[a4u2fear]

Ok, I have posted some systems in the past, but this thread is not about any of those.

Would like your guys' opinions on the following:

I found a way to estimate each team's points per game (everyone has their own).
We'll call the sum of these two estimates = X, which will be compared to the Vegas Total for the game

This estimate X is calculated for weeks 9 to 17 (to allow some teams performances to become more clear).

I calculated this estimate from 1985 until 2011 (again weeks 9 to 17)

Using linear regression, I found an equation for the comparison of my estimate (this is both teams estimates combined so it is effectively my estimate for the game) and the actual outcome total of the game.

Now that I had the linear equation, I used this equation to estimate the total from 2012 and 2013 (weeks 9 to 17).

Findings:
When the linear equation is 2 points or more lower than the Vegas Total, the game goes under 60.8%, or for 59 wins and 38 losses.

When the equation is 4 points or more lower than the Vegas Total, the game goes under 63.9%, or 30 wins and 17 losses.
Obviously the 2 point difference is more profitable, but this one is still good none-the-less.

When the equation is 6 points or more higher than the Vegas Total, the game goes over 80% or 8 wins and 2 losses.

WHen the equation is 7 points higher or more higher than the Vegas Total, the game goes over 100%, or 4-0.

Any thoughts on this? I think it's very significant.

I also did the same tests using weeks 7 to 17, but the performance was a little worse (but still very good).

[a4u2fear]

How can I effectively test this even more? Create the regression equation from 1985 to 2010 then test 2011,2012,2013? And then again from 1985 to 2009 then test 2010-13 etc. etc.?

[a4u2fear]

any thoughts? Another huge part of this:

During weeks 9-17 from 2012-13, the over under was 136-136-2. So Vegas was dead nuts on. Yet, during these weeks in games where my formula estimated a difference in the line being 2 pts higher than it should, I found 97 occurrences, where the formula went 60.8%!!!!! In just these two years.

Huge.

[Miz]

you should check to see if the line movement agrees with your predictions. If so, then you can feel better about your results being real. Also, if you include too much old data, you introduce noise into your results. THink about how different the game was in 1985, rules, game strategy etc. Maybe try using the previous 2 years as your training set, optimize on the next year and blindly go forward and test the year after that. A trap many people fall into is tweaking on the same set as they are testing. Dig through old posts in here and you'll find that many people have asked good questions and gotten good responses. As of late, not so much.

[a4u2fear]

you should check to see if the line movement agrees with your predictions. If so, then you can feel better about your results being real. Also, if you include too much old data, you introduce noise into your results. THink about how different the game was in 1985, rules, game strategy etc. Maybe try using the previous 2 years as your training set, optimize on the next year and blindly go forward and test the year after that. A trap many people fall into is tweaking on the same set as they are testing. Dig through old posts in here and you'll find that many people have asked good questions and gotten good responses. As of late, not so much.

agree on the noise from older years and I have tested other ranges (i.e. 2000 - present) and developed a modeling accordingly.

I'll take a look at line movement, but it will take a bit and report back. thanks.

[a4u2fear]

Miz,

What do you mean by movement agrees with predictions? How do I classify this? Sometimes line goes up and sometimes down but the system says play the under. Shoudl I count how many times the closing line moved lower and it was also a play in my system vs how many times it didn't?

Also-what if I looked at the model and used only the closing lines as the sides I would bet? In other words, all of my plays would be decided on or about 1 hour prior to each game which the total is usually set at a certain number.

[matthew919]

1. Where are you getting your lines? Presumably these are openers that you are testing against? If not, they should be.

2. Checking the agreement means that you select the subset of games that you would have bet (against the opener) and comparing the open/close for those games. If your method is valid, lines should move in accordance 60+% of the time.

3. Always beware of bad data. Select a subset of lines for your plays, and check them manually.

[Miz]

Matthew's right. 60% of the time is a good rule of thumb. Generally as the difference between your prediction and the line gets larger, the agreement percentage will increase too

[a4u2fear]

So, there were 102 games I bet on (using the method) for 2012-13 where my line was 2 points or lower than the Vegas line.

Out of these 102 games, the closing line moved higher than the opener 54 times, lower 35 times, and neither 13 times, or 60.6% moved higher.

I'm not sure if this validates it either way?

[a4u2fear]

I don't understand why checking picks against the opening line vs closing when betting the total matters. My bookie sets the lines one hour prior to each game and rarely are there significant moves in totals at that time. So I am almost always betting against the closing line, so why not model it as such?

if you're checking it against the opening and I find an edge, I wouldn't be able to bet the opening line anyway.

[marcoforte]

Warning, the rules in NFL have changed so much that total betting is much more difficult. I used to bet tons of totals with winning seasons through 2008 using a database back to 1985. Since then have been a loser. I would suggest limiting your data to 2008 and later for the nfl. I use binary logistic regression as my tool of choice.

[statnerds]

I don't understand why checking picks against the opening line vs closing when betting the total matters. My bookie sets the lines one hour prior to each game and rarely are there significant moves in totals at that time. So I am almost always betting against the closing line, so why not model it as such?

if you're checking it against the opening and I find an edge, I wouldn't be able to bet the opening line anyway.

first, you really need to get more outs than a Local hanging lines an hour before game time. Get more Locals too though as mine gives me any college game i want at 11 a.m. EST and does not update lines through out the day.

second, not sure how to explain this well enough without getting too wordy. checking your numbers against the openers allows you to then see what the Market did as well. If you find a large number of totals you would have bet move against you, that is bad news. if the opposite is true, ie the Market is moving toward your number, that is good news. to illustrate, let's say you found 200 openers you would have played against. 122 moved with you, 58 moved against you and 20 finished flat. the Market has validated your model, but only on a past results basis.

finally, long term it is all about the number. get the best number you can get. line shopping. paying the least vig you can on a play. use your own data above to see what a major difference this could make. if you are hitting 60% against the closers, what % increase could you expect if you were hitting openers and produced a long term positive CLV ~1 point?

GL

[James Marques]

For the last decade or so, both college and NFL games have followed a very accurate logistic model with point spread as the independent variable and favorite winning-percentage as the dependent variable. In fact, the NFL R-squared value is about .91 (college is slightly less, .87).

If you are able to figure out what F's winning percentage would be over U using a separate methodology (using the formula above wouldn't tell you anything you didn't already know), you could work the above equation backwards and solve for point spread. I calculated the above regression for college from 2004-2011, and then used a separate method to calculate favorite win % for 2012 and 2013. 2012 wasn't all that good 52.9%, but last year it hit about 55.7% ATS and (surprisingly) 58.3% on totals. Now, of course, I didn't bet anything on either season... I just wanted to see how it would do. I did of course make my yearly "ultra-sharp" $20 bet on Georgia Tech to win the national championship at 300:1, but sadly I just missed it. : )

I'm excited to post some plays this coming season and see if some of the minor tweaks I've been making will pay off. I may even take some 5Dimes poker winnings over to the book and bet 11 games a week and see how I do.

[nash13]

if you can provide the data i can test it in spss and show you if there is a bias based on the year you use or not.