| title | tags | authors | affiliations | date | bibliography | ||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
AtsPy: Automated Time Series Forecasting in Python |
|
|
|
22 April 2020 |
paper.bib |
AtsPy, an open source automated time series framework is developed as a working prototype to showcase the ability of state of the art univariate time series methods.
A Python-centric view on the recent growth of time series tools shows the development of the Prophet by Facebook[@Taylor:2017], the GluonTS[@Alexandrov:2019] toolkit by Amazon, and the ForecastTCN algorithm by Microsoft. Among others, these tools have put forecasting methods in the hands of the everyday user. We have also seen contributions from academia and independent developers with algorithms and packages like N-Beats, Auto-Arima, and TBATS[@Hyndman:2011],. The majority of these have been implemented in AtsPy which is hosted on GitHub.The recent surge in automated time series methods is the direct result of new algorithms (TBATS), procedures (Prophet), and frameworks (GluonTS). In the following section we will seek to understand how these methods have led to the automation of time series forecasting and also discuss how existing method can be used to automate predictions.
Can be seen as a univariate instantiation of GluonTS with an emphasis on model diversity. AtsPy is built on top of Auto-Arima, TBATS, Prophet, and GluonTS. It is an extremely fast method to test which model best fits your data. AtsPy's final innovation is an ensemble time series protocol developed with the LightGBM flavour Gradient Boosting Machine and extracted time series features.