[cyberbabble]

I have a 60 win - 40 loss record in a 50/50 situation. This is good enough to continue following even though the sample size is smallish.

I consult Mr. Binomial and he says I would get 60 win or better about 3% of the time using random picking. Still looks promising.

I am realistic enough to know 60% winners is probably not going to last long term, but I would be happy with say a 55% win rate long term.

Given the sample of 100, can I calculate some sort of confidence number that I am 55% or better long term?

[Wrecktangle]

If you had static distributions to draw from, ordinarily I'd say using a z-score test off a binomial dist would give you some idea. But given that every sport I follow has changed in the last few years (some drastically like the NBA), I'd say no such test exists that will give you any confidence going forward.

Maybe one the Frequentists will chime in here; they usually have some "classical" test or approach that typically fails in the long run.

[cyberbabble]

I was wondering if this could be viewed as a sampling type situation.

Use the famous infinite bag of marbles. I pull out a 100 marbles and get 60 white and 40 black. Can I say that there is a confidence number that the bag has 55% or more white marbles?

Thanks.

[cyberbabble]

Bag of marbles is not really infinite. Just has a million marbles and marbles are drawn one at a time and replaced.

[MonkeyF0cker]

If you had static distributions to draw from, ordinarily I'd say using a z-score test off a binomial dist would give you some idea. But given that every sport I follow has changed in the last few years (some drastically like the NBA), I'd say no such test exists that will give you any confidence going forward.

Maybe one the Frequentists will chime in here; they usually have some "classical" test or approach that typically fails in the long run.

You think binomial z-scores aren't classical/frequentist? LOL.

[Wrecktangle]

You think binomial z-scores aren't classical/frequentist? LOL.

Not everything you guys use is crap.

[cyberbabble]

Model: past statistics + some regression method ---> predict the future.

You are saying modeling is flawed because the statistics are based on a dynamic, on going situation and the model gets farther from reality over time. At some point the model no longer has any value. The modeler doesn't realize this until too late. Obviously a risk when using a model.

What is the way to handle this? Give up on the modeling idea? Pursue scalping, middling, steam chasing, etc? Switch to slot machines?

[hutennis]

I was wondering if this could be viewed as a sampling type situation.

Use the famous infinite bag of marbles. I pull out a 100 marbles and get 60 white and 40 black. Can I say that there is a confidence number that the bag has 55% or more white marbles?

Bag of marbles is not really infinite. Just has a million marbles and marbles are drawn one at a time and replaced.

Thanks.

What gives you the right to assume that marbles are replaced with the same ones you draw out?
Is universe built on the idea of being fair to you specifically?
What if someone (lets call it randomness god) is constantly and secretly changing composition of marbles in a bag?
Like grabbing 100k or so blacks from the bottom before your every draw and without letting you know replacing them with whites. And he does it every time randomly.
How many draws and what results do you need to have confidence you are looking for?

[cyberbabble]

The original question with infinite marbles was bad. You forgot to point out that if the bag has an infinite number of white and an infinite number of black then the number of white = number of black. The ratio of white to black has no real meaning.

If I start worrying about the randomness god fixing NFL games to confuse me, I need a new hobby or maybe a shrink.