LstmHyperparams

chronax.models.lstm_model.LstmHyperparams · inherits PydanticBaseModel

Pydantic model defining the hyperparameters for the LstmModel.

Attributes

Attribute	Type	Default	Description
units	int	32	Number of LSTM units
layers	int	2	Number of LSTM layers
dropout	float	0.3	Dropout rate
learning_rate	float	-	Learning rate for optimizer
batch_size	int	32	Batch size for training
epochs	int	500	Number of training epochs
context_length	int	32	Context length
prediction_window	int	8	Prediction window

LstmModel

chronax.models.lstm_model.LstmModel · inherits BaseModel

Multivariate LSTM model implementation. This model extends the univariate LSTM to handle multiple target variables simultaneously. It uses a single output layer that predicts forecast_horizon * n_targets values (flattened). Optional past and future covariates are concatenated on the feature axis. context_length and prediction_window are clamped during training to ensure at least one sliding window fits the available data.

__init__(self, params: Dict[str, Any], settings: Dict[str, Any])

Initializes the LstmModel, validating hyperparameters against LstmHyperparams.

Parameter	Type	Default	Description
params	Dict[str, Any]	-	(undocumented)
settings	Dict[str, Any]	-	(undocumented)

train(self, y_context: np.ndarray, y_target: np.ndarray, timestamps_context: np.ndarray, timestamps_target: np.ndarray, x_context: Optional[np.ndarray] = None, x_target: Optional[np.ndarray] = None, **kwargs: dict) -> LstmModel

Train the Multivariate LSTM model on given data.

Uses sliding window, multi-step learning for multiple targets.

Parameter	Type	Default	Description
y_context	np.ndarray	-	Past target values, shape (num_steps, num_targets)
y_target	np.ndarray	-	Future target values (for supervised labels)
timestamps_context	np.ndarray	-	Timestamps for context (unused here)
timestamps_target	np.ndarray	-	Timestamps for target (unused here)
x_context	Optional[np.ndarray]	None	Optional covariates aligned with y_context
x_target	Optional[np.ndarray]	None	Optional covariates aligned with y_target
**kwargs	dict	-	Must include tuning_loss

Returns: LstmModel (the fitted model instance).

predict(self, y_context: np.ndarray, timestamps_context: np.ndarray, timestamps_target: np.ndarray, x_context: Optional[np.ndarray] = None, x_target: Optional[np.ndarray] = None, **kwargs: dict) -> np.ndarray

Make predictions with the trained Multivariate LSTM model.

Predicts the required number of steps ahead for all targets using non-overlapping multi-step windows. Does not use own predictions as further inputs for target channels (y); known future covariates are used when provided.

Parameter	Type	Default	Description
y_context	np.ndarray	-	Context/history values (n_steps, n_targets)
timestamps_context	np.ndarray	-	Timestamps for context (unused)
timestamps_target	np.ndarray	-	Timestamps for target (unused)
x_context	Optional[np.ndarray]	None	Covariates aligned with y_context (required if model was trained with covariates)
x_target	Optional[np.ndarray]	None	Future covariates aligned with timestamps_target (required if trained with covariates)
**kwargs	dict	-	(undocumented)

Returns: np.ndarray (Model predictions with shape (forecast_steps, n_targets)).