pymc_forecast.gaussian#

Correlated-over-time observation noise: the Gaussian prefix conditional.

For elementwise noise the forecast distribution is the horizon marginal (what predict() does implicitly). When the observation noise is a MultivariateNormal over the time axis — GP-style correlated residuals — the forecast must instead be the Gaussian conditional given the observed prefix. This module ports numpyro_forecast’s _mvn_prefix_condition (Cholesky of the prefix block, triangular solves for the conditional mean and covariance, jitter + symmetrization) to PyTensor and wires it into a predict-style primitive.

Univariate series only (the MVN event dimension is time), matching upstream.

pymc_forecast.gaussian.conditional_mvn(loc, cov, observed_prefix, *, jitter=1e-06)[source]#

Condition a length-t+f MVN on its first t observed values.

Parameters:

  • loc – Full-horizon mean, shape (t + f,).

  • cov – Full-horizon covariance, shape (t + f, t + f).

  • observed_prefix – Observed values, shape (t,).

  • jitter – Diagonal floor added to the prefix and conditional covariances.

Returns:

(cond_mean, cond_cov) – Mean (f,) and covariance (f, f) of the forecast conditional.

pymc_forecast.gaussian.predict_mvn(h, loc, cov, *, jitter=1e-06)[source]#

Register MVN-over-time observation/forecast variables.

The correlated-noise counterpart of predict(): the likelihood is MvNormal over the observed window (the prefix block of cov) and — when forecasting — the "forecast" variable is the exact Gaussian conditional given the observed prefix, not the horizon marginal.

Parameters:

  • h – The horizon of the current model build.

  • loc – Full-horizon mean with shape (duration,) (time on axis 0).

  • cov – Full-horizon covariance with shape (duration, duration) — e.g. a GP kernel gram matrix over the time positions.

  • jitter – Diagonal floor for numerical stability.

Raises:

HorizonError – For multivariate data (the MVN event dim must be time) or when forecasting without observed data.