From Facebook's github repository

This branch explored the time series reconstruction using a prediction tool developed by Facebook, based on a variant of STL decomposition.

Prophet paper: Sean J. Taylor, Benjamin Letham (2018) Forecasting at scale. The American Statistician 72(1):37-45

This repository has been tested on Ubuntu 16.04 LTS.

Prophet and his dependencies are available on Windows and OSX/MacOs too. Refer to Prophet's github page for detailed installing instructions.

Make sure compilers (gcc, g++, build-essential) and Python development tools (python-dev, python3-dev) are installed. PyStan and Prophet packages need to be compiled and require the appropriate build toolchain.

Be sure to have your dataset.csv into the ./input folder, run your server and you're ready to go.

Recommended: create a dedicate virtual environment for this repository.

Install the required packages using pip, there's a specific order to observe while installing modules.

    pip install -r requirements.txt

Be sure to have your csv files in the input folder.

Scripts available

• prophet_demo.py takes as input a complete dataset, erases a significative portion of values, then proceeds to reconstruct them and plot the result.

• crossvalidation.py scans the input folder and collect all the csv inside. Then proceeds to erase a portion of values (16/4/2/1 weeks long, see global parameters) and reconstruct that erased values for all datasets. Produces as output a report with RMSE, integral of values and standard deviation.

• plot.py an useful pre-baked script to visualize Prophet's output dataframe.