I was asked this question in a recent interview for the position of a Data Scientist: Which probability distribution will you use to model outliers ?

I told him outliers are like rare events which can be modelled by a Poisson distribution. I pretty sure I'm wrong and the interviewer seemed to think the same. But I don't know the answer to this.

Please advise.

  • 1
    $\begingroup$ Based on standard deviations of a standard normal distribution it is well possible to describe outliers (usually something beyond 3 standard deviations) en.wikipedia.org/wiki/Standard_score. Opinion: There is no „one true“ answer for the question described by you, its just „testing“. Speaking about right-skewed poisson will be okay, I guess. $\endgroup$
    – Peter
    Nov 7, 2020 at 11:29

1 Answer 1


I think the answer is Gaussian distribution. This is a famous approach that is used in Anomaly Detection. What you do is to fit your feature to the Gaussian distribution and the samples which have the probability below the specific threshold are labeled as an outlier.

Quoting from the paper Modeling Outlier Score Distributions:

Many existing unsupervised outlier detection algorithms calculate some kind of score per data object which serves as a measure of the degree of outlier. Scores are used in ranking data points such that the top n points are considered as outliers. For example, the statistical-based approach proposed in [4], uses a Gaussian mixture model to represent normal behavior and each datum is given a score on the basis of changes in the model. A high score indicates a high possibility of being an outlier.

Example of usage - 1.

Example of usage - 2


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.