Time Series Forecasting with FB Prophet

Recently, Facebook released an internal package for time series analysis called Prophet for R and Python.  It uses an additive model where non-linear trends are fit with yearly, weekly and daily seasonality and permits holidays to be modelled and included or excluded.  The linked page above includes a background white paper and instructions for installation in both languages.

The package offers the advantages of both a relatively simple interface out-of-the-box, and an ability to accommodate missing data, outliers and the like.  It's also tweakable by those with domain knowledge.

I wanted to take a look and understand a little more about how it works. I've used a toy dataset in this case on monthly shampoo sales over 3 years as it's more intuitive for understanding patterns.  I hope to repeat this exercise with a larger N dataset in the next few weeks.

The below embedded Jupyter Notebook runs through the analysis.