Prior vs Posterior

Visualize how model parameters shift from prior assumptions to learned posteriors after fitting.

Overview

Prior vs Posterior shows how each variable's effect distribution changes after model fitting. By comparing the assumed priors with the learned posteriors, you can assess whether the model has enough data to update its beliefs and whether the fitted effects align with domain expectations.

Visualizations

Toggle between two plot types using the radio buttons in the header:

Box — box plots showing the distribution spread (median, quartiles, outliers).
Density — smoothed density curves for a more detailed shape comparison.

Plots are organized into tabs by variable group (Media, Price, Distribution, Holiday, Season, Trend, etc.).

Transformations

The Transformation selector controls what is shown on the X axis:

Transformation	Description
Effect	The raw impact the variable has on the target
Impact Ratio	Effect normalized by the target sum: `effect / sum(abs(target))`
Mean	Mean effect per epoch
Return	Effect normalized by the variable size: `effect / sum(abs(variable))`

Filtering

Click Filters to narrow results by:

Variables — select specific variables
Groups — select variable groups
Targets — select target variables

When to Use

Model validation — if posteriors barely shift from priors, the model may lack sufficient data for that variable.
Sanity checking — verify that effect directions match business intuition (e.g., media spend should have a positive effect).
Comparing models — run prior vs posterior on different model versions to see how parameter estimates evolve.