estimates
according to some back of the envelope calculations, I need at least 40k € for a year to get anything done (this includes two persons working together, me and somebody else of equal skill level, and an office, no equipment) not really based on a Gantt chart, but whatever can be done must fit into this amount
re: estimates
@amiloradovsky lognormal from percentile intervals for the rescue!
https://www.squiggle-language.com/docs/Guides/DistributionCreation#to
@amiloradovsky It is easy enough to run guesstimates:
But, as they say in docs, you gonna need a bit of training to get overconfidence under control.
The good news here, is the `x to y` model is less underconfidence-sensitive so you can set "protect thy ass" upper ranges and still get an realistic, actionable estimate.
once upon a time I also wrote a program for least-square optimization (ERP-like)
I don't know anything about things like maximum likelihood estimation or chi-squared test, or how to apply anything I know from PT in practice
generally sophisticated statistical estimates feel like free money -- trying to get more precision from the same information -- well you'd need actual data for it to be valuable, and the estimates are just that: adding sophisticated theory on top of guesses doesn't make the predictions any more accurate
I might need the tools like Squiggle at some point, but I'd prefer them to be on-premise and not a webapp, and also if learning them I need to start from the very beginning, from the very basic practical examples
@amiloradovsky squiggle is just a bunch of js, should work offline.
You don't need to have a statistics insight, just combine those intervals and get the risk profile: given X amount of resources I will fuck that up in Y% of cases.
if I say estimate how long (how many man-hours) a task will take, the only source of information I have is it's structure (WBS) and the sense of all the leaf subtasks being about the same size; I'm not sure how reasoning in terms of distributions may even help
@amiloradovsky those "hours" numbers, are lower/upper bound? Mean/median? Anyway, there's a missing variance information. And usually that's what screws your deadlines. Sufficiently high average numbers aren't actionable.
@amiloradovsky i think using lognormal makes sense given it represents the long tail of "ways this can go wrong" while yielding to central limit nicely.
I just thought that ELI5 of it can be something like "estimating order of magnitude". Something like "weeks" is anywhere from 7 days to a month.
@dpwiz re central limit: this kinda suggests that the distribution for the entire task/project has to be normal, or in this case log-normal Poisson distributions may also play a role here, contributing to the delays associated with failures, either of hardware or say sick leave of the employees
(usually I prefer thinking about mean, because it's easier to compute and may be generalized to higher dimensions, unlike CDFs)
thinking in terms of ranges would be more accurate indeed: say each subtask may take from x to y (hours), where we assume that the chance of it taking less than x or more than y is say <0.05 (or 1/4)
then from this we may compute the range for the whole task/project
naively one might just add the lower and upper bounds separately, but I guess statistically this wouldn't be quite correct
another question is whether assuming the distribution to be log.normal is always adequate