SOFTSSharp

chronax.models.SOFTSSharp · inherits BaseForecaster

SOFTSSharp (Self-Organizing Feature Transformation for Time Series) model. A univariate time series forecasting model based on the SOFTS architecture, implemented using Flax/NNX. This model is designed for short-term forecasting and uses a combination of feature transformation and self-organizing layers.

Attributes: * uses_exog: False. This model does not use exogenous variables. * alias: 'softs-sharp'.

__init__(self, h: int, input_size: int, hidden_size: int = 64, num_layers: int = 2, dropout_rate: float = 0.1, learning_rate: float = 0.001, max_epochs: int = 100, early_stopping_patience: int = 10, seed: int = 42)

Initializes the SOFTSSharp forecaster.

Parameter	Type	Default	Description
h	int	-	Forecast horizon.
input_size	int	-	The lookback window size (context length).
hidden_size	int	64	Dimension of the hidden layers.
num_layers	int	2	Number of SOFTS layers.
dropout_rate	float	0.1	Dropout probability.
learning_rate	float	0.001	Learning rate for the optimizer.
max_epochs	int	100	Maximum number of training epochs.
early_stopping_patience	int	10	Patience for early stopping.
seed	int	42	Random seed for reproducibility.

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

Trains the SOFTSSharp model.

The model is built and trained using the provided time series data y. Exogenous variables X are ignored as uses_exog is False.

Parameter	Type	Default	Description
y	jnp.ndarray	-	Univariate time series data (T, 1).
X	jnp.ndarray	None	Ignored. Exogenous variables.

Returns: Self (the fitted forecaster; sets self.model_).

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

Generates point forecasts using the fitted model.

Parameter	Type	Default	Description
h	int	-	The forecast horizon (must match self.h).
X	jnp.ndarray	None	Ignored. Exogenous variables.
level	list[float]	None	Prediction intervals levels (e.g., [80, 95]). Currently ignored as SOFTSSharp is deterministic.

Returns: dict A dictionary containing the forecasts. Keys: * mean: jnp.ndarray (h, 1) - Point forecasts.

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

Fits the model and generates forecasts in one step.

If fitted is True, it uses the existing self.model_ for prediction. Otherwise, it calls fit(y) followed by predict(h).

Parameter	Type	Default	Description
y	jnp.ndarray	-	Historical time series data (T, 1).
h	int	-	Forecast horizon.
X	jnp.ndarray	None	Ignored. Historical exogenous variables.
X_future	jnp.ndarray	None	Ignored. Future exogenous variables.
level	list[float]	None	Prediction intervals levels. Ignored.
fitted	bool	False	Whether the model is already fitted.

Returns: dict Forecast results, same structure as predict.