DeepAR is for forecasting one-dimensional time series data. It is able to find the frequencies and seasonalities.
Input
- jsonl (timestamp, dynamic or categorical features, target)
Training
- Use all the data when possible
- Inference on a small window in the future
- Train on many series if possible
Hyperparameters
- Context length
- Epochs
- Batch size
- Learning rate
- Hidden dimension/cells
CPU can be sufficient, especially for inference.