About Facebook Prophet

Prophet

Prophet is a open source forecasting tool for time series data. It is implemented by facebook's research team. The motivation behind building this tool is help people with minimum knowledge produce quality forecasts. Usually, there is a trade off between quality forecasts and the level of knowledge an analyst should posses to tweak the model for performance. This tool tries to increase the performance of the model even for an immature analyst since it uses human interpretable parameters and hyperparameters.

This tool has its boundaries. This tool is built to suite for particular occasions. The model is best suitable when the data has the following characteristics:

• the historical data must be at least an year.
• strong seasonality.
• in case of missing values or outliers.
• trends that are non-linear growth curves.

Why Prophet?

Prophet can be a very efficient tools in the above circumstances. It is also the tools used in Facebook in production. Few answers to why prophet are,

• Very effective for few situations.
• The interface is really simple to use.
• Available in both R and Python.
• The results are produced fast (since they use stan models).
• Resistant to outliers.
• Can deal with missing values.
• Non expert interpretable hyperparameters.

Interface

The data structure used in Prophet is dataframe. The input data is a dataframe with just two columns. The first column named 'ds' contains the series of dates in date or datetime. The second column named 'y' contains the values we intend to predict over time. It should be numeric. Remember, the names of the columns need to be only 'ds' and 'y', else it would throw an error. That's it done! The underlying model is generalized additive model. The calculation is done on stan which is the reason for quick results.