A Bayesian hierarchical time series forecasting model based on Metaβs Prophet.
This method forecasts all bottom series in a hierarchy at once, using a MultivariateNormal as the likelihood function and LKJ priors for the correlation matrix.
This forecaster is particularly interesting if you want to fit shared coefficients across series. In that case, shared_features parameter should be a list of feature names that should have that behaviour.
Parameters
Name
Type
Description
Default
trend
Union[BaseEffect, str]
Trend component of the model.
"linear"
feature_transformer
BaseTransformer
Transformer for features preprocessing.
None
exogenous_effects
optional
Effects to model exogenous variables.
None
default_effect
optional
Default effect specification.
None
shared_features
optional
Features shared across time series.
None
noise_scale
float
Scale parameter for the noise distribution.
0.05
correlation_matrix_concentration
float
Concentration parameter for the correlation matrix.