DeepAR: Probabilistic Forecasting with Autoregressive Recurrent Networks

DeepAR is a Julia module designed for probabilistic forecasting using autoregressive recurrent networks, based on the concept introduced in the paper DeepAR: Probabilistic Forecasting with Autoregressive Recurrent Networks.

Installation

Currently, DeepAR is not registered in Julia's General registry. To install, clone the repository directly.

Features

DeepAR allows the training and use of deep learning models for time series forecasting. Key features include:

• Time series forecasting using autoregressive models.
• Training and prediction with custom hyperparameters.

Usage

Importing the Module

To use DeepAR, first import the module along with required dependencies:

using DeepAR

Hyperparameters

Define the hyperparameters for the DeepAR model using DeepARParams:

• η::Float64: Learning rate for model training. Default is 1e-2.
• n_mean::Int: Number of samples used for computing the predictive mean. Default is 100.

Training

Train a DeepAR model with the train_DeepAR function:

• Parameters:
  • model: The DeepAR model.
  • loaderXtrain: DataLoader with input sequences.
  • loaderYtrain: DataLoader with target sequences.
  • hparams: DeepARParams instance.
• Returns a vector of loss values.

Forecasting

Generate forecasts using the forecasting_DeepAR function:

• Parameters:
  • model: Trained DeepAR model.
  • ts: Time series data.
  • t₀: Starting time step for forecasting.
  • τ: Number of time steps to forecast.
  • n_samples: Number of samples per forecast (default: 100).
• Returns a vector of forecasted values.

Example

  using DeepAR

  model = Chain(
      RNN(1 => 10, relu), RNN(10 => 10, relu), Dense(10 => 16, relu), Dense(16 => 2, identity)
  )

  train_data = Float32.(randn(1, 1001))
  deepAR_params = DeepARParams(; η=0.01, epochs=100, n_mean=100)

  loaderXtrain = DataLoader(train_data[1:end-1], batchsize=1000)
  loaderYtrain = DataLoader(train_data[2:end], batchsize=1000)

  # Train the model
  loss = train_DeepAR(model, loaderXtrain, loaderYtrain, deepAR_params)

Contributing

Contributions to the DeepAR module are welcome. Please submit issues and pull requests on the repository.

License

DeepAR is licensed under MIT License. Please check the repository for more details.

---

For more detailed information and updates, refer to the DeepAR GitHub repository.