[forecast][GluonTS] method in the time-series forecast

IJN-Kasumi 1939–1945
2 min readFeb 14, 2020

--

A boat that sails for living and culture. (Image source: Author)

Background

Support both data researchers and production use cases in the novel design and a fully-managed service, respectively. The further introduction is on the below website from Amazon(R) company. Of course, please obey their policy using this package framework.

Installation

The installation process is handy without too much trouble upon the mentioned website content. However, please consider using the mxnet package in 1.4.1 version from an unexpected issue. The issue log and command are below for your own reference.

issue log: “mxnet.base.MXNetError: vector<T> too long”

issue reference: https://github.com/awslabs/gluon-ts/issues/495

command: pip install — upgrade mxnet==1.4.1 gluonts

Test Example

Let’s check out the example code to make sure the environment ready. Some example files are on the GitHub website for reference. Some files are tested properly.

This article tests the example code from this website with slight modification.

Example code

Example test code from the above website. It is for environment test purposes only. The parameter-setting are the freq, prediction_length, and epochs in the Line21.

Example code. It should be 100% executable

The model training time is 46.57 seconds in the i7–5820K CPU machines. During the model training, all CPU cores are working but no signs on GPU’s work in the resource monitor report. The test result is below with the shades of the forecast possibility and median curve.

Example code result. image source: own environment

Conclusion and future work

This framework is strength in the

  • a. easy installation with stable solvers
  • b. forecast possibility report
  • c. support all CPU cores working, and
  • d. instinctive tuning parameters.

In the future, the trained model explanation and polynomial equation test case should be done to check its accuracy.

--

--