Nixtla's neuralforecast is a python library for time series forecasting with deep learning models. This app provides AutoRNN functionality, implemented using PyTorch, PyTorchLightning and Ray. This package automatically tests different settings, identifying over multiple iterations which parameters perform the best. The results for the champion model are then provided. If trained using CPU, then a maximum of 32 time series will be processed, while if trained using GPU a maximum of 128 will be processed instead. Models are trained using all series, meaning that with more series included, performance is likely to improve. Models can be saved, and then applied to similar time series, or trained further.

Please provide data in three columns, in the following order:

• unique_id for each time series (up to 16 time series can be provided in one file, and trained simultaneously--if unique_id is not included it is assumed that the file contains only one time series)
• date-time column can be in any format (date, time, number, other), but must be in ascending order.
• value. The value of the time series at each time