2
$\begingroup$

Finding the best distribution that fits a data sample seems to be a though problem since there is no cookie-cutter solution.

Although automated fit softwares exist, it remains suboptimal to use a list of well-known distribution.

I heard about the function fitdistr() in the MASS package but some people argue that there is no best way to fit data properly with this kind of method as well as using likelihood methods for all data sample.

My questions are: - Do those packages or automated fit softwares work for business analysis where strict accuracy doesn't matter ? - What is the investigation method to find the distribution of a data sample without using a predefined list of distribution ? Is it based on guess ? Experience ? Methodology ? If it is based on methodology, what is it ?

I'm not a scientist so I may have difficulties to understand the concepts. I'm in business and don't need to be 100% accurate in my analysis. I just want to find a way to find the distribution of my data without force-fitting them with a normal distribution or that kind of wrong analysis. I'm using R.

$\endgroup$
  • 2
    $\begingroup$ The problem with "good enough" is that mileage varies. Expectations minus reality equals disappointment. What do you expect from your fit? Every approach has problems - every one of them breaks down somewhere. I think you need to be much more clear about how much wrong you can tolerate, and what your expectations are before a good answer can be provided. $\endgroup$ – EngrStudent Aug 6 '15 at 12:14
  • $\begingroup$ It is difficult to say exactly how much wrong I can tolerate. I would say, unlike medicine or industrial manufacturing, I don't need to be 99% or 95% sure I modeled my data the right way. If I'm right 80%-90% of the time, it's good. The pattern I want to discover in the data is more a trend than specific numbers. $\endgroup$ – Synleb Aug 6 '15 at 12:22
  • 1
    $\begingroup$ Glen_b has just written a response to another question with advice that appears to be relevant here: see stats.stackexchange.com/questions/164925/…. We would need to explore the same issues with you, starting with why do you want to fit a distribution to your data? (Distribution fitting scarcely constitutes finding "trends" or "patterns"--and if either is your aim, there are far better procedures to employ.) $\endgroup$ – whuber Aug 6 '15 at 12:28
  • 1
    $\begingroup$ Whuber: my purpose is to do predictive modeling. So I guess finding the distribution of the data allows you to find the probabilities associated with a possible outcome. Building models also often require you to specify the distribution of the data to find the right parameter. And finally, knowing the distribution of the data allows you to know which machine learning algorithm to use. $\endgroup$ – Synleb Aug 6 '15 at 13:10

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.