AutoGluon Time Series FAQ#

What forecasting tasks can AutoGluon be used for?#

AutoGluon can generate probabilistic multi-step-ahead forecasts for one or multiple univariate time series. For example, you can use AutoGluon to forecast daily sales of multiple products over the next month.

AutoGluon also supports additional information, such as time-independent static features (e.g., location of the store) and time-dependent covariates (e.g., price of the product each day). See the In Depth Tutorial for more details.

Currently, AutoGluon does not support features such as hierarchical forecasting and forecast explainability.

How can I get the most accurate forecasts?#

To maximize the forecast accuracy, set the predictor.fit() argument presets="best_quality" and provide a high time_limit.

Are there any restrictions on the data that I can pass to TimeSeriesPredictor?#

See “What data format is expected by TimeSeriesPredictor?” in the In Depth Tutorial

Can I use GPUs for model training?#

Yes! All deep learning models used by autogluon.timeseries support GPU training. The models will be automatically trained on a GPU if (1) your machine has a GPU and (2) you installed a PyTorch version with CUDA support. Multi-GPU training is not yet supported.

What machine is best for running AutoGluon TimeSeries?#

AutoGluon can be run on any machine including your laptop. It is not necessary to use a GPU to train TimeSeriesPredictor, so CPU machines are fine. Using a machine with more CPU cores and more RAM will lead to faster training and allow you to quickly generate forecasts for larger datasets. For example if using AWS instances for Tabular: we recommend M6 instances instances, where a m6i.24xlarge machine should be able to handle most datasets.

Issues not addressed here#

First search if your issue is addressed in the tutorials, documentation, or Github issues (search both Closed and Open issues). If it is not there, please open a new Github Issue and clearly state your issue and clarify how it relates to the module.

If you have a bug, please include: your code (ideally set verbosity=4 which will print out more details), the output printed during the code execution, and information about your operating system, Python version, and installed packages (output of pip freeze). Many user issues stem from incorrectly formatted data, so please describe your data as clearly as possible.