> Looks like our anomaly detector is doing a decent job.
I don't agree. Calling ~5% of your dataset anomalies purely by construction always seemed like a crappy way of doing things to me. Look At those stars towards the end of the curve, they don't look like anomalies to me.
It seems like a lousy job. When you put up code samples make sure that code is working. I found at least 2 instances where the variable was missing by just reading the article. Please make sure when you share code, that it works. Also, a shared iPython notebook on github/juypter network would do better.
Thanks for the article.