1

I have to build model to predict churn and when reading related work on the internet I have realized that in most of the cases the AUC is used as a metric to compare different models. That's surprising to me because, in general I have used different metrics to access the performance of various classifiers. Is there something that makes churn prediction so unique that applying AUC is a better idea than using other accuracy metrics? Or I'm simply missing something?

ononono
  • 75
  • 3
  • 1
    Interesting question and a +1. What metrics do you prefer? – Dave Aug 24 '21 at 14:37
  • I would know a perfect answer for Species Distribution Modelling (predicting probability of occurrence), this is my area, but that answer probably wouldn't apply to predicting churn.... – Tomas Aug 24 '21 at 14:58
  • 1
    I'm with @Dave: we can probably give you better answers if we know what metrics you would use instead. In the meantime, [What does it mean that AUC is a semi-proper scoring rule?](https://stats.stackexchange.com/q/339919/1352) may be enlightening. – Stephan Kolassa Aug 24 '21 at 15:03
  • 2
    Normally there is a business case where you eg want to prioritise those most likely to churn (ie ranking). personally I would go for logloss where you aim to get accurate probability, so that you can calculate expected loss from churning ... – seanv507 Aug 24 '21 at 16:36
  • Thank you for your comments. In general I prefer using metrics like precision, recall balanced accuracy or f1 score. In this specific case the main objective (as it is a real business case) is to identify the customers that will most likely churn. – ononono Aug 25 '21 at 10:58
  • Probability exactly measures the colloquial “likelihood” of an event. – Dave Aug 25 '21 at 11:19

0 Answers0