[neila]

I have a question concerning Power Rankings for those of you who do simulations for NBA. Basically, do you use some sort of Power Ranking in your model? The problem I am having is in direct comparison of teams playing out of division. Teams playing out of division only play each other a couple of times a year, so direct comparisons becomes problematic. You have the ongoing problem of not enough recent data and long term past data being irrelevant. For example, if team A only plays team B twice a year, I have to incorportate ten years of data to get twenty games of head to head competition. And, what happened ten years ago is probably not reflective of the two teams today.
Just wondering if Power Rankings would be beneficial here. Any thoughts, comments?

[Indecent]

I have a question concerning Power Rankings for those of you who do simulations for NBA. Basically, do you use some sort of Power Ranking in your model? The problem I am having is in direct comparison of teams playing out of division. Teams playing out of division only play each other a couple of times a year, so direct comparisons becomes problematic. You have the ongoing problem of not enough recent data and long term past data being irrelevant. For example, if team A only plays team B twice a year, I have to incorportate ten years of data to get twenty games of head to head competition. And, what happened ten years ago is probably not reflective of the two teams today.
Just wondering if Power Rankings would be beneficial here. Any thoughts, comments?

Honestly, I think you need to drastically re-think your approach.

[neila]

Honestly, I think you need to drastically re-think your approach.

Actually, this is not an "approach". It is a refinement or rather an adjustment. So, can I can conclude from your answer that you don't consider how NBA teams match up against one another to be of any predictive value?

[Justin7]

neila,

inter-team games from a year ago are almost irrelevant. Games from 2-3 years ago are 99% irrelevant.

Power rankings look at all games a team plays in the current year (or sometimes in the previous year) to determine the overall strength of a team.

Individual match-ups between teams have some impact, but your approach seems to be minimizing the current year's data, and overestimating prior years' data. A different approach to try might be to emphasize this year's data, and quantify how how the match-ups between the two teams affect the relative power rankings.

[neila]

Thank you both for your responses. The model I have built is very flexible. I can vary the data set from the present back to 1990. That is as far back as my database goes. The problem I'm having is that it is too restrictive, i.e., not enough bets. Backtesting for the past 5 years only yields 71 bets when my line is 6 or more points from the actual spread. Less than 6 is not profitable. I'm looking for ways to open it up without sacrificing too much of an edge. Ideally, I'd like to get at least that many bets in a season. By the way Justin, I'm enjoying the book.

[Peregrine Stoop]

look into the APBR group
while SBR is king in many areas, there are some that are publicly a few strides ahead

[neila]

look into the APBR group

Thanks P, great site. Much reading there.

[Justin7]

The problem I'm having is that it is too restrictive, i.e., not enough bets. Backtesting for the past 5 years only yields 71 bets when my line is 6 or more points from the actual spread. Less than 6 is not profitable. I'm looking for ways to open it up without sacrificing too much of an edge. Ideally, I'd like to get at least that many bets in a season. By the way Justin, I'm enjoying the book.

This is a common problem. Don't get greedy. If you find 14 good bets in NBA each year that hit 60%, and you do the same for 4 other sports, your bankroll starts growing without a lot of work.

[pat venditto]

Power rankings is whoreshit. Or can it really be useful?

[Wrecktangle]

I'll take the "or"

[Wrecktangle]

neila, try a few differing sizes of "look-back" games and watch how your mean squared error (MSE) changes. Take the one that minimizes the MSE. Make sure you have at least two seasons in your in-sample trial.

[aleem15]

Personally never looked at power rankings but going to look into it a little more.

[neila]

neila, try a few differing sizes of "look-back" games and watch how your mean squared error (MSE) changes. Take the one that minimizes the MSE. Make sure you have at least two seasons in your in-sample trial.

Thanks W, I just checked that formula and I believe I can easily incorporate that into my model. I'm already monitoring how closely the pointspreads I'm generating are to the actual outcomes of the games and of course, how they compare to the books lines.

This model is becoming a monster. It is churning out all this information and now I have to learn to interpret what it is telling me. My next big assignment is to refine my backtesting procedures so I can have the program step through the variables and find the better parameters all the while being careful not to "fit" the model to any particular data set. I may have to start feeding this thing.

[Wrecktangle]

neila, another thing to try: use Mean Absolute Error rather than squared error. Depending on the technique used, MAE may turn up better results.