I think that the effect is from the fact that you're sampling from all possible outcome history paths. To realise the positive expected value, your sample needs

I think that the effect is from the fact that you're sampling from all possible outcome history paths. To realise the positive expected value, your sample needs to include some of the increasingly rare possibilities where you gain enormously. This means that if your population is large compared to the number of branches (2^rounds) then you probably do sample that and the overall gain is positive. If your number of branches is large compared to your population, you probably don't sample the branches where you make lots of money, and your overall gain is negative.