---
title: "Validating and comparing interpolation methods"
description: "Define prediction tasks, compare validation designs, and select methods without overstating map-wide accuracy."
output: rmarkdown::html_vignette
vignette: >
  %\VignetteIndexEntry{Validating and comparing interpolation methods}
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
---

```{r setup}
library(potentiomap)
data("synthetic_wells")
p <- ps_make_points(synthetic_wells[1:16, ], "x", "y", "gw_elevation",
                    "well_id", "EPSG:26916")
```

Validation designs represent different prediction tasks. Spatial separation is
appropriate when transfer to unsampled areas matters; random folds answer a
different question.

```{r validation}
v <- ps_validate(p, c("IDW", "TPS"), design = "kfold", folds = 3,
                 prediction_mode = "direct", seed = 12)
v$fold_manifest[, c("fold_id", "training_count", "validation_count")]
comparison <- ps_compare_methods(v, metric = "rmse")
comparison$ranking[, c("method", "rmse", "finite_coverage", "rank")]
```

These scores are conditional on the recorded folds; they are not universal
method rankings or automatic map accuracy.