---
title: "Advanced Predictive Diagnostics"
output: rmarkdown::html_vignette
vignette: >
  %\VignetteIndexEntry{Advanced Predictive Diagnostics}
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
---

```{r setup, include=FALSE}
knitr::opts_chunk$set(collapse = TRUE, comment = "#>")
```

This article extends the prediction layer with diagnostics that remain
descriptive rather than becoming automatic acceptance or rejection rules.

```{r}
library(gp3bayes)

p <- c(0.05, 0.20, 0.75, 0.90)
y <- c(0, 0, 1, 1)

binary_confusion_table(p, y)
binary_roc_curve(p, y)
binary_precision_recall_curve(p, y)
binary_calibration_error(p, y, bins = 2)
```

```{r, eval=requireNamespace("ggplot2", quietly=TRUE)}
plot_binary_roc(binary_roc_curve(p, y))
plot_binary_precision_recall(binary_precision_recall_curve(p, y))
```

For a fitted model, posterior predictive discrepancy checks retain the entire
replicated distribution:

```{r, eval=FALSE}
pp <- predict_model(
  fit,
  type = "predictive",
  include_group_effects = TRUE,
  ndraws = 1000
)

check <- posterior_predictive_statistic(pp, statistic = "mean")
ppc_statistic_table(check)
plot_ppc_statistic(check)
```

Duration models additionally support predictive Q-Q and tail checks:

```{r, eval=FALSE}
duration_qq_table(pp)
plot_duration_qq(pp)

tail <- duration_tail_check(pp, threshold = 2000)
plot_duration_tail(tail)
```

None of these diagnostics certifies adequacy automatically.