@spec build_component(Axon.t() | nil, map()) :: Axon.t()

Builds the AR component layer(s).

Supports both linear AR (single dense layer) and deep AR-Net (multiple hidden layers with ReLU activation followed by linear output).

Parameters

• input - Axon input node from build_network_input/1.
• config - Model configuration map.

Returns

An Axon layer when AR is enabled, Axon.constant(0) otherwise.

@spec build_input(map(), list(), non_neg_integer()) :: Nx.Tensor.t()

Builds AR input tensor for prediction given training data and prediction dates.

For each prediction date, looks up the previous lags values from the training data to use as AR features. Returns zeros for dates that don't have enough history.

Parameters

• training_data - Map with :dates (list of dates) and :y_normalized (list of values)
• prediction_dates - List of dates to build AR inputs for
• lags - Number of lagged values to include

Returns

A tensor of shape {n_predictions, lags} with AR features.

@spec build_network_input(map()) :: Axon.t() | nil

Creates the Axon input node for the AR component.

Parameters

• config - Model configuration map with :ar key.

Returns

An Axon input node when AR is enabled, nil otherwise.

@spec create_lagged_inputs(Nx.Tensor.t(), non_neg_integer()) ::
  {Nx.Tensor.t(), Nx.Tensor.t()}

Creates lagged input features and corresponding targets for AR training.

Given a time series y and number of lags, creates sliding windows where each window contains lags consecutive values, and the target is the next value.

Parameters

• y - A 1D tensor of time series values.
• lags - Number of lagged values to use as features.

Returns

A tuple {lagged, targets} where:

• lagged - Tensor of shape {n_samples, lags} with lagged features
• targets - Tensor of shape {n_samples, 1} with target values

Examples

iex> y = Nx.tensor([1.0, 2.0, 3.0, 4.0, 5.0])
iex> {lagged, targets} = Soothsayer.AR.create_lagged_inputs(y, 3)
iex> Nx.shape(lagged)
{2, 3}

@spec get_weights(Soothsayer.Model.t()) :: %{
  required(String.t()) => %{kernel: Nx.Tensor.t(), bias: Nx.Tensor.t()}
}

Extracts raw AR layer weights from a fitted model.

For linear AR models, returns the output layer weights. For deep AR-Net models, returns all layer weights including hidden layers.

Parameters

• model - A fitted Soothsayer.Model struct with AR enabled.

Returns

A map of layer names to weight structs containing :kernel and :bias tensors.

Examples

iex> weights = Soothsayer.AR.get_weights(fitted_model)
%{"ar_dense_out" => %{kernel: #Nx.Tensor<...>, bias: #Nx.Tensor<...>}}