0
$\begingroup$

What options exist in order to forecast when next observation will be an outlier in a time series? Initially, I thought to train a simple forecasting model, which turned out to decently predict the normal values, but not predict the outliers, which I am more interested in for business purposes. Note that I want to predict when next observation will be anomalous, not detect previously anomalous observation. enter image description here

I cant find any literature about this topic online. Am I just not capturing some feature that might prove a better predictor or is there some methodology that may help?

$\endgroup$
2
  • 1
    $\begingroup$ If it's predictable, it is, by definition, not an anomaly. $\endgroup$ Jul 29, 2022 at 6:58
  • $\begingroup$ Do you want to detect the anomalous points? Or to predict their value (hard/impossible by definition as Rob says, because if it is possible, then they are not really anomalies, under that model). $\endgroup$
    – Jon Nordby
    Jul 31, 2022 at 16:55

2 Answers 2

1
$\begingroup$

One way to go about it:

You identify the positions in the time series where the outliers have occurred and predict a yes / no (binary classification) for the future i.e. whether a particular observation in the future will be an outlier or not. The feature engineering involved with this approach might be slightly tricky and will be heavily dependent on what data you have.

$\endgroup$
2
  • 1
    $\begingroup$ Thanks., I may try this option, as the other answers deal with removing the outliers, which is the opposite of what I want. $\endgroup$
    – M Reyes
    Jul 29, 2022 at 20:31
  • $\begingroup$ Happy to be of help. $\endgroup$ Apr 17 at 14:38
1
$\begingroup$

handling outliers can be managed by : 1. ignoring them 2. scaling the data reducing the affect of the outliers on your data 3. using a higher level curve fitting algorithm 4. adding more categories and aggregating the data into an average.

If the data was not normal distributed, you would apply a feature transformation

In such cases, the extreme values could be identified and removed in order to make the distribution more Gaussian. These extreme values are often called outliers

Taking the square root and the logarithm of the observation in order to make the distribution normal belongs to a class of transforms called power transforms.

$\endgroup$

Your Answer

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

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