The header image is my recent Time-Series anomaly (negative) detection project. It is univariate. Green is predicted values, blue is raw data, and red triangles are predicted anomalous point. The model is ARIMA 317 011, derived by following guidelines in Forecasting: principles and practice. It was modeled using latest 180 days of data but seemed to fit the overall trend. The model is refitted every day. Whenever there’s a numerical error, auto-arima would be used instead.
There are a few more TS to model and I am thinking of trying out VARS.