forecaster.ForecastingModel

Abstract base class for forecasting models.

Usage

Source

forecaster.ForecastingModel()

Subclasses implement model(), which must call predict() exactly once. The instance itself is the (pure) NumPyro model function with signature model_instance(covariates, data=None): the forecast horizon is inferred from the shapes (future = covariates.shape[-2] - data.shape[-2]).

This is the object-oriented façade over the functional API: time_series() and predict() delegate to the free functions in numpyro_forecast.functional, passing the current ~numpyro_forecast.functional.models.Horizon.

Attributes

Name Description
duration Total horizon length t + future (in time steps).
future Number of forecast time steps f (0 while training).
t_obs Number of observed (in-sample) time steps t.

duration

Total horizon length t + future (in time steps).

duration: int

future

Number of forecast time steps f (0 while training).

future: int

t_obs

Number of observed (in-sample) time steps t.

t_obs: int

Methods

Name Description
__call__() Run the model as a NumPyro model function.
markov_time_series() Sample a Markov (state-space) latent over the full horizon.
model() Define the generative model and call predict() exactly once.
predict() Register the observation/forecast sites for the model.
predict_glm() Register GLM-style observation/forecast sites from a latent predictor.
time_series() Sample a time-varying latent over the full horizon.

__call__()

Run the model as a NumPyro model function.

Usage

Source

__call__(covariates, data=None)
Parameters
covariates: Array

Covariates with time at axis -2 spanning the full horizon.

data: Array | None = None
Observed data with time at axis -2 (None for prior sampling).

markov_time_series()

Sample a Markov (state-space) latent over the full horizon.

Usage

Source

markov_time_series(
    name, init_carry, transition, xs=None, *, plates=(), reparam_config=None
)

Thin wrapper over numpyro_forecast.functional.models.markov_time_series() that threads this model’s train/forecast horizon. In-sample steps run in a scan under site name; the forecast horizon runs in a second scan under f"{name}_future" seeded by the final in-sample carry, so the guide never sees the future site.

Parameters
name: str

Sample-site name for the in-sample scan; the forecast scan uses f"{name}_future".

init_carry: Any

Initial carry PyTree fed to the first transition step.

transition: Transition

Callable (carry, x_t) -> (dist_t, carry_fn): dist_t is the per-step observation distribution (its per-step shape must carry the trailing observation dimension) and carry_fn(z_t) builds the next carry from the sampled latent z_t. The wrapper owns the sample statement, so the Markov structure cannot be broken by resampling.

xs: Array | None = None

Optional exogenous inputs spanning the full horizon with time at axis -2; split and moved into scan layout internally. None for autonomous dynamics.

plates: Sequence[tuple[str, int]] = ()

(name, size) pairs opened inside the scan body around the sample statement (the only placement NumPyro accepts around a scan).

reparam_config: Mapping[str, Reparam] | None = None
Optional site-name to ~numpyro.infer.reparam.Reparam mapping applied inside the scan body.
Returns
Array
The latent over the full horizon in package layout (*plate_batch, duration, obs) (time at axis -2).

model()

Define the generative model and call predict() exactly once.

Usage

Source

model(zero_data, covariates)
Parameters
zero_data: Array | None

Zeros shaped like the data extended to the covariate duration (shape/dtype only; None during pure prior sampling).

covariates: Array
Covariates with time at axis -2 and shape (*batch, duration, cov).

predict()

Register the observation/forecast sites for the model.

Usage

Source

predict(noise_dist, prediction)

Thin wrapper over numpyro_forecast.functional.models.predict().

Parameters
noise_dist: dist.Distribution

Zero-centered observation noise (e.g. Normal(0, sigma)).

prediction: Array
Deterministic mean with time at axis -2, shape (*batch, duration, obs).

predict_glm()

Register GLM-style observation/forecast sites from a latent predictor.

Usage

Source

predict_glm(obs_dist_fn, latent)

Thin wrapper over numpyro_forecast.functional.models.predict_glm().

Parameters
obs_dist_fn: Callable[[Array], dist.Distribution]

Link mapping the full-horizon latent predictor to the observation distribution (e.g. lambda eta: Poisson(jnp.exp(eta))).

latent: Array
The deterministic latent predictor over the full horizon, time at axis -2.

time_series()

Sample a time-varying latent over the full horizon.

Usage

Source

time_series(name, dist_fn, *, reparam=None)

Thin wrapper over numpyro_forecast.functional.models.time_series().

Parameters
name: str

Base sample-site name for the in-sample latent.

dist_fn: Callable[[], dist.Distribution]

Zero-argument callable returning the per-step prior distribution.

reparam: Reparam | None = None
Optional reparameterization (e.g. LocScaleReparam) applied to both the in-sample and forecast sites.
Returns
Array
The latent over the full horizon with time at axis -2.