Logistic regressions

[jarhead60]

Has anyone worked with logistic regressions for ats winners/losers?

Any thoughts or hints as to what to use for independent variables??

[uva3021]

Yes. Typically there is very little to no causal correlation between certain statistics and ATS record. If there happens to be a high relationship, its probably accidental. Though one thing that may show to be connected is pythagorean win percentage vs actual win percentage. If the pythagorean is higher, than it could be reflected in their ATS record since they are scoring more points than what can be inferred from their SU record.

[jarhead60]

yes, hard to find good independent variables...I have had some luck with setting up for home teams, and using home visitor difference with iv's as non-conference game, time zone change between home and visitor, last game home or away, and average turnover per game.

Get very good p values, very good chi square model fit, but probabilities for each game range from .31 to .68

sample size was 2569 games

looking backward, assume bet on if prob .55 or higher and bet against if prob .45 or lower, would have won 589 out of 1023 bets. did not do out of sample test


My theory of why most box score data in nfl is useless is that the plays called (and therefore most of the box score data ie yards, passes, rushes. etc) are dependent on the score, field position and time remaining. If that data is dependent on the score, then that historical data can't predict the future score.

[uva3021]

Originally Posted by jarhead60

yes, hard to find good independent variables...I have had some luck with setting up for home teams, and using home visitor difference with iv's as non-conference game, time zone change between home and visitor, last game home or away, and average turnover per game.

Get very good p values, very good chi square model fit, but probabilities for each game range from .31 to .68

sample size was 2569 games

looking backward, assume bet on if prob .55 or higher and bet against if prob .45 or lower, would have won 589 out of 1023 bets. did not do out of sample test


My theory of why most box score data in nfl is useless is that the plays called (and therefore most of the box score data ie yards, passes, rushes. etc) are dependent on the score, field position and time remaining. If that data is dependent on the score, then that historical data can't predict the future score.

this is a very good point, and i think you are on the right track, stick with situational and time zones, maybe isolate situations that could act as independent values that not only possess a solid rational approach but backed by statistical support

after bye weeks, time zones, road wins as underdogs, weather, etc...