would be a good bet that you will pick out a blue ball. However you actual odds of picking 16 blue balls in 16 weeks is about 1 in 656. If you pick balls 16 times over and over again, on average you will get 10.67 blue balls per set. Occaisionally you will get 16, even more rarely you will get 0. usually you will end up with 10 or 11. But over enough sets you will end up with an average of 10.67.

No disagreement on the statistics, but, of course, the complicating factor is that the number of blue and red balls for each week are different and that difference will vary after each week as "on field" factors vary.

 

[FourierSeries]

I think this is one Cortana prediction that we'd all like to see come to fruition. Nate Silver agrees as well. I would not be surprised at all if it happens.

 

[zydecochris]

As long as we're discussing prediction algorithms, Cortana (Windows virtual assistant) outperformed Nate Silver's predictions last week, going 13-3. It correctly predicted NE over Pittsburgh and the Rams upset over Seattle. Last season it had a 67% accuracy and predicted NE would win the SB back in December.

Cortana is picking Buffalo this week.

This discussion of football prediction algorithms reminds me of that wonderful year in 2007 when Bill Simmons was in a contest with his wife in predicting football winners versus the line. He didn't really have a statistical approach, but you could tell from his columns and his podcasts that he was devoting a GIGANTIC amount of time and energy in trying to get his predictions right. His wife, had a little "Sportsgal" column where she would make her picks, and her reasoning was always something completely non-football related, something like "I don't like the name of the dolphins, it is too cuddly, I'll pick against them" or something about their uniform color. Of course, she absolutely pulverized Bill Simmons in the contest, showing that a little knowledge can be a dangerous thing when it comes to betting (or the stock market).

http://sports.espn.go.com/espn/page2/story?page=simmons/071228