0
$\begingroup$

I would like seek advice on how to build an efficient approach to identify outliers in a financial series taking into account also related series. For example, let's assume the there is a very important growth in yields of country A, so that the last observation of my series is the highest ever recorded. With a Local Outlier Factor (LOF) or Isolation Forest algorithms, I may probably find that my n observation is flagged as outlier. However, let's say that also country B and C show the same growth, although with different levels, so that, by visual inspection, I would consider that my n observation in country A series is not an outlier as it fits the same trend (growth of yields as in all the countries).

Which kind of (mix of) algorithm/approches would you suggest in this case? Thank you.

$\endgroup$

1 Answer 1

0
$\begingroup$

What are you going to do with the outliers? An answer to that question determines how best to detect them.

A week ago the UK pension funds nearly went bankrupt. This was an "outlier". It was also the most important financial event so far this century. In some but not all scientific contexts, "ignoring outliers" is reasonable. In finance, it is the outliers that matter.

$\endgroup$
1
  • $\begingroup$ Thank you. I agree but I would like to minimise false positives (wrong outliers). For instance, if all countries show an increase in yields, then I would expect that these latest high values in the series are to be considered as inliers, not outliers, as the fit into the same pattern. From visual inspection, you would also agree. My question then is how to take into account related time series (i.e. similar countries) in order to assess the likelihood of these apparent outlier values to be treated as such (by adding some sort of "cross-sectional" / cross countries controls) $\endgroup$ Commented Oct 8, 2022 at 9:47

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.