MLP

chronax.models.mlp_model.MLP · inherits BaseForecaster

MLP: Multi Layer Perceptron (flax.nnx port of neuralforecast.MLP). The simplest neural forecaster: the scaled insample window (plus any future-known exogenous inputs, flattened) feeds num_layers fully connected ReLU layers and a raw linear head that emits all h horizon steps at once — no recurrence, no attention. fit accepts a single series (T,) or an N-series panel (T, n_series); a 2-D fit cross-learns one global network over all columns (channel-independent — each column is forecast from its own tail context) and predictions follow the input rank. Future-known exogenous inputs are supported (uses_exog = True; shared across columns on a 2-D fit); historical and static exog are not modeled. Point ("mae"/"mse"/"huber"), multi-quantile (MultiQuantileLoss), or Gaussian-mixture (GMM) losses; with a GMM head the model is a probabilistic forecaster whose intervals come from seeded Monte-Carlo samples of the predictive mixture in original units. float32 throughout.

Attributes: * uses_exog: True * alias: str * conformal_params: ConformalIntervals | None * model_: MLPNet | None

__init__(self, h, input_size=-1, num_layers=2, hidden_size=1024, max_steps=1000, learning_rate=1e-3, windows_batch_size=1024, scaler_type='identity', loss='mae', quantile_sort=True, random_seed=1, alias='MLP')

Parameter	Type	Default	Description
h	-	-	(undocumented)
input_size	-	-1	If less than 1, defaults to 3 * h.
num_layers	-	2	(undocumented)
hidden_size	-	1024	(undocumented)
max_steps	-	1000	(undocumented)
learning_rate	-	1e-3	(undocumented)
windows_batch_size	-	1024	(undocumented)
scaler_type	-	"identity"	(undocumented)
loss	-	"mae"	(undocumented)
quantile_sort	-	True	(undocumented)
random_seed	-	1	(undocumented)
alias	-	"MLP"	(undocumented)

fit(self, y, X=None, *, futr_exog=None) -> Self

Trains the MLP model using pooled windows from y. Supports univariate or multi-series input (y). If y is multi-series, one global network is trained across all series.

Parameter	Type	Default	Description
y	jnp.ndarray	-	Time series data, shape (T,) or (T, n_series). Must be long enough to form at least one window (T >= input_size + 1).
X	None	None	Not supported. Raises NotImplementedError.
futr_exog	jnp.ndarray	None	Future-known exogenous inputs, aligned with y. Shape (T, F).

Returns: Self (the fitted forecaster; sets self.model_). Raises: NotImplementedError, ValueError.

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

Generates forecasts for horizon h using the fitted model and the context stored during fit.

Parameter	Type	Default	Description
h	int	-	Forecast horizon. Must be less than or equal to the h used during initialization.
X	None	None	(undocumented)
futr_exog	jnp.ndarray	None	Future-known exogenous inputs for the forecast horizon, shape (h, F). Required if the model was fit with futr_exog.
level	list[int]	None	Confidence levels (e.g., [80, 95]) for prediction intervals.

Returns: dict. Keys include: * "mean": jnp.ndarray (Point forecast. Shape (h,) for 1-D fit, (h, n_series) for 2-D fit. Derived from median for quantile loss, analytic mean for distribution loss, or raw output for point loss). * "lo-L": jnp.ndarray (Lower bound of the L% prediction interval, present if level is provided). * "hi-L": jnp.ndarray (Upper bound of the L% prediction interval, present if level is provided).

forecast(self, y, h, X=None, X_future=None, *, futr_exog=None, level=None, fitted=False) -> dict

Stateless fit-then-predict. X is unsupported (MLP models future-known exog only); X_future = future-known exog for the horizon (h, F); futr_exog = its history (T, F).

Parameter	Type	Default	Description
y	jnp.ndarray	-	Time series data used for fitting.
h	int	-	Forecast horizon.
X	None	None	Not supported. Passed to fit.
X_future	jnp.ndarray	None	Future-known exogenous inputs for the forecast horizon, passed as futr_exog to predict. Shape (h, F).
futr_exog	jnp.ndarray	None	Historical future-known exogenous inputs, passed as futr_exog to fit. Shape (T, F).
level	list[int]	None	Confidence levels for prediction intervals.
fitted	bool	False	If True, computes and returns fitted values (one-step ahead predictions on the training set). Only supported for 1-D fits without temporal exog.

Returns: dict. Same keys as predict. If fitted=True, includes: * "fitted": jnp.ndarray (Fitted values).

conformity_scores(self, y, X=None) -> jnp.ndarray

Calculates conformity scores for the provided series y.

Parameter	Type	Default	Description
y	jnp.ndarray	-	Time series data. Must be 1-D.
X	None	None	(undocumented)

Returns: jnp.ndarray (Conformity scores). Raises: ValueError.