Facebook's core data science team has developed and released an automatic tool for large-scale forecasting. It does a particularly great job with highly seasonal time series, and series with complex trends. Prophet's internal algorithm for detecting trend breakpoints is particularly interesting. Even though it works with any periodicity, it works best with daily data.
Although the algorithm can be used with little to no time series knowledge, experienced users can tweak many of its parameters.
According to its documentation, it has four components:
- Trend detection via a piecewise linear trend, or nonlinear growth curves. This is relevant because some of the time series tasks that Facebook has encountered are phenomena where the data reaches saturation.
- A yearly seasonal component.
- A monthly seasonal component.
- A list...
