DLinear

chronax.models.DLinear · inherits BaseForecaster

Univariate DLinear forecaster (JAX/Flax-NNX port of neuralforecast.DLinear). Decomposes each input window into trend (edge-padded moving average of size moving_avg_window) and seasonal (residual), applies one linear head to each, and sums (Zeng et al., 2023). Defaults match neuralforecast 3.1.7 (moving_avg_window=25, max_steps=5000, learning_rate=1e-4, identity scaler, MAE loss) so the benchmark harness compares both libraries at native settings.

Maintenance status: Active univariate forecaster. Integrates with the BaseForecaster interface, including conformal prediction intervals via predict(level=...), pickle round-trip, and forecast(fitted=True).

Attributes

Attribute	Value	Description
uses_exog	False	Indicates that this model does not support exogenous variables.
alias	"DLinear"	The default alias for the model instance.

__init__(self, h: int, input_size: int = -1, moving_avg_window: int = 25, max_steps: int = 5000, learning_rate: Union[float, Callable[[int], float]] = 1e-4, windows_batch_size: int = 1024, loss: Union[str, LossFn] = 'mae', scaler: Union[str, Scaler] = 'identity', random_seed: int = 1, alias: str = 'DLinear')

Initializes the DLinear forecaster configuration.

Parameter	Type	Default	Description
h	int	-	Forecast horizon.
input_size	int	-1	If < 1 (default), expands to 3*h.
moving_avg_window	int	25	Size of the edge-padded moving average used for trend decomposition. Must be odd and positive.
max_steps	int	5000	(undocumented)
learning_rate	Union[float, Callable[[int], float]]	1e-4	(undocumented)
windows_batch_size	int	1024	(undocumented)
loss	Union[str, LossFn]	"mae"	(undocumented)
scaler	Union[str, Scaler]	"identity"	"identity" (default, matches neuralforecast) or "robust".
random_seed	int	1	(undocumented)
alias	str	"DLinear"	(undocumented)

fit(self, y: jnp.ndarray, X: jnp.ndarray | None = None) -> Self

Fit on a 1-D series.

Parameters:

Parameter	Type	Default	Description
y	jnp.ndarray	-	1-D series to fit.
X	jnp.ndarray \| None	None	Exogenous variables (not supported).

Returns: Self (the fitted forecaster; sets self.model_). Raises: * NotImplementedError: on exog. * ValueError: if y is not 1-D or shorter than input_size + h. * RuntimeError: on divergence.

predict(self, h: int, X: jnp.ndarray | None = None, level: list[int | float] | None = None) -> dict

Forecast h steps ($1 \le h \le \text{self.h}$). level triggers the inherited conformal path.

Parameters:

Parameter	Type	Default	Description
h	int	-	Forecast horizon (must be $1 \le h \le \text{self.h}$).
X	jnp.ndarray \| None	None	(undocumented)
level	list[int \| float] \| None	None	Prediction interval levels (triggers conformal prediction).

Returns: dict Keys: {"mean": jnp.ndarray}. If level is provided, includes keys for prediction intervals (e.g., lower_XX, upper_XX).

forecast(self, y: jnp.ndarray, h: int, X: jnp.ndarray | None = None, X_future: jnp.ndarray | None = None, level: list[int | float] | None = None, fitted: bool = False) -> dict

Stateless fit-then-predict.

Parameters:

Parameter	Type	Default	Description
y	jnp.ndarray	-	(undocumented)
h	int	-	(undocumented)
X	jnp.ndarray \| None	None	Exogenous variables (not supported).
X_future	jnp.ndarray \| None	None	Exogenous variables (not supported).
level	list[int \| float] \| None	None	(undocumented)
fitted	bool	False	If True, adds one-step-ahead fitted values to the result.

Returns: dict Keys: {"mean": jnp.ndarray}. Optionally includes prediction intervals and {"fitted": jnp.ndarray} if fitted=True.