NFL ATS Model

[mebaran]

Stayed up tonight putting together an NFL model which I will tweak the first couple weeks. For the past two seasons, the model's average line error is around 11.3.

Is that about right? It's a bit worse than the book lines, and doesn't include any technical injury data. Is this comparable to what's "out there" in the modelling world?

Thanks for any input.

[buby74]

If you go to www.thepredictiontracker.com you will see that an 11.3 point error would put your system in the bottom half of the 50 or so systems that he tracks. So not too bad if it is just your first attempt but nothing remarkable. How do you set your line?

[donkson]

your model took one night?

[mathdotcom]

7PM - Turn on computer
8PM - Collect some stats
9PM - make a model
10PM - feeling of satisfaction

9AM - Beat the toughest market in all of N American sports

[mbs4]

By average line error, you mean for a game with a -3 line, your model had it set at something like -14.3?

[mathdotcom]

He means difference b/w his predicted spreads and the game's actual spread

Too bad you don't get extra money for covering by 50 points rather than half a point

[Bigbill365]

thats a nice site to get an idea of what everyone is thinking that should be a good tool i use and phil steeles prodictions

[mebaran]

your model took one night?

6-8 hours to build/automate last night, yes. But I've been researching and testing methods for a few weeks now.

[mebaran]

7PM - Turn on computer
8PM - Collect some stats
9PM - make a model
10PM - feeling of satisfaction

9AM - Beat the toughest market in all of N American sports

Trust me, I have no illusions of grandeur here, Mathy. It was more an exercise in model building than anything else, but for me to have come up with something in which I can squeeze out some edge on the absolutely horrendous lines, I'm fairly satisfied with my first NFL work.

With pretty significant filters on spreads, the model went 19-14-3 in 2011, and 24-17-2 in 2010. So I'm cool with just running it and learning from the process.

[mathdotcom]

Trust me, I have no illusions of grandeur here, Mathy. It was more an exercise in model building than anything else, but for me to have come up with something in which I can squeeze out some edge on the absolutely horrendous lines, I'm fairly satisfied with my first NFL work.

With pretty significant filters on spreads, the model went 19-14-3 in 2011, and 24-17-2 in 2010. So I'm cool with just running it and learning from the process.

And there's the catch

In 2012 you could find a total of x such that if you bet the over on every total of x you would have made a profit.

But would anyone, even some of the biggest squares, really think that's a recipe for success in 2013?

[mebaran]

I think I misspoke. The "filter" is eliminating games a play if the book's line differs from my projected line by x amount. I'm not playing something like "home favorites, when the wind is <20mph, and the dewpoint was....", nothing like that.

[MonkeyF0cker]

I think I misspoke. The "filter" is eliminating games a play if the book's line differs from my projected line by x amount. I'm not playing something like "home favorites, when the wind is <20mph, and the dewpoint was....", nothing like that.

Actually, you are.

[cyberbabble]

You're doing OK so far. Look at it as a homework assignment. Learn some stuff, see if you like the class, and maybe then take the advanced course next. NFL is tough, but you can use what you learn to try arena football (I don't even know if it exists anymore) or Canadian football. These are supposed to be easier than NFL.

Here's a few things to think about.
Time for a real time test. You need to freeze the model, no midseason changes. You can use prior weeks data to project this weeks game, but there should be no changes to the regression equation or whatever you are using. You need to make picks based on the model and not try to second guess it. If you want to add a few rules on top of the model, decide on them now. They need to be specific. For example, it would be OK to pass on the model selection if the starting quarterback is out for the game. But then what if it's the running back out instead of the quarterback? A good rule might be don't bet the first four weeks so that the teams can settle into their routines and you have four weeks of fresh data for the model. The idea of waiting a few weeks is pretty standard for modelers. You need to make any such rules now and follow them. Probably best to just wait four weeks and then take the models picks for this season.

You need to make some sort of bets as you go through the season. Bet points in the SBR book or maybe $1 bets at an online book. There will be some ambiguous situations or close decisions throughout the season. It is important to 'pull the trigger' and make the bet to get a good test of your model. Keep track of why you did what you did so you can evaluate the results correctly. You don't want to say "I'll think about this bet for now" and then after the game is over decide that you were going to make the bet and then count a win that you didn't really have.

You probably developed the model against either the open or the closing line. Your bets will be against a changing line which might matter. Your bets may be into a line different from -110/-110 and this will make a difference.

[mebaran]

You're doing OK so far. Look at it as a homework assignment. Learn some stuff, see if you like the class, and maybe then take the advanced course next. NFL is tough, but you can use what you learn to try arena football (I don't even know if it exists anymore) or Canadian football. These are supposed to be easier than NFL.

Here's a few things to think about.
Time for a real time test. You need to freeze the model, no midseason changes. You can use prior weeks data to project this weeks game, but there should be no changes to the regression equation or whatever you are using. You need to make picks based on the model and not try to second guess it. If you want to add a few rules on top of the model, decide on them now. They need to be specific. For example, it would be OK to pass on the model selection if the starting quarterback is out for the game. But then what if it's the running back out instead of the quarterback? A good rule might be don't bet the first four weeks so that the teams can settle into their routines and you have four weeks of fresh data for the model. The idea of waiting a few weeks is pretty standard for modelers. You need to make any such rules now and follow them. Probably best to just wait four weeks and then take the models picks for this season.

You need to make some sort of bets as you go through the season. Bet points in the SBR book or maybe $1 bets at an online book. There will be some ambiguous situations or close decisions throughout the season. It is important to 'pull the trigger' and make the bet to get a good test of your model. Keep track of why you did what you did so you can evaluate the results correctly. You don't want to say "I'll think about this bet for now" and then after the game is over decide that you were going to make the bet and then count a win that you didn't really have.

You probably developed the model against either the open or the closing line. Your bets will be against a changing line which might matter. Your bets may be into a line different from -110/-110 and this will make a difference.

Wow, a constructive comment in the sharktank? Who knew?

Thanks for the feedback! The model predictions start week 4, so I'm good there, and as far as line movement goes, I'm probably not banging away at a spread I like hours before game time. My play is jumping on some openers, pending any blatant injuries.

Trust me, I "like the class" . I've tried modeling three different sports, and though I lack some of the upper echelon math/programming skills, I'm working at it, I enjoy it, and I'm doing better than a breakeven bettor. So for now, I'm expanding my knowledge base with the final goal of being able to translate this experience into any sport I can get data on.

Underwater basket weaving....here I come (anyone willing to sell an underwater basket weaving database to me? Must go back at least 8 years, and include time of possession data!)

[MonkeyF0cker]

If you're just looking for some encouragement, then yes, it was a very constructive comment. Sadly, it's filled with poor advice.

You have nothing to bet. There is no statistical rationale as to why certain subsets of lines that are output by your model would outperform the market. You're doing nothing different than system players. You are pecking through past results and attempting to find criteria/patterns that have generated profit over a limited sample.

If you take cyberbabble's "constructive" advice, all you'll be doing is flipping coins and wasting your time.

Learn and move on. Improve upon it or scrap it. But there is absolutely no reason to paper trade it in any way.

[Spektre]

Mebaran,

Don't worry about MDC. His charms grows on you after a ...ok, nevermind. He's pretty much an idiot 24/7.

It looks like you've been around a while here though to be considered a rookie. It woudl be nice if you culd post your major steps in a blog or some such for the help of those of us on the starting end of the curve.

[MonkeyF0cker]

Mebaran,

Don't worry about MDC. His charms grows on you after a ...ok, nevermind. He's pretty much an idiot 24/7.
It looks like you've been around a while here though to be considered a rookie. It woudl be nice if you culd post your major steps in a blog or some such for the help of those of us on the starting end of the curve.

Exactly. Well, except that he's about the only one to offer any solid advice in this thread.

It's funny how posters that come asking for advice only listen to the advice that they were hoping to hear in the first place.

Anyways... carry on. I didn't mean to interject in your panhandling.

[cyberbabble]

You are on your way to being an HTT regular. You've had the MDC and MF discussions. I don't know what happened to the third musketeer, hutennis. You need a hutennis discussion and seal of disapproval to complete your membership application.

You've done some research and it's hard to know what to suggest. I tried to bring up a few things you might not have considered.

I said DON'T "paper trade". Make real bets for a token amount or for points in the SBR book here. This is very important for testing the model. It forces you to make a bet/not bet decision and provides a realistic betting line to measure your results against.

Don't tweek the model or rules during the season. That pretty much invalidates the testing. If you think of something you would you would like to do differently, then start with a new model, back test it, and then continue through the season with two models. This leads to the problem that if you make several models probably one will work OK and you become a victim of selection bias which also pretty much invalidates the testing. On the other hand you need to do real time testing and it will be a year before you can start with a different model so you might as well do as much as you can right now.

The books to read are Stanford Wong, King Yao, Elihu Feustel (Justin7 here). Not real heavy on math/stats, but the minimum reading list.

If you don't know the name Ganchrow, search the forum for his posts. Very heavy on math/stats but skim them anyway. You will pick up the names of some other posters worth reading. These guys were before my time but it is interesting to see what HTT was like and how good it was in the old days. I don't know what happened to these people. I think some people still read stuff here but they don't post much. Maybe they all meet in some secret forum that we're not allowed to join.

Props are generally considered the best thing to start with if you are at least semi-serious. I have two good seasons doing NFL props. If things go well then by midseason I will be betting enough to care whether I win or lose. So that's the only positive personal experience I have. My modeling attempts have all beat 50/50 over reasonable sample sizes but not by enough to cover the vig.

Good luck and continue posting.

[MonkeyF0cker]

I said DON'T "paper trade". Make real bets for a token amount or for points in the SBR book here. This is very important for testing the model. It forces you to make a bet/not bet decision and provides a realistic betting line to measure your results against.

So you are advocating BETTING an unprofitable model? Umm. You're serious? How ingenious.

Mathy and I have been here since the beginning.

And you obviously know what you're talking about.

[mathdotcom]

IQ(Players Talk) > IQ(HTT) for the first time in the history of SBR

Even the moderator of the HTT can't pass the skill testing questions

[mebaran]

If you're just looking for some encouragement, then yes, it was a very constructive comment. Sadly, it's filled with poor advice.

You have nothing to bet. There is no statistical rationale as to why certain subsets of lines that are output by your model would outperform the market. You're doing nothing different than system players. You are pecking through past results and attempting to find criteria/patterns that have generated profit over a limited sample.

If you take cyberbabble's "constructive" advice, all you'll be doing is flipping coins and wasting your time.

Learn and move on. Improve upon it or scrap it. But there is absolutely no reason to paper trade it in any way.

What should I be looking for then? Getting back to the original post, do I need to decrease my error to below the market's, or just get reasonably close?

What would be your minimum criteria to start using a model in live betting (avg. error, sample size tested, etc)? Any info would be helpful.

Appreciate the input.

[MonkeyF0cker]

If you're going to beat the market, your error needs to be smaller than the market's. If I were you, I'd honestly forget about the NFL for now. You're trying to take on the champ in your first fight. Try something (a lot) smaller in terms of limits.

[mebaran]

If you're going to beat the market, your error needs to be smaller than the market's. If I were you, I'd honestly forget about the NFL for now. You're trying to take on the champ in your first fight. Try something (a lot) smaller in terms of limits.

I'm definitely going about this in a roundabout way. Yeah its tough to model a spo4t that has such high limits as it attracts groups with more capital and sophistication than I. Was a learning experience this time around for sure.

I'll try my hand at other sports in the future for sure,bt with the availability of data for NFL, it was a good opportunity to try some analysis.

[mathdotcom]

Fukk, go to any online sportsbook

Load up CRIS, log the fukk in, and look at what they let you bet on

I don't know about you, but when I log in there are more options than "NFL". Even if there were only data for NFL (DUH, there is data for every sport), 'NFL' is but one of 7.

So if you want to get serious instead of playing around, open your eyes

Tired of these PLAYGROUND BETTORS

[hutennis]

I'm definitely going about this in a roundabout way. Yeah its tough to model a spo4t that has such high limits as it attracts groups with more capital and sophistication than I. Was a learning experience this time around for sure.

I'll try my hand at other sports in the future for sure,bt with the availability of data for NFL, it was a good opportunity to try some analysis.

I think there is a monstrous confusion here.
You are confusing the process you are involved in with modeling.
It is nothing of a sort.

What you are doing can, at best, be called "hunting for patterns".

You arrange past results in some specific way and, unavoidably, start getting "winning patterns".
Then you say to yourself " If wins has been coming in this situations with this frequency in the past, then history must repeat itself and wins will come in the same situations in the future. I have a model for predicting future results."

No, you don't. You simply found a pattern that worked in the past. One of millions of patterns that you could have found.

Whatever you have done is just one of many, many, many things that are needed to be done and thus can not possibly be the valid way to accomplish what you are looking for.

Here is what odds makers do. In a nutshell of course.

They take a ton of data,for every aspect of the game, for player and for team, past and present, (including the data you've been working with), then they rely on very sophisticated and heavily scientific formula/algorithm to analyse this mountain of data any possible way they can think of (including the way you analysed it),then, based on those analysis, they run simulations. Computer plays the game in question millions of times and spits out generated results.
From those results fair odds are being derived for ML, spreads, totals, what have you.
The above process is what I would call building and using a predictive model.

What you are doing has as much in common with it as folding paper plane has in common with building Boeing 747.
And it has the same chance of coming out ahead.

Please, don't get me wrong.
I'm not trying to put you or your efforts down.
I'm simply trying to point out what a horrendous waste of time this is.
You would be so much better off if you'd put the same time and effort into line shopping, it's not even funny.

[mebaran]

Fukk, go to any online sportsbook

Load up CRIS, log the fukk in, and look at what they let you bet on

I don't know about you, but when I log in there are more options than "NFL". Even if there were only data for NFL (DUH, there is data for every sport), 'NFL' is but one of 7.

So if you want to get serious instead of playing around, open your eyes

Tired of these PLAYGROUND BETTORS

Sorry. My fault for trying something I was interested in. Thanks for the info, as cryptic as you are being.

[mathdotcom]

Sorry. My fault for trying something I was interested in. Thanks for the info, as cryptic as you are being.

Next time start your thread with "Hi I'm mebaran and I just like playing around, I am not interested in any way in trying to make a profitable model."

[357vegas]

Can someone email me a sample of a football model? Please show me an example of what a model looks like and how it works.

thanks

[mebaran]

Can someone email me a sample of a football model? Please show me an example of what a model looks like and how it works.

thanks

Try looking up some academic papers on the subject to get ideas. Some are helpful.

[brewers7]

7PM - Turn on computer
8PM - Collect some stats
9PM - make a model
10PM - feeling of satisfaction

9AM - Beat the toughest market in all of N American sports

This was just F'ing hilarious...Laughed my ass off...No offense to Meberan, but he kind of set "Mathy" up for that hysterical retort...I only get here maybe once every 2 weeks, but always appreciate the quippy responses from "Mathy"...

I also enjoyed post #27 of this thread...And again, no offense Meberan...

Maybe one day I'll post a lot of my NBA theories (I don't do models) based on my 31 years of capping that sport and have a give-and-take with "Mathy" and hutennis in here as I think it would be interesting...But I want to finish my NBA database first (getting close) and check out some budding theories that pop up from time to time as the NBA is the most situational sport to cap (in my opinion) and I am all about situational spots...

GL with the model, Meberan...

Although I agree with Mathdotcom 100% here and have told people this for more than a decade and that is that the NFL is the toughest sport to beat out of the 4 major N. American sports to bet on...By far...Although there are situational spots in the NFL, too, just not many of them with the piddly 16-game schedule and having teams play once a week (save the bye weeks)...

[mathdotcom]

Don't even read hutennis's posts. I feel dumber as a result of not following my own advice.

- "Mathy"

[Maverick22]

Don't even read hutennis's posts. I feel dumber as a result of not following my own advice.

- "Mathy"

How is being a douche working out for you?

[mathdotcom]

I am sorry you feel inferior because you aren't able to contribute here

[evo34]

Have you ever interacted with (or even met) an "odds maker"? If so, who? If not, refrain from speculating what they do and do not do to arrive at opening prices.

I think there is a monstrous confusion here.
You are confusing the process you are involved in with modeling.
It is nothing of a sort.

What you are doing can, at best, be called "hunting for patterns".

You arrange past results in some specific way and, unavoidably, start getting "winning patterns".
Then you say to yourself " If wins has been coming in this situations with this frequency in the past, then history must repeat itself and wins will come in the same situations in the future. I have a model for predicting future results."

No, you don't. You simply found a pattern that worked in the past. One of millions of patterns that you could have found.

Whatever you have done is just one of many, many, many things that are needed to be done and thus can not possibly be the valid way to accomplish what you are looking for.

Here is what odds makers do. In a nutshell of course.

They take a ton of data,for every aspect of the game, for player and for team, past and present, (including the data you've been working with), then they rely on very sophisticated and heavily scientific formula/algorithm to analyse this mountain of data any possible way they can think of (including the way you analysed it),then, based on those analysis, they run simulations. Computer plays the game in question millions of times and spits out generated results.
From those results fair odds are being derived for ML, spreads, totals, what have you.
The above process is what I would call building and using a predictive model.

What you are doing has as much in common with it as folding paper plane has in common with building Boeing 747.
And it has the same chance of coming out ahead.

Please, don't get me wrong.
I'm not trying to put you or your efforts down.
I'm simply trying to point out what a horrendous waste of time this is.
You would be so much better off if you'd put the same time and effort into line shopping, it's not even funny.

[chunk]

This was just F'ing hilarious...Laughed my ass off...No offense to Meberan, but he kind of set "Mathy" up for that hysterical retort...I only get here maybe once every 2 weeks, but always appreciate the quippy responses from "Mathy"...

I also enjoyed post #27 of this thread...And again, no offense Meberan...

Maybe one day I'll post a lot of my NBA theories (I don't do models) based on my 31 years of capping that sport and have a give-and-take with "Mathy" and hutennis in here as I think it would be interesting...But I want to finish my NBA database first (getting close) and check out some budding theories that pop up from time to time as the NBA is the most situational sport to cap (in my opinion) and I am all about situational spots...

GL with the model, Meberan...

Although I agree with Mathdotcom 100% here and have told people this for more than a decade and that is that the NFL is the toughest sport to beat out of the 4 major N. American sports to bet on...By far...Although there are situational spots in the NFL, too, just not many of them with the piddly 16-game schedule and having teams play once a week (save the bye weeks)...

Haven't had a lot of time to spend in here lately because it "tis the season", but just a couple of observations/questions. After 31 years of capping, why are you just finishing your database? Not my strength, but agree that it is conducive to situations.

Agree that the NFL is the toughest to beat of the majors. Anyone that ventures into that minefield too often is asking for damage. There are opportunities...30 to 40/yr max. Based on reasonable return of course.