[Justin7]

No. A model that beats the line move 55% of the time might show a win-rate of 51%.

[JustinBieber]

Just out of interest what do you think your edge would be if you could beat the closing line(no vig) by a few cents? Like if a no vig line closed at -105 and you were allowed to bet -102 every time.

[MarketMaker]

Just out of interest what do you think your edge would be if you could beat the closing line(no vig) by a few cents? Like if a no vig line closed at -105 and you were allowed to bet -102 every time.

The implied odds of a no vig line of -105 are a probability of success of 51.2195%. To breakeven at -102 you must win 50.495% of the time. Your edge could then be defined as 0.7245%. According to kelly you should wager 1.46% of your bankroll on each bet.

[JustinBieber]

The implied odds of a no vig line of -105 are a probability of success of 51.2195%. To breakeven at -102 you must win 50.495% of the time. Your edge could then be defined as 0.7245%. According to kelly you should wager 1.46% of your bankroll on each bet.

ty, so basically this would be a lot better than a model which predicts 55% of line movements correctly?

[MarketMaker]

That depends. In something like the NFL where something like 16% of totals close the same as they open I imagine predicting 55% would yield better than a 51% win% as Justin suggested is possible. To answer your question someone would have to figure out how to use line move success rate to predct win%.

[MarketMaker]

Correlation coefficient significance?

Anyone have an idea what the significance of a correlation coefficient of .591256 is when comparing the data sets of model line vs market line differential in relation to actual line move, sample size is 91.

[Indecent]

I'll have to strongly disagree.

A standard modeling technique is to pigeon hole. What is the differential between your model's line, and the market line? group a: 0-2.5; b: 3-5.5; c: 6-8.5, ... If your model is good, you will see an increasing return (with fewer and fewer plays) as the differential increases.

Since the comparison is based on the market line and the market is adjusting/changing, how much should you use most recent results when evaluating long term edge? It seems likely that more recent results should be factored more heavily in estimating long term results (I think the market today is more similar to 2008 than 1998), but it doesn't seem clear what should be a good "cutoff" date for evaluations like this, or the best way to determine it.

I've been thinking about this recently as some models of my models are performing very well in the 0-2.5 range (with a 10% decrease in frequency) over the last over the last 4 years (2 of which were not used to train it) and I've been trying to determine what is causing the sudden decrease in frequency and increase in accuracy. Also, since I'm seeing such a dramatic increase in accuracy over these 4 years when compared to the previous 10 years, I'm not sure what that says about the reliability of my model moving forward. Or is it unwise to use the lower values (0-2.5,2.5-5) at all due to the market fluctuations? I don't want to do everything right up until this point and then put too much value in (relatively) short term results.

As far as the model is concerned, I'm certain I followed all basic principals to avoid over fitting issues so that's no concern, and the performance is actually better on the years it was not trained on. The 10+ accuracy stays very solid over all the years and both the ats winner and overall winner accuracy increases steadily in each progression after 0-2.5.

Am I being overly analytic/paranoid in being concerned in the dramatic difference when predicting close to the spread? Should I ignore those results altogether and focus on the overall record of the model? Or should I just shut up and celebrate?