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

`inst/benchmarks/run-benchmark.R` reproduces a benchmark of `fastgbm` against `gbm`,
`xgboost`, and `ranger` on six real survival datasets from the `biostatlab` package
(`pbc`, `heart_failure`, `breast`, `colon_cancer`, `crc_mondaca2020`, `framingham`),
under an equal-hyperparameter-grid regime with repeated 70/30 train/test splits.

```r
# from the package root
Rscript inst/benchmarks/run-benchmark.R
```

This writes, under `inst/benchmarks/`:

* `benchmark-results.csv` -- one row per (dataset, model, repeat), with training time,
  prediction time, and Harrell's C-index;
* `benchmark-summary.csv` -- median/IQR per (dataset, model), plus paired
  win/loss/tie counts of `fastgbm` against each competitor;
* `parallel-speedup.csv` -- `threads = 1` vs. `threads = <hardware concurrency>`
  training time on the largest dataset (`framingham`);
* `session-info.txt` -- R/package versions, OS, seed, and hyperparameters, for
  reproducibility.

The full write-up, with tables and figures generated directly from these CSVs, is in
`paper/fastgbm-benchmark.qmd` (renders to `fastgbm-benchmark.pdf` via `quarto render`).
Results are reported honestly (win/loss/tie per dataset), not as a universal
superiority claim -- see the paper's Comparative Summary section.