Posterior Plot Generators

Overview

This page documents three visual diagnostic generators used after model fitting:

• posterior predictive fit checking,
• posterior forest plots,
• prior-versus-posterior comparison.

They provide fast visual evidence for fit quality and parameter identifiability.

1) Posterior Predictive Check

Function Signature

from src.diagnostics.posterior_predictive_plot import generate_ppc_plot

generate_ppc_plot(
    model: Any,
    config: dict[str, Any],
    results_dir: str,
) -> str | None

Artefact

Filename	Stage folder	Description
posterior_predictive_check.png	30_model_assessment/	ArviZ posterior predictive density overlay (num_pp_samples=100) against observed data.

Behaviour

• If posterior_predictive is missing, the function attempts to generate it via pm.sample_posterior_predictive(..., extend_inferencedata=True).
• Returns saved path on success, otherwise None.

2) Posterior Forest Plots

Function Signature

from src.diagnostics.posterior_forest import generate_posterior_forest_plot

generate_posterior_forest_plot(
    model: Any,
    config: dict[str, Any],
    results_dir: str,
) -> str | None

Artefacts

Filename	Stage folder	Description
posterior_forest.png	20_model_fit/	Forest plot for beta_channel with 94% HDI.
posterior_forest_all_params.png	20_model_fit/	Forest plot for available key parameters among beta_channel, alpha, lam, with 94% HDI.

Behaviour

• Uses HDI_PROB = 0.94.
• Returns the main posterior_forest.png path when available.

3) Prior-Posterior Comparison

Function Signature

from src.diagnostics.prior_posterior_comparison import generate_prior_posterior_comparison

generate_prior_posterior_comparison(
    model: Any,
    config: dict[str, Any],
    results_dir: str,
) -> str | None

Artefact

Filename	Stage folder	Description
prior_posterior_comparison.png	20_model_fit/	ArviZ distribution comparison (plot_dist_comparison) of prior and posterior.

Behaviour

• If prior is missing on idata, the function attempts to generate prior samples (pm.sample_prior_predictive(samples=500)) and extend idata.
• Returns saved path on success, otherwise None.

Usage Example

from src.diagnostics.posterior_forest import generate_posterior_forest_plot
from src.diagnostics.posterior_predictive_plot import generate_ppc_plot
from src.diagnostics.prior_posterior_comparison import generate_prior_posterior_comparison

generate_posterior_forest_plot(driver.model, driver.config, driver.results_dir)
generate_prior_posterior_comparison(driver.model, driver.config, driver.results_dir)
generate_ppc_plot(driver.model, driver.config, driver.results_dir)

Relationship to Workflow Stages and Gates

• Stages: 20_model_fit/ and 30_model_assessment/.
• These plots support gate review and model criticism but are not standalone machine gate fields.
• They should be interpreted alongside convergence_report.json, calibration_report.json, and pareto_k_summary.json in 50_diagnostics/.