Setup

Clone this repo and enter the directory:

$ git clone https://github.com/dashaub/m4.git
$ cd m4

Install docker. On Windows/OSX, you may need to modify preferences of the Docker daemon to utilize more than 2 CPU cores and allow up to 8GB RAM usage.

Build the docker image:

docker build .  --compress -t m4:5.0 --build-arg BUILD_DATE=2018-05-08

Note that the BUILD_DATE should not be modified if a reproducible build is desired since this pins the versions of the packages installed to those available at this date.

Forecasts

Creating the forecasts is computationally expensive and could take several weeks. Run the image:

docker run -it --name forecastHybrid m4:5.0 /root/m4/forecastM4.sh

Running the job like this will forecast each series serially and can take 2-3 weeks on a fast CPU. Instead you may wish to launch several individual forecasting processes simultaneously by launching a container for a single series.

To produce forecasts for only a single series (e.g. Daily), pass the series name as an argument. It can also be especially useful to change the container name to something descriptive if you wish to launch several containers simultaneously.

docker run -it --name forecastHybridDaily m4:5.0 /root/m4/forecastM4.sh Daily

If the process completes successfully, you will receive a message in the console:

Point forecasts and prediction intervals created successfully!
Pressing enter will exit and erase all results.

Before exiting, you will wish to copy the results from the Docker container to your filesystem. The forecasts will appear inside the docker container in the mean/ directory; the upper/lower 95% prediction intervals will appear in upper/ and lower/, respectively. In another terminal (not inside the docker container), use docker cp to copy the forecasts from the container to your filesystem. The mean/, upper/', and lower/` directories of your local filesystem will now contain the forecasts and prediction intervals, and you can safely exit and destroy the container.

To copy all of these results back to the host filesystem, run the following commands (where forecastHybrid is the name of the docker continer):

$ docker cp forecastHybrid:/root/m4/mean .
$ docker cp forecastHybrid:/root/m4/upper .
$ docker cp forecastHybrid:/root/m4/lower .

The forecasts can be combined into a single file:

cat upper/*.csv > upper_submission.csv
cat lower/*.csv > lower_submission.csv
cat mean/*.csv > mean_submission.csv

To prevent exponential upper-bound growth in the prediction intervals, it can be useful to run cleanUpper.R to adjust series that exhibit this explosive growth.