3 ms·

You're not wrong; normal distributions are OK but not ideal for estimating durations -- it can lead to weird artifacts if you don't restrict yourself to just th

by diiq 5y ago

You're not wrong; normal distributions are OK but not ideal for estimating durations -- it can lead to weird artifacts if you don't restrict yourself to just the mean. (https://blog.vistimo.com/post/169546687605/estimation-math https://blog.vistimo.com/post/169546687605/estimation-math) And once you're committed to using monte carlo, you're free to include one-off events -- sick days, vacations, social loafing, etc, that are really hard to model analytically, since they are roughly uniform in time -- the longer the project runs, the more days off make the project run longer.

Grustaf 5y ago
I suppose you could model those one-off events with their own normal distributions (with 0 variance), rather than modifying the main distribution? But sure, Monte Carlo is more flexible, and not really harder to implement I suppose, since you're already on a computer!.
iainmerrick 5y ago
That blog post is very clear, nice work! As another alternative, could you convolve the probability distributions numerically? I’m not sure of the trade-offs between that approach and Monte Carlo, but if it’s something you’ve considered I’d be interested to hear your thoughts.
- diiq 5y ago
  Once I took a few steps down the simulation path, it seemed so easy and powerful that I didn't look back. There is some computational cost, but for me it's more than made up for by how easy it is to, for instance, say "I'd like to model sick days"; * let the government tell me how often people are sick (https://www.bls.gov/cps/cpsaat47.htm https://www.bls.gov/cps/cpsaat47.htm, see "Computer and # mathematical occupations", absence rate, due to illness or injury), * make my simulated workers take days off at that rate. And everything else, all the statistics etc, can stay exactly the same. And I can have a lot of trust that it's producing sane results because I can look at a run, and say yes, that's a plausible number of sick days. Monte carlo requires you to be clever once, setting up the simulation, and then you almost never have to be clever later when you want to introduce new twists. Any other statistical model you have to keep being cleverer for every additional complication.

by diiq 5y ago

Grustaf 5y ago

I suppose you could model those one-off events with their own normal distributions (with 0 variance), rather than modifying the main distribution? But sure, Monte Carlo is more flexible, and not really harder to implement I suppose, since you're already on a computer!.

iainmerrick 5y ago

That blog post is very clear, nice work! As another alternative, could you convolve the probability distributions numerically? I’m not sure of the trade-offs between that approach and Monte Carlo, but if it’s something you’ve considered I’d be interested to hear your thoughts.

diiq 5y ago
Once I took a few steps down the simulation path, it seemed so easy and powerful that I didn't look back. There is some computational cost, but for me it's more than made up for by how easy it is to, for instance, say "I'd like to model sick days"; * let the government tell me how often people are sick (https://www.bls.gov/cps/cpsaat47.htm https://www.bls.gov/cps/cpsaat47.htm, see "Computer and # mathematical occupations", absence rate, due to illness or injury), * make my simulated workers take days off at that rate. And everything else, all the statistics etc, can stay exactly the same. And I can have a lot of trust that it's producing sane results because I can look at a run, and say yes, that's a plausible number of sick days. Monte carlo requires you to be clever once, setting up the simulation, and then you almost never have to be clever later when you want to introduce new twists. Any other statistical model you have to keep being cleverer for every additional complication.