chronax.models.rnn.forecaster

The following reference is generated from forecaster.py.

`chronax.models.rnn.forecaster`

High-level fit/forecast adapter around the Chronax RNN.

`RNNForecaster`

chronax.models.rnn.forecaster.RNNForecaster · inherits BaseForecaster

High-level fit/forecast wrapper around :class:chronax.models.rnn.RNN.

init(self, h, input_size=-1, encoder_hidden_size=128, encoder_n_layers=2, encoder_activation='tanh', encoder_bias=True, encoder_dropout=0.0, decoder_hidden_size=128, decoder_layers=2, futr_exog_size=0, hist_exog_size=0, stat_exog_size=0, output_size=1, recurrent=False, cell_type='elman', layer_norm=False, *, max_steps=1000, learning_rate=0.001, batch_size=32, random_seed=0, alias='RNN', loss='mae', scale=True, grad_clip=1.0, window_sampling=True, use_lr_schedule=True, weight_decay=0.0, val_fraction=0.1, val_check_steps=100)

Initializes the RNN Forecaster.

Parameter	Type	Default	Description
`h`	`int`	-	forecast horizon.
`input_size`	`int`	`-1`	history window length; -1 (default) uses `3 * h`.
`encoder_hidden_size`	`int`	`128`	forwarded to :class:`RNNConfig`.
`encoder_n_layers`	`int`	`2`	forwarded to :class:`RNNConfig`.
`encoder_activation`	`str`	`"tanh"`	forwarded to :class:`RNNConfig`.
`encoder_bias`	`bool`	`True`	forwarded to :class:`RNNConfig`.
`encoder_dropout`	`float`	`0.0`	forwarded to :class:`RNNConfig`.
`decoder_hidden_size`	`int`	`128`	forwarded to :class:`RNNConfig`.
`decoder_layers`	`int`	`2`	forwarded to :class:`RNNConfig`.
`futr_exog_size`	`int`	`0`	forwarded to :class:`RNNConfig`.
`hist_exog_size`	`int`	`0`	forwarded to :class:`RNNConfig`.
`stat_exog_size`	`int`	`0`	forwarded to :class:`RNNConfig`.
`output_size`	`int`	`1`	forwarded to :class:`RNNConfig`.
`recurrent`	`bool`	`False`	forwarded to :class:`RNNConfig`.
`cell_type`	`str`	`"elman"`	forwarded to :class:`RNNConfig`.
`layer_norm`	`bool`	`False`	forwarded to :class:`RNNConfig`.
`max_steps`	`int`	`1000`	number of optimiser steps performed by :meth:`fit`.
`learning_rate`	`float`	`1e-3`	Adam learning rate.
`batch_size`	`int`	`32`	number of windows per training step.
`random_seed`	`int`	`0`	PRNG seed for parameter init and training shuffling.
`alias`	`str`	`"RNN"`	display name for external reporting.
`loss`	`Union[str, LossFn]`	`"mae"`	registered name (`"mae"`, `"mse"`) from :mod:`chronax.models.rnn.loss` or a callable `(y, y_hat) -> scalar`.
`scale`	`bool`	`True`	if True, standardise each series before training and undo the scaling on the forecast. Strongly recommended for raw real-world data (e.g. airline passengers, hourly temperature).
`grad_clip`	`float`	`1.0`	global gradient-norm clip threshold (0 = disabled).
`window_sampling`	`bool`	`True`	if True, each training step samples a random `input_size + h` window. When no exogenous variables are present, all valid windows are pre-extracted into a device cache at the start of `fit`; each step then becomes a cheap fancy-index instead of 128 Python `_sample_window` calls.
`use_lr_schedule`	`bool`	`True`	if True, wrap Adam with a warmup + cosine-decay schedule decaying to 1 % of `learning_rate`. Set False to use a constant learning rate (matches NeuralForecast's default optimizer behaviour).
`weight_decay`	`float`	`0.0`	(undocumented)
`val_fraction`	`float`	`0.1`	(undocumented)
`val_check_steps`	`int`	`100`	(undocumented)

`fit(self, y, *, hist_exog=None, futr_exog=None, stat_exog=None, verbose=False) -> Self`

Train the model in-place on y.

Parameters:

Parameter	Type	Default	Description
`y`	`SeriesLike`	-	a single 1D series or a list of 1D series (panel).
`hist_exog`	`Optional[Sequence[jnp.ndarray]]`	`None`	per-series `[T_i, X]` historic covariates, or None.
`futr_exog`	`Optional[Sequence[jnp.ndarray]]`	`None`	per-series `[T_i + h, F]` future covariates that cover both history and the forecast horizon, or None.
`stat_exog`	`Optional[Sequence[jnp.ndarray]]`	`None`	per-series `[S]` static covariates, or None.
`verbose`	`bool`	`False`	print loss every `max(1, max_steps // 10)` steps.

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

`forecast(self, y=None, h=None, *, hist_exog=None, futr_exog=None, stat_exog=None) -> jnp.ndarray`

Produce horizon predictions.

Parameters:

Parameter	Type	Default	Description
`y`	`Optional[SeriesLike]`	`None`	optional series to forecast for. If None, uses the same series fit() was last called with — currently unsupported (the forecaster does not retain training data); pass it explicitly.
`h`	`Optional[int]`	`None`	optional explicit horizon. If None, uses `self.config.h`. (We do not yet support changing the horizon at predict time; this argument is accepted for API symmetry only.)
`hist_exog`	`Optional[Sequence[jnp.ndarray]]`	`None`	(undocumented)
`futr_exog`	`Optional[Sequence[jnp.ndarray]]`	`None`	(undocumented)
`stat_exog`	`Optional[Sequence[jnp.ndarray]]`	`None`	(undocumented)

Returns: jnp.ndarray of shape [h] for univariate input or [n_series, h] for panel input. Raises: RuntimeError, ValueError.

`predict(self, h=None, X=None, level=None) -> dict`

Forecast from the series passed to :meth:fit.

Satisfies the BaseForecaster contract: unlike :meth:forecast, this does not take y explicitly — it reuses the series cached at fit time.

Parameters:

Parameter	Type	Default	Description
`h`	`Optional[int]`	`None`	Forecast horizon. Defaults to `config.h` when None.
`X`	`Optional[jnp.ndarray]`	`None`	Reserved for future exogenous regressors; unused.
`level`	`Optional[List[Union[int, float]]]`	`None`	Not yet supported for this model.

Returns: dict: {"mean": jnp.ndarray}. Raises: RuntimeError, NotImplementedError.

`fit_predict(self, y, **kwargs) -> jnp.ndarray`

Fit on y and immediately forecast for the same series.

Parameters:

Parameter	Type	Default	Description
`y`	`SeriesLike`	-	(undocumented)
`**kwargs`			(undocumented)

Returns: jnp.ndarray.

`fitted`

bool

Returns whether the forecaster has been fitted (self._state is not None).

`with_config(self, **overrides) -> RNNForecaster`

Return a fresh (unfitted) forecaster with overridden config fields.