NSSDNResult.forecast#

NSSDNResult.forecast(steps=1, *, exog=None, time=None)[source]#

Recursive multi-step forecast from the end of the fitted sample.

The network is warmed up on the sample under teacher forcing, the last observation \(y_{n-1}\) is fed as the first future input, and every later input is the previous emission,

\[h_{n+s} = F_\psi\bigl(h_{n+s-1}, \hat y_{n+s-1}, \Delta t_{n+s}\bigr), \qquad \theta_{n+s,k} = \theta_{n+s-1,k} + \omega_{n+s,k}\,\Delta t_{n+s}, \qquad s = 1, \ldots, S,\]

with \(\hat y_{n} \equiv y_{n-1}\) for the first step. Error accumulates with the horizon; the validation notebook found this mode far weaker than the conditional one-step mode under the filtered alignment, and the paper reports both for that reason.

Parameters:

  • steps (int) – Periods ahead \(S\), at least one.

  • exog (ArrayLike | None) – (steps, p) future covariates; required when the fit used them. A one-dimensional array is reshaped to (steps, -1).

  • time (ArrayLike | None) – (steps,) future grid after time[-1]; defaults to continuing at the sample’s last step size, so a regular grid needs nothing.

Return type:

NDArray[double]

Returns:

(steps,) forecasts on the original scale.

Raises:

  • SpecificationError – If steps is not a positive integer, or time does not have steps entries.

  • DimensionError – If the future covariates are missing or mis-shaped for a fit that used them.

Example

>>> import numpy as np
>>> y = np.sin(0.2 * np.arange(120.0))
>>> res = NSSDN(y, k=2, state_dim=8, schedule=(Stage(1e-2, 20),), seed=0).fit()
>>> res.forecast(3).shape, res.forecast(1).shape
((3,), (1,))
>>> res.forecast(0)
Traceback (most recent call last):
    ...
ns_sdn.exceptions.SpecificationError: steps must be a positive integer; got 0.
On this page