DeepAR - MXNet

DeepAR is an autoregressive recurrent neural network that forecasts probability distributions for the next forecast horizon values given the preceding context length values. It also uses lagged values and time features, automatically computed based on the selected time frequency. This implementation relies on MXNet.

Training parameters

Batch size

The size of the batches to be used for training and prediction.

Epochs

Use default nb. of batches per epoch

Default value is the number of batches for which on average each time step appears in one sample per epoch.

Number of batches per epoch

Random state

Using a fixed random seed allows for reproducible result.

Model parameters

Use default context length

Set the context length to the model's default value (context length = forecast horizon).

Output distribution

Try Student’s t distribution Try Gaussian distribution Try Negative binomial distribution

Distribution to use to evaluate observations and sample predictions.

Encode identifiers

Encode time series identifiers and use them as external features.

Cell type

Try Long short-term memory (LSTM) Try Gated recurrent unit (GRU)

Dropout cell type

Try ZoneoutCell Try RNNZoneoutCell Try VariationalDropoutCell Try VariationalZoneoutCell

Scale target values

For each time series, rescale the target by its average absolute value.

Minimum scale

Default scale (used if the time series has only zeroes).

Number of parallel samples

Number of evaluation samples per time series, to increase parallelism during inference.