chronax.models.tide.model

`TiDEConfig`

chronax.models.TiDEConfig

Hyperparameters for :class:TiDE.

Attributes

Attribute	Type	Default	Description
`h`	`int`	`24`	Forecast horizon.
`input_size`	`int`	`48`	History window length `L`.
`hidden_size`	`int`	`512`	MLP hidden width for encoder/decoder blocks.
`decoder_output_dim`	`int`	`32`	Per-step output dimension of the dense decoder.
`temporal_decoder_dim`	`int`	`128`	Hidden size of the temporal decoder MLP block.
`dropout`	`float`	`0.3`	Dropout rate (0 = disabled).
`layernorm`	`bool`	`True`	Insert LayerNorm after each MLPResidual block output.
`num_encoder_layers`	`int`	`1`	Number of stacked MLPResidual encoder layers.
`num_decoder_layers`	`int`	`1`	Number of stacked MLPResidual decoder layers.
`temporal_width`	`int`	`4`	Projected feature dimension for temporal covariates.
`futr_exog_size`	`int`	`0`	Number of future exogenous features `F`.
`hist_exog_size`	`int`	`0`	Number of historic exogenous features `X`.
`stat_exog_size`	`int`	`0`	Number of static exogenous features `S`.
`output_size`	`int`	`1`	Outputs per step — 1 for point forecasts.

`MLPResidual`

chronax.models.MLPResidual · inherits flax.linen.Module

MLP with skip connection and optional LayerNorm.

Mirrors NeuralForecast's MLPResidual exactly:

h = relu(Dense(hidden_size)(x))
h = Dense(output_dim)(h)
h = Dropout(h)
out = h + Dense(output_dim)(x)   # skip
if use_layernorm: out = LayerNorm(out)

Operates on the last axis, so both [B, D] and [B, T, D] inputs are handled correctly without reshaping.

`init(self, hidden_size, output_dim, dropout_rate=0.0, use_layernorm=True)`

Parameter	Type	Default	Description
`hidden_size`	`int`	-	(undocumented)
`output_dim`	`int`	-	(undocumented)
`dropout_rate`	`float`	`0.0`	(undocumented)
`use_layernorm`	`bool`	`True`	(undocumented)

`call(self, x, deterministic=True) -> jnp.ndarray`

Parameter	Type	Default	Description
`x`	`jnp.ndarray`	-	(undocumented)
`deterministic`	`bool`	`True`	(undocumented)

Returns: jnp.ndarray (undocumented)

`TiDE`

chronax.models.TiDE · inherits flax.linen.Module

Time-series Dense Encoder (TiDE).

Input insample_y is expected as [B, L, 1] (consistent with the rest of Chronax). The trailing feature dim is squeezed inside the forward pass before concatenating covariates and running the encoder stack.

`init(self, config)`

Parameter	Type	Default	Description
`config`	`TiDEConfig`	-	:class:`TiDEConfig` specifying the full architecture.

`call(self, insample_y, hist_exog=None, futr_exog=None, stat_exog=None, deterministic=True) -> jnp.ndarray`

Parameter	Type	Default	Description
`insample_y`	`jnp.ndarray`	-	`[B, L, 1]` (undocumented)
`hist_exog`	`Optional[jnp.ndarray]`	`None`	`[B, L, X]` (undocumented)
`futr_exog`	`Optional[jnp.ndarray]`	`None`	`[B, L+h, F]` (undocumented)
`stat_exog`	`Optional[jnp.ndarray]`	`None`	`[B, S]` (undocumented)
`deterministic`	`bool`	`True`	(undocumented)

Returns: jnp.ndarray ([B, h, output_size])

TiDEConfig

Attributes

MLPResidual

__init__(self, hidden_size, output_dim, dropout_rate=0.0, use_layernorm=True)

__call__(self, x, deterministic=True) -> jnp.ndarray

TiDE

__init__(self, config)

__call__(self, insample_y, hist_exog=None, futr_exog=None, stat_exog=None, deterministic=True) -> jnp.ndarray