[Dark Horse]

Action Points To Percentage

Like many, I use action points to help determine the strength of models. But the problem with action
points is that you can't add them all up and divide them by the number of games, because 3 pts has
a different value depending on whether the total is 36 or 47.

So I tried to define a calculation to translate action points into percentages by recalibrating the score
into what it would have been if the game had ended exactly on the game total.
Example. Line: Fave -7 with 47 total (inherent score 27-20, with 7 pts being 14.9% of the total).
Result: Dog wins 23-20. Translated to 47 pts that score would be 25.14 - 21.86 (total 47).
So the dog beat the favorite by 3.28 pts against the 47 total, or 7.0%.

Am I correct to reason that the dog beat the spread not just by 10 action points, but by 21.9%? (the
dog covered the 14.9% plus an extra 7.0%). In this example, the 21.9% advantage would translate
as 60.95 % - 39.05%.

If the same +7 dog had lost the game 20-26 (20.43-26.57) the ATS win would have an action points
to percentage number of 4.55% with a distribution of 52.275% - 47.725%.


======================================== =============================

ATS And APTP (Action Points To Percentage)

Hypothesis. If action points can be used to determine the strength of a model in a more reliable way
than just the ATS %, then action points to percentage could, possibly, give a (better) read on the
influence of luck.

The ATS % and the APTP numbers are not the same, but they do point in the same direction in that
they both qualify the strength of a model. Therefore, if a long term 57% ATS model is accompanied
by a long term 60% APTP number, the influence of luck would have been less than if a 57% ATS
model were accompanied by a 54% APTP number. If this reasoning is accurate, or going in the right
direction, then one may be able to identify when a model is over or underperforming, and adjust bet
size accordingly.

Basically, one would have two graph lines for every model, and I believe those two may be better
than one. Especially in fields of smaller sample sizes, such as sports betting, this could be a useful
approach.

======================================== =============================

Thoughts?

[dialup_king]

There will be some big errors in certain sports. In NFL, If the total is 36 and the home team is favored by 3, they won't be favored by 3(47/36)~3.9 in a game with 47 total. It won't even be 3.5.

The posted total is median. After the game ask yourself what would the median total be if the final score was the average total

[Dark Horse]

There will be some big errors in certain sports. In NFL, If the total is 36 and the home team is favored by 3, they won't be favored by 3(47/36)~3.9 in a game with 47 total. It won't even be 3.5.

Books don't like moving off 3.

Factor in the juice?

[buby74]

In college football the total has almost no effect on the moneyline for a given pointspread. in theory if you take all games that had a 7 point favourite the moneyline odds should be worse at lower totals when in fact there is very little correlation
So maybe you don't need to worry about it.

[Dark Horse]

I mentioned this to a friend with a degree in math and statistics. He explained that what I did is called normalizing.

Thinking along these lines, I would suggest an approach that could identify edges well before the Z-score can. With the game scores normalized against the posted totals, take the league's average for the percentage by which every winning team over an entire season (or seasons) beats the spread. That's 'ground level'. Then do the same for every winner in your model. The difference in percentage will give an indication of how strong your model is; because it compares your winners against all winners. So if every winning team in the league over an entire season beats the spread by 5% (ground level), and the winners in your model beat the spread by 10% you got 'something'. In my (still experimental) view this difference could be a more reliable indicator than a 'flat' ATS distribution, and could be especially useful for bet sizing.

It can be broken down more. If a model picks dogs, compare your winning dogs to all winning dogs, etc. There's always a ground level. And a model's success rate can be compared to it.

[pedro803]

Dark Horse, if I understand what you are saying correctly, then I think what you are calling 'ground level' is sometimes called 'baseline' in quantitative analysiss parlance -- there has been some discussion of this somewhere on the tank, e.g. when including a team's avg points per game in a model then league average points per game should also be included in the model, it is a little hard to 'get' but it seems that a lot of experienced statistical analysts agree that this 'baseline' is a must for inclusion if you want a good model.