An alternative approach to time series decomposition is to use an additive model, in which a time series is represented as a combination of patterns on different time scales (daily, weekly, monthly, yearly, and so on) together with the overall trend. Facebook's Prophet does exactly that, along with more advanced functionalities such as accounting for changepoints (rapid changes in behavior), holidays, and much more. A practical benefit of using this library is that we are able to forecast future values of the time series, along with a confidence interval indicating the level of uncertainty.

In this recipe, we will try fitting Prophet's additive model to daily gold prices from 2000-2004 and predicting the prices over 2005.