Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

Depending on how much precision you'd like for your answer, Monte Carlo methods can converge unbearably slowly, since the standard deviation is proportional to 1/sqrt(N) as you add more samples. I recall once having to run 100 billion simulations to get the mean of a certain quantity to within a few decimal places. I'd take an infinite series over that ant day of the week (well, as long as it's not something as slow as the Leibniz series for pi/4).

Of course, such precision relative to the standard deviation is rarely necessary for most practical situations.



As a friend in grad school used to say... "An infinitely long Monte Carlo simulation recapitulates the underlying probability distribution perfectly, but we don't want to wait that long.":


Yeah, it’s the practicality. Most business decisions don’t need more than three sigfig [source?]. We only tend to think about precision on things that already have it.


Unless you are trying to prove the existence of a subatomic particle, I am not sure what business decisions have anything close to that level of certainty.


Kinda depends on what you mean by "business decisions". Repetition can bring in a high need for confidence. You don't want to have your life depend on a piece of equipment that has a one in a million chance of failure if it gets tested a thousand times an hour. Pricing things can also need a lot of certainty, like in HFT. But most things aren't either of those cases and people are commonly off by an order of magnitude.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: