VDC · Python library · NeurIPS 2026

Vine Denoising Copula

One pretrained network estimates every pair-copula of a vine. VDC maps the empirical histogram of each edge to a copula density grid in a single forward pass, projects it to uniform marginals, and plugs it into standard vine recursion for likelihoods, sampling and information analysis.

VDC overview: training once on simulated copulas, fitting a vine tree by tree with the frozen estimator, and the operations the fitted vine supports

Start here

Models and computation

Train and reproduce

Results at a glance

Bivariate accuracy versus latency, vine likelihood, fitting time versus dimension and S&P100 likelihood
Bivariate accuracy and latency, vine likelihood, fitting time against multithreaded pyvine-TLL, and held-out likelihood on S&P100 returns (NeurIPS 2026, Figure 2).

Project overview · Source code · Report an issue