I am new to the anomaly detection world and am dealing with a project to detect real-time anomalies for a time-series in a fraud detection schema. I read the answer by Rob Hyndman here and like the simplicity of it. However, I have to concerns. 1) I read that STL is not scalable for real-time analytics (see here for example). What is the best decomposition choice here beside fourier. 2) How can I use Rob Hyndman for a new/real-time event by using this approach? For example, let say we are time
t and a new event happens (like a credit card activity). How can I detect if this is an anomaly or not given the dynamic and non-stationary nature of my time-series?