Rating algorithm that natively predicts moneyline?

[Spektre]

The more I think about this the more I think the answer is no, but I'd like some input from the pros here.


Most rating algorithms chew up some stats and spit out a rating. If you condition the scale of your system, that rating could corresponds to an expected point spread between the two teams. So while some systems give a numeric rating to a team on an arbitrary scale, some can scale it natively to a point spread.


When making a moneyline predictive algorithm, that doesn't seem possible. What generally happens is the numeric rating is provided and then, through back testing of the dataset, a histogram is formed which determines the probability of teams with these specific ratings winning over each other.


Has anyone developed or seen a system that "natively" determines a percent chance of winning without resorting to histograms?

[hus7ler]

idk what histograms are but i am deductung that you want a rating system that doesnt use previous games as part of the variable. this might be an oxymoron because you can only get those ratings based on prev games played.

nvm, misunderstood ur post. why dont you test one out urself and give us the results

[evo34]

Huh? Any model should be able to estimate odds of winning if it is able to estimate odds of covering a spread.

[Spektre]

Huh? Any model should be able to estimate odds of winning if it is able to estimate odds of covering a spread.

How so, without using a histogram of past results?

[evo34]

Not sure what to tell you. Most models (in my experience) predict the odds of winning, and then specific MOV odds are derived from histograms -- not the other way around.

[Waterstpub87]

I don't know why you would need a histogram. It is just a graphical representation. You would need a distribution.

At any rate, if you knew the point spread, you could convert it to money line with the calculator in the tools section on this board.

I guess you are doing something like ranking teams on an ordinal scale, and saying something like " Number 1 team should be a 3 point favorite over Number 4 team". I don't know why you couldn't do something like "Number 1 teams win 60% of the time against a Number 4 team".

It isn't clear what you are asking based on your first paragraph. I would assume that most models don't pump out ratings, they pump out expected point spreads, ml and totals between two teams. It seems kind of backwards to rate teams, then back test to get point spreads.

[Spektre]

Not sure what to tell you. Most models (in my experience) predict the odds of winning, and then specific MOV odds are derived from histograms -- not the other way around.

Interesting. Perhaps you could point me to a simple example of such a system. I have never seen one that natively derives a percent change of winning.

I don't know why you would need a histogram. It is just a graphical representation. You would need a distribution.

At any rate, if you knew the point spread, you could convert it to money line with the calculator in the tools section on this board.

I guess you are doing something like ranking teams on an ordinal scale, and saying something like " Number 1 team should be a 3 point favorite over Number 4 team". I don't know why you couldn't do something like "Number 1 teams win 60% of the time against a Number 4 team".

It isn't clear what you are asking based on your first paragraph. I would assume that most models don't pump out ratings, they pump out expected point spreads, ml and totals between two teams. It seems kind of backwards to rate teams, then back test to get point spreads.

Yes, you are correct. The distribution that creates the histogram is what is needed. I believe the calculator you refer to uses a distribution of past games to calculate the moneyline from the point spread. That is what I am intersted in seeing NOT happen.

As an input to many algorithms are past game scores. Indeed the simpest of them (Massey, Colley for example) that is the only input. An output can therefore easily be scaled to a point spread. and be used to bet such. The moneyline however cannot (that I have seen) be derived in this manner.

In general a point spread (or a rating readily converted to a pointspread) is found and then the moneyline is calculated from this based on a past distribution of games.

[Baeoz]

Interesting. Perhaps you could point me to a simple example of such a system. I have never seen one that natively derives a percent change of winning.



Yes, you are correct. The distribution that creates the histogram is what is needed. I believe the calculator you refer to uses a distribution of past games to calculate the moneyline from the point spread. That is what I am intersted in seeing NOT happen.

As an input to many algorithms are past game scores. Indeed the simpest of them (Massey, Colley for example) that is the only input. An output can therefore easily be scaled to a point spread. and be used to bet such. The moneyline however cannot (that I have seen) be derived in this manner.

In general a point spread (or a rating readily converted to a pointspread) is found and then the moneyline is calculated from this based on a past distribution of games.

When calculating the moneyline or probability from the differences of such ratings, do you use all match results, or only results that you have match odds data for, or results from a similar division/level? I've calculated normal curves based on past data for tennis using ratings, then used that to determine the probability, and then used that inverse to determine the odds/moneyline and it's worse than throwing a brick up in the air blindfolded and hopping you hit a target. It's just better to bet with the higher rated player with odds greater than $2. Sometimes you come out in front.

The thing that bothers me, is the results in the past are normally distributed based on rating difference and pretty much same mean and standard deviation across all levels, so it seems at first blush a reasonable exercise.

[evo34]

@Spektre

Yeah, if you're using off the shelf power ratings, I guess you start with margin of victory. If you create your own model, I have no idea why it wouldn't be able to predict win odds. Why would you not be able to use past scores to estimate future odds of winning?

[Cookie Monster]

The Elo system, used first for chess has been adapted for use in many other sports. It designed to compute the winning probabilities of a team, given a rating difference. So, it is a system natively designed to compute moneyline instead of spread. Sagarin and fivethirtyeight base their ratings on Elo.

IMO, a system based on spreads is worse, for two reasons:
1.- In some sports such as football, the spread distribution is uneven. The spread 3 is worth much more than the 5. A continuous spread-based power rating does not account for that.
2.- The point spread difference is not transitive, it breaks at the extremes. If A team is favored 10 points over team B, and team B is favored by 15 over team C; A should not be directly favored by 25 over C. The winning probability ratings system use a logistic function to give diminishing returns for greater rating difference, while linear spread system does not.