`TiDEForecaster`

chronax.forecaster.TiDEForecaster · inherits BaseForecaster

High-level fit/forecast adapter for Chronax TiDE. Mirrors neuralforecast.NeuralForecast(models=[TiDE(...)]).fit(df).predict() with a familiar fit(y) -> self / forecast(y) -> ndarray interface.

init(self, h, input_size=-1, hidden_size=512, decoder_output_dim=32, temporal_decoder_dim=128, dropout=0.3, layernorm=True, num_encoder_layers=1, num_decoder_layers=1, temporal_width=4, futr_exog_size=0, hist_exog_size=0, stat_exog_size=0, output_size=1, *, max_steps=1000, learning_rate=0.001, batch_size=1024, random_seed=0, alias='TiDE', loss='mae', scale=True, global_scale=False, use_std_scaler=True, grad_clip=0.0, weight_decay=0.0, use_lr_schedule=False, window_sampling=True, val_fraction=0.1, val_check_steps=100, early_stop_patience_steps=-1)

Initializes the TiDE forecaster configuration and training parameters.

Parameter	Type	Default	Description
`h`	`int`	-	Forecast horizon.
`input_size`	`int`	`-1`	History window length; -1 (default) uses `3 * h`.
`hidden_size`	`int`	`512`	forwarded to :class:`TiDEConfig`.
`decoder_output_dim`	`int`	`32`	forwarded to :class:`TiDEConfig`.
`temporal_decoder_dim`	`int`	`128`	forwarded to :class:`TiDEConfig`.
`dropout`	`float`	`0.3`	forwarded to :class:`TiDEConfig`.
`layernorm`	`bool`	`True`	forwarded to :class:`TiDEConfig`.
`num_encoder_layers`	`int`	`1`	forwarded to :class:`TiDEConfig`.
`num_decoder_layers`	`int`	`1`	forwarded to :class:`TiDEConfig`.
`temporal_width`	`int`	`4`	forwarded to :class:`TiDEConfig`.
`futr_exog_size`	`int`	`0`	forwarded to :class:`TiDEConfig`.
`hist_exog_size`	`int`	`0`	forwarded to :class:`TiDEConfig`.
`stat_exog_size`	`int`	`0`	forwarded to :class:`TiDEConfig`.
`output_size`	`int`	`1`	forwarded to :class:`TiDEConfig`.
`max_steps`	`int`	`1000`	Total optimiser steps.
`learning_rate`	`float`	`1e-3`	Adam(W) peak learning rate.
`batch_size`	`int`	`1024`	Windows sampled per training step.
`random_seed`	`int`	`0`	PRNG seed for parameter init and window sampling.
`alias`	`str`	`"TiDE"`	Display name for external reporting.
`loss`	`Union[str, LossFn]`	`"mae"`	Registered name (`"mae"`, `"mse"`) from :mod:`chronax.models.tide.loss` or a callable `(y, y_hat) -> scalar`.
`scale`	`bool`	`True`	Apply per-window `RobustScaler` during training and inference. Strongly recommended; matches NeuralForecast's `scaler_type`.
`global_scale`	`bool`	`False`	(undocumented)
`use_std_scaler`	`Optional[bool]`	`True`	(undocumented)
`grad_clip`	`float`	`0.0`	Global gradient-norm clip (0 = disabled).
`weight_decay`	`float`	`0.0`	AdamW weight decay (0 = plain Adam).
`use_lr_schedule`	`bool`	`False`	Warmup-cosine LR schedule; False = constant LR.
`window_sampling`	`bool`	`True`	Use all rolling windows (True, recommended) vs only the last window per series (False).
`val_fraction`	`float`	`0.1`	Fraction of windows held out for validation (0 = none).
`val_check_steps`	`int`	`100`	Evaluate validation loss every this many steps.
`early_stop_patience_steps`	`int`	`-1`	Stop when val loss does not improve for this many consecutive checks; -1 = disabled.

`fit(self, y, *, verbose=False) -> Self`

Train the model in-place on y.

All rolling windows are extracted from every series long enough to yield at least one window. Per-window RobustScaler is applied inside the JIT-compiled train_step_windows.

Parameters:

Parameter	Type	Default	Description
`y`	`SeriesLike`	-	A single 1-D series or a list of 1-D series (panel mode).
`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) -> jnp.ndarray`

Produce point forecasts from the last input_size of each series.

Per-window RobustScaler statistics are computed on the last input_size observations of each series (matching the scaler applied during training) and the forecast is inverse-transformed before return.

Parameters:

Parameter	Type	Default	Description
`y`	`Optional[SeriesLike]`	`None`	Series to forecast for (required).
`h`	`Optional[int]`	`None`	Horizon override — must match `config.h` if provided.

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

`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: NotImplementedError: TiDEForecaster.predict(level=...) is not yet supported.

`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 (undocumented).

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

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

Parameters:

Parameter	Type	Default	Description
`**overrides`		-	(undocumented)

Returns: TiDEForecaster (undocumented).