Consider a baseball model that requires the starting lineups to be released before making a prediction.
Now say we have two games at 6.00pm and another two games at 7.30pm.
At 6 we have our predictions for the first games but the starting lineups have not been released for the 7.30 games so we do not know our edge for these. It is also the case that the 6pm games will not have finished by 7.30, so we will not know the outcome of our initial bets when it comes to betting on the 7.30 games.
My question then is how do we use Kelly to bet optimally on all these games?
A couple of iffy solutions are:
1) Assume we have a constant edge of x% (calculated from past performance) for games that we do not yet have the starting lineups for and we can then plug this into our initial 6pm Kelly calculation.
2) A better solution would be to guess what the starting lineups are going to be if they have not yet been released. However this relies on being able to guess well and is more cumbersome if you want the model to be fully automated as guessing is a very human thing.
N.B. The problem gets even trickier if we have a chain of games say 6pm, 7.30pm and 9pm such that 6pm games haven't finished by 7.30, 7.30 haven't finished by 9 but 6.30 games HAVE finished by 9.
Now say we have two games at 6.00pm and another two games at 7.30pm.
At 6 we have our predictions for the first games but the starting lineups have not been released for the 7.30 games so we do not know our edge for these. It is also the case that the 6pm games will not have finished by 7.30, so we will not know the outcome of our initial bets when it comes to betting on the 7.30 games.
My question then is how do we use Kelly to bet optimally on all these games?
A couple of iffy solutions are:
1) Assume we have a constant edge of x% (calculated from past performance) for games that we do not yet have the starting lineups for and we can then plug this into our initial 6pm Kelly calculation.
2) A better solution would be to guess what the starting lineups are going to be if they have not yet been released. However this relies on being able to guess well and is more cumbersome if you want the model to be fully automated as guessing is a very human thing.
N.B. The problem gets even trickier if we have a chain of games say 6pm, 7.30pm and 9pm such that 6pm games haven't finished by 7.30, 7.30 haven't finished by 9 but 6.30 games HAVE finished by 9.
