Anomaly detection in time series with the Prophet model

Insofar human activities constitute a major facet among all possible aspects that describe the world, they represent a valuable source of data for statisticians. Data coming from anthropic behaviours are often influenced by temporal factors, such as the month of the year, the day of the week or the presence of holidays and thereby tend to show periodic structures, sometimes along with underlying trends. However, observations that deviate away from the expected pattern occur more frequently than one may think. In this regard, a relevant task for the analyst is to detect anomalies and possibly identify the origins of these alienated points. For this purpose, the Prophet model, a Bayesian mixed linear-nonlinear model designed for time series forecasting, comes to the rescue. We conduct a practical analysis using Prophet on a real-world dataset. Additionally, we explore several methodologies to evaluate the anomaly score.