Hey Ganch, some stats help needed.. thanks

[newb411breaker19]

Hi all,

I originally sent this post in a private message to ganchrow, but he wisely suggested I post this question in the think tank, so that others may benefit from this question.

"hey ganchrow,

hope you're doing well. I was looking at dr. bob's website the other day and i'll post the part that i need some help on.. only if you have the time though, thanks a lot

"I suggest creating a spread sheet of your power rating or math model predictions that contain the actual line (in terms of the home team, where favorites are negative numbers. Thus, a 7 point home favorite would be -7), your power rating/math model prediction (also in terms of the home team, but use positive numbers if you favor the home team and negative numbers if your ratings forecast the home team to lose), the line differential between your line and the actual line (your line + actual line, so positive numbers represent a play on the home team and negative numbers represents your model picking the road team), actual game point differential (home score - road score) and home pointspread result (1 if the home team covered the spread, 0.5 for a push, and 0 for a spread loss by the home team). After compiling a year or two of actual predictions and results - not back-fitted predictions using games that you used to derive your model - you can begin to see if your model is better than the actual line. Simply use statistical software, or analysis available on Excel, to create a regression equation predicting home team spread result as a function of the line differential of your power ratings/math model from the actual line. For instance, I have 6 years using my NFL math model and the equation to predict the chance that the home team covers the spread is .505 + 0.0128xLD, where LD is the line differential between my math model prediction and the line. So, for every point differential, I can add 1.28% to my chance of winning (which is about 50% of the actual value of a points of true line value - so the difference between my model and the actual line is about 50% the mistake of the oddsmakers). If my model projects a 4 points home favorite to win by 10.0 points, then they would have a 58.2% chance of covering based on the past predictability of my math model (.505 + 0.0128x6.0 = .5818), without accounting for any positive or negative situations applying to that game (situational analysis is explained later)."

I was just wondering how you would go about setting this up in excel to determine the regression equation.

I guess what makes this method a problem would be whether or not the market itself is getting sharper over time, which I'm assuming happens from year to year.

thanks a lot for your help, I really appreciate it

newb411"

Thanks everyone

[Data]

http://office.microsoft.com/en-us/excel/HA011119631033.aspx?pid=CL100570551033

[Ganchrow]

Data's link provides detailed instructions for how one might perform such a regression.

A major word of caution, however, is that it simply assumes a linearity that clearly doesn't exist. This is especially apparent in, for example, the NFL, where a 1-point line differential between 2.5 and 3.5 is very different from a 1-point line differential between PK and 1.

[newb411breaker19]

hey ganch, yeah thanks, I am aware of the huge difference in pointspread push frequencies, I think the push percentage on 3 is 10%, and of course certain numbers like 7, 10, 14, 4, etc.. are more important..

I actually emailed dr. bob once on this specific question because when he used to show his math model numbers, I didn't understand his expected win percentages for certain scores, but he explained that those percentages were strictly based on the point differential based on the past predictibility of his math model

thanks data for the link, i appreciate it.

[Justin7]

You might write your own regression. Instead of mean-squared, take the sum of the square-roots. This will de-emphasize extreme results... And you are more concerned about close results where the spread matters.

[Ganchrow]

Instead of mean-squared, take the sum of the square-roots. This will de-emphasize extreme results... And you are more concerned about close results where the spread matters.

Unless you fully understand and are confident in the statistical ramifications of such a methodology, I'd highly recommend against this.