Is a winning model at one sportsbook applicable to other books?

[podonne]

Greetings,

Was struggling with this question today. To properly parameterize and backtest a model you have to provide the spread\odds of every game. Ideally I'd imagine that you want all those spreads\odds to come from the same book, and not be "best available" or "average". So, say I find a +EV model using historical lines from Pinnacle. Is it reasonable to assume that I could take the same model and apply it to another book (with the same juice), like say BetOnline? Since I'm a US player, Pinnacle is out for me.

My first answer was that when my Pinnacle model took a side in a game, I could bet anywhere so long as the line was equal or "better" than Pinnacle, and that makes sense, but its also an additional parameter in the model that is much harder to backtest (something like "bet only when Pinnacle is not the best available line").

Any thoughts? Do you build models\systems that apply only to particular sportsbooks and fail on others?

Thanks,
podonne

[Pancho sanza]

Leave lines out of it, find fair value of whatever you're modelling and then you'll have a strike price.

[Duff85]

Greetings,

Was struggling with this question today. To properly parameterize and backtest a model you have to provide the spread\odds of every game. Ideally I'd imagine that you want all those spreads\odds to come from the same book, and not be "best available" or "average". So, say I find a +EV model using historical lines from Pinnacle. Is it reasonable to assume that I could take the same model and apply it to another book (with the same juice), like say BetOnline? Since I'm a US player, Pinnacle is out for me.

My first answer was that when my Pinnacle model took a side in a game, I could bet anywhere so long as the line was equal or "better" than Pinnacle, and that makes sense, but its also an additional parameter in the model that is much harder to backtest (something like "bet only when Pinnacle is not the best available line").

Any thoughts? Do you build models\systems that apply only to particular sportsbooks and fail on others?

Thanks,
podonne

If you have a model that can beat Pinny over the long term (on a major limit sport)... retire to Costa Rica and print moneyz.

[podonne]

Leave lines out of it, find fair value of whatever you're modelling and then you'll have a strike price.

I've heard this a few times, but I don't believe you can ignore the lines.

If your line and the book's line disagree, then either your line is a better reflection of reality or the book's is. The only way to tell which is the case is to test on new data, betting whenever your line and the book's line disagree and seeing if you end up with more money.

Hence, you need a source for lines to compare your line against, and we're back to my original question.

[podonne]

If you have a model that can beat Pinny over the long term (on a major limit sport)... retire to Costa Rica and print moneyz.

It rains a lot in Costa Rica. Then again I live in Seattle, so who am I to talk? :-)

Seriously, though, I'd have to demonstrate that my models at Pinny worked on new (as in future) data to retire on it, and I can't do that unless I can bet at Pinny, so catch 22.

[Pancho sanza]

I've heard this a few times, but I don't believe you can ignore the lines.

If your line and the book's line disagree, then either your line is a better reflection of reality or the book's is. The only way to tell which is the case is to test on new data, betting whenever your line and the book's line disagree and seeing if you end up with more money.

Hence, you need a source for lines to compare your line against, and we're back to my original question.

So create your model from data through 2008, then test your model on 2009/2010 lines.

[podonne]

So create your model from data through 2008, then test your model on 2009/2010 lines.

Right, so I pick a book as the source of my lines (Pinny) and the model is positive, my question is, can I immediately use that model at another book if I don't have access to Pinny?

I'm hoping the answer is not that I have to train, test and use a seperate model for each book, since that is alot of work.

[greva]

Shouldn't matter if you shop for the best available price/line

[Juret]

test it vs the whole market in the first place

[Inspirited]

How would you split only 5 seasons of data between a model set and test set? Would you take games up to a certain point in chronological order to create the model and test against the game remaining or would you randomly select games for the model set and then test against the games not selected?

[podonne]

Shouldn't matter if you shop for the best available price/line

That was my first thought, but it could happen that Pinny is the book with the best line, so in that case I have a model I trained on Pinny but I have to take a worse line, and I don;t know how that efects the performance of the model.

Its possible that I'm overeacting to the situation. I'm hoping that someone would respond with "the odds at all the big books are really close and they move in tandem, so as long as you're dealing with big books, you can train on one and use on another without giving up much EV."

But no one said that, so now I'm worried about it.

[podonne]

test it vs the whole market in the first place

Well, you can't, you've got to pick one line\cost per game. One line\cost means you have to decide where to get the specific line\cost from. Either you pick a book and use it's lines, or you look at all the books' lines and use the best one, the worst one, or an average. Either way, you have to pick a single line, you can't test against "the whole market".

[podonne]

How would you split only 5 seasons of data between a model set and test set? Would you take games up to a certain point in chronological order to create the model and test against the game remaining or would you randomly select games for the model set and then test against the games not selected?

Good question, depends on whether the order of the games matters.

If you were doing a regression model, randomly sampling the data into would be your best bet, since the order doesn't matter so much.

I use 2 years of training and 3 years of testing, but that's because alot of what I do is time-series dependent and deals with looking at trends during the prior x periods and applying it to the next x periods. So the order of bets is very important to my process.