BayesianPredictionSettings

BayesianPredictionSettings()

Versioned settings for the complete Bayesian workflow.

Attributes

Name	Type	Description
model_config		Frozen configuration rejecting extra and nonfinite values.
method	Literal['independent', 'correlation adjusted']	Independent precisions or a joint Gaussian conditional expectation.
series	Literal['combined', 'backtests', 'predictions']	Historical, forward, or stitched estimates.
values	tuple[str, …]	Output names to publish.
minimum_observations	int	Minimum error count; a prior still requires two.
use_absolute_errors	bool	Absolute growth errors; false uses relative errors.
stdev_weighted_consensus	bool	Prefer analyst dispersion over historical errors.
consensus_count_exponent	float | None	Optional exponent on positive analyst counts.
consensus_stdev_variance_multiplier	float	Positive dispersion variance scale.
consensus_bias_lambda	float	Fraction of the filtered consensus bias to remove.
consensus_bias_process_noise	float	Local-level Kalman process variance.
consensus_bias_measurement_noise	float	Local-level Kalman measurement variance.
winsorize_fraction	float | None	Share of each tail clipped from historical growth and error inputs, each window on its own bounds. None clips nothing.
include_model	bool	Include model observations in the independent update.
interval_method	Literal['student_t', 'empirical']	Student-t shape, or point-in-time empirical quantiles of past standardised residuals pooled across every series.
interval_width	float	Interval coverage.
interval_df	float | None	Student-t degrees of freedom. Omit all three shape parameters to use the calibration, which covers the dispersion-weighted profile at the calibrated bias lambda only.
interval_loc	float | None	Student-t location.
interval_scale	float | None	Student-t scale.
empirical_interval_minimum	int	Fewest past residuals an empirical interval reads its quantiles from.
directional_confidence_cap	float | None	Ceiling on directional confidence below consensus at short horizons with non-positive consensus growth. None leaves it uncapped.
directional_cap_horizon_days	int	Longest horizon, in days before publication, the cap applies to.
start_date	date | None	First period included in the training sample.
relative_to	Literal['publish', 'pd'] | None	Optional publication-relative anchoring, with pd as an alias.
relative_days	int | Literal['latest'] | None	Nonpositive horizon or latest live prediction horizon.
preset	str | None	Name of the preset these settings come from, kept when a preset is derived under a new version. None when built directly.
fitted_through	date | None	Last date the data behind any fitted value covered. None for chosen values. Filled from the calibration’s date when the interval comes from the calibration.
version	str	Audit identifier for these settings.