So now I know smart Harvard is...they have an interesting formula for calculating who will be in the Super Bowl. I'd like to see if they are right at the end of the season. Good article man!!

 

Leaving aside the merit (or lack thereof) of using approximate value (AV) to estimate a player's value (ideally, to win %), the reason the Dolphins came in 3rd is probably due to this quote:


"I aggregated each team's per game approximate value of what I considered to be the "core" makeup of an NFL team: QB, RB, 2 WR, TE, Top 2 OL, the Top-4 "Front Seven" defensive players, and the Top-2 players from the secondary."

..

Notice the "Top 2 OL" in there? Yeah, when you only look at the top two on our OL, our OL looks GREAT!! Wish we could play with an OL whose strength was best estimated by our top two players on the OL.

Anyway, you get the idea.. they are only ranking teams based on how balanced the distribution of "top players" are on the team. That being said, I do hope their predictions are accurate.

 

Coaching totally excluded.

 

Check out how high their formulas have the Jets as well. The problem, along with what cbrad pointed out, is that it doesn't take into account what a big deal having a horrible QB can be. Unlike baseball, where you could argue that all eight position starters are of fairly equal importance, that's obviously not the case in the NFL. I terrible QB makes an average team a bad one, and a great QB gets an average team into the playoffs.

The system also doesn't take into account that, even if a player is very good, older players are more likely to spend time injured than younger ones. A team loaded with vets, especially ones who have suffered serious injuries in the past, is likely to lose important players as they go along.

The whole thing is an interesting idea, but it doesn't see things that people would.

 

This is partly true, partly not. It tends to weight QB's more because the weights were chosen to "look right". So, if you look at AV for all the Patriots players on that nearly perfect 2007 team that went undefeated except for the (THANK YOU GIANTS!!) Super Bowl, you'll see this:

Brady = 26
Moss = 21
Light = 21
Welker = 17
Vrabel = 15
Wilfork = 13
Mankins = 13
etc..

So you do get the extra weight on the QB because they want the stats to come out "right". However.. the weights are entirely subjective (I mean they really are). And you'd have a hard time arguing that the resulting AV's actually represent the relative importance of the QB to the final outcome. I mean, was Brady 26/21 times more important than Moss to the Patriots 2007 season? Very likely no. So in that sense you're right: AV probably doesn't capture the relative importance of the QB in the modern NFL.

 

Link to the study:

http://harvardsportsanalysis.org/2015/07/a-way-too-early-prediction-of-the-nfl-season/

 

Well if that first list is anything to go by, then the AFC East will be the best division in football and it's not even close.

All four teams are in the top 10

 

I do think three of the four teams should be among the top 10 defenses and the one that isn't is the defending champ.

The formula was reportedly predictive from the previous year so that gives it some credibility.

 

As has been the case for decades...it's been close at times though.

 

I'm guessing you're referring to this statement from the study:

"To make sure this was a sound method, I tested it out on last year's data and ran a regression to see if AV was predictive of the end-of-regular season ELO ratings as reported by FiveThirtyEight. Aggregated AV was indeed significant with a T-stat of 8.57. It was also a strong predictor of ELO, as the regression returned a .72 R-Squared value."

..

That quote makes it sound like he used 2013 AV ratings to predict 2014 end-of-year ELO rankings (there's nothing wrong with ELO rankings btw.. it's probably the best ranking method out there), BUT I am pretty sure he did not do that. I'm almost 100% positive he used 2014 AV rankings to predict 2014 end-of-year ELO rankings.

Why?

Because if you can predict end-of-the-year ELO rankings with 72% probability you can make a killing in sports betting. ELO rankings tells you the probability of any team beating any other team (thus, it's not just rank order.. ELO rankings includes information to calculate probability of winning). If you can get that right 72% of the time (or be correlated that high), that's more than the % correct you need in sports betting to consistently win (often, just getting 55% correct may be enough.. certainly 60%+ is a no brainer).

So no, I don't think that quote means the formula was predictive of the previous year (the quote can be interpreted both ways).

 

Predicting NFL win totals at 72% does not mean you can make money gambling. You are ignoring the fact that Vegas predicts at around the same rate. Question is whether you can predict at a significantly higher rate than sportsbooks.

Also consider that when it comes to futures bets, there is a huge opportunity cost - your money is being held by the books for 6 months, which means you aren't able to generate profit from it. Additionally, most sportsbooks have pretty low limits on futures bets.

Sent from my SAMSUNG-SM-G920A using Tapatalk

 

Pretty fu##in cool nerds if you ask me.. I like it, we have a lot of talent

 

You're not predicting win totals. You're predicting probability of A beating B for any two A and B that are ranked. ELO is capable of that. If this were just predicting win totals that's another story. Saying you're predicting ELO rankings implies you are predicting probability of winning for any two teams that play each other.

And when you have that information, there is no opportunity cost. You bet the day of the match based on the moneylines/odds and get your winnings immediately after the game.

 

Even then my point still stands. Sportsbooks are predicting the winner of the game ~70% of the time.

Sent from my SAMSUNG-SM-G920A using Tapatalk

 

No you're not getting it. The probability of winning directly translates into a moneyline (let's just assume the bookie keeps nothing for the moment). If you can accurately predict the actual probability of winning, then you can compare this to the moneyline and bet accordingly (or not bet at all). If the moneyline is too small you bet the other way, etc..

Adding in the effect of the bookie having to make money is where you get that 55% I was talking about. That's not 55% you predict the winner correctly. That's 55% of the time you correctly predict the probability of winning. Whether you actually bet or not depends of course on the moneyline itself. So sportsbooks could predict the winner of the game correctly 90% of the time and it wouldn't matter.. it's the moneylines = probability of winning that matters.

 

A system based off elo ratings will not win money in the long term, despite it's ability to predict the winner 70% of the time, or picking the right probability 99% of the time. If it predicted better than the sportbooks, the sportsbooks would simply use that to set their moneylines.

By definition, a system such as the ELO system will not beat a market-based moneyline. Sportsbooks will always beat systems such as ELO.

Predicting the probability 55% of the time doesn't matter if the sportsbooks predicts it right 55% of the time as well.

Sent from my SAMSUNG-SM-G920A using Tapatalk

 

Two things:

1) Again, it's not about predicting the winner 70% of the time.. it's about predicting the probability of winning 70% of the time.

2) My whole argument was exactly what you were trying to say in this last post: what they claim cannot be true because IF it was true and IF it was novel, you'd make a killing in betting. IF it's true, and if it's NOT novel, then bookies would use it as you point out. Of course, they're claiming they have a novel method.

 

No, it can be true. Thats my point. They said they can predict the winner of the game 70% of the time.

Sent from my SAMSUNG-SM-G920A using Tapatalk

 

No, read the quote again. They said the regression (straight line fit) with ELO is 0.72. That means that on average they're essentially predicting the probability of winning 70% of the time. There's no implication about predicting who wins the game per se (because it's irrelevant).

 

You can't predict a probability.

Sent from my SAMSUNG-SM-G920A using Tapatalk

 

That's exactly what a lot of statistics does. In most cases you can just interpret it as predicting the frequency of occurrence, which can obviously be tested.

btw.. physics has the best example: predicting the probability of a subatomic particle being in a certain location. Quantum mechanics predicts the probability distribution there.

 

Also - probabilities can never be wrong, unless you're predicting a 100% probability.

Sent from my SAMSUNG-SM-G920A using Tapatalk

 

Two things:

1) They can be statistically speaking (so probabilistically) wrong, and that's what in practice is used in science/engineering etc..

2) Even predicting 100% or 0% you can never prove wrong after a finite number of trials. The probability of choosing any positive integer from the set of all possible positive integers is 0%. How do you know that the true answer is or isn't 0% after an infinite number of trials? No matter how many finite trials you have, you can't in principle prove or disprove it.

 

And what is the occurrence happening in an NFL game. It's either win or lose. You can only measure accuracy by wins and losses.

Sent from my SAMSUNG-SM-G920A using Tapatalk

 

In the case of betting, the accuracy is measured by profit over time, not whether a team won or lost in a single game.

 

We aren't talking about betting, we are talking about the predictive ability of ELO. How do we measure the accuracy?

Sent from my SAMSUNG-SM-G920A using Tapatalk

 

this thread is just for folks who have ivy league educations obviously..

 

Well, my argument WAS about betting.

But I'll answer your question anyway. If you have a ton of games between different pairs of opponents whose ELO rankings for the sake of argument do not change over time (in ELO they do but that is an added variable one can add if one wants), then you have the predicted probability of winning for every match.

What you now do is calculate the probability of the actual (observed) wins/losses by all teams assuming they came from the predicted probability distributions. If you don't mind using a computer, you can do that by using something called Monte Carlo simulations (so let the computer assume your predicted probability distributions are the actual ones, and randomly sample from them). For each simulation, you will get a result of wins/losses by all the teams in all the matches.

Do the simulation over and over again until you get a distribution of wins/losses for all the teams. You can now do traditional hypothesis testing on that distribution to determine how likely it was that the observed win/loss records came from your predicted distribution. That's how you would test your prediction.

 

It is a Harvard study lol..

Hey as long as Stringer Bell is asking questions, I gotta answer, and in some cases I can't answer without some technical detail. You can just skip over those posts though :wink2:

 

I think its all a fascinating discussion, but i'll beat your conclusions based on numbers in the long run using my techniques theories and knowledge.

 

"In baseball, my theory is to strive for consistency, not to worry about the numbers. If you dwell on statistics you get shortsighted, if you aim for consistency, the numbers will be there at the end." Tom Seaver

Here's to more Dolphins' consistency in playing well and winning!

(And less consistency in mediocrity)

 

My conclusion was only that the study's claim of validating their method (using AV to predict ELO rankings) on last year's data doesn't mean what it at first sight seems. The wording was ambiguous and you could interpret it as saying they used 2013 data to predict 2014 data (which would be impressive) OR that they used 2014 data to predict 2014 data (which isn't that impressive). I'm pretty sure it means they used 2014 to predict 2014 because if it was using 2013 to predict 2014, they'd be able to beat the bookies by a large margin.

 

anyone have a link of the guys from harvard doing that list last year ?

 

Well, I guess we'll see how smart these actually are.

 

This is a best case scenario. Tannehill improves, Suh is his usual dominant self, the WRs mostly replicate what they did on other teams, the guards are serviceable and don't get Tannehill killed, and no starter misses much time. That's possible (some teams don't have many significant injuries). But it seems more unlikely than not. I think those 8.5 lines are more realistic, with injury luck (though there's probably some skill there) bumping us over or under that line.

 

Yeah, this is the right question to ask. Just gotta remember that the rank order is not what's important. It's the numbers (the ELO ratings) themselves, which gives you a prediction for each game. In the NFL, ELO may predict the final ranking of teams in terms of record to be different than the actual rank order because you have different divisions with varying levels of ELO-rated competition.

I'd say if they can beat the average of the two moneylines per game (basically tells you what the bookies/bettors think will occur if you remove bookie profits), then they've got something. If not, then just go with the moneylines as your best prediction (or if you want predictions of end-of-season record, go with futures).

 

I actually think that I'd be on cloud 9 the whole time. Like a dream.