Benchmarks

The benchmark suite has two jobs: validate numerical behavior and measure runtime. They are kept separate on purpose. A timing comparison is only meaningful after both implementations have been shown to represent and fit the same model.

VineCopulas.jl uses R rvinecopulib as the external reference implementation.

Goal	Command	Main output
Correctness	benchmarks/correctness/run_correctness_gate.sh	strict PASS/FAIL reports
Evaluation speed	benchmarks/run_main.sh	timing and numerical summaries
Fitting speed	benchmarks/fitting/run_fit.sh	compact fitting comparison

What is covered?

The correctness gate covers pair-copula primitives, fixed general R-vines, fixed-structure fitting, and automatic Dißmann/Kruskal structure selection. The performance suite separately measures evaluation, transforms, simulation, and fitting.

The standard evaluation battery uses Gaussian, Clayton, Gumbel, and Frank D-vines at  . Student-t is kept in a separate study because its runtime profile is dominated by scalar Student-t CDF and quantile evaluations.

Reading the results

Timing tables use a plain-language final column such as Julia 1.56× faster, R 1.33× faster, or near parity. The raw ratio is therefore not needed to interpret a table.

All reported timings are local measurements, not hardware-independent performance claims. Generated CSV files contain the detailed values; the documentation keeps only the results needed to understand the current performance profile.

For the exact commands and generated files, see Reproducing results.