Granger Causality in Expectiles: an M-vine copula test

Authors: Roberto Fuentes-Martínez, Irene Crimaldi Date: 2026-03-24 Paper ID: arxiv:2603.23294

Summary

This paper introduces a novel, model-free measure of Granger causality defined in terms of expectiles, which generalizes the traditional mean-based approach to capture distributional dependencies across the conditional distribution. The authors develop a multivariate test statistic for $d+1$ time series utilizing M-vine copula models to account for complex, non-linear, and non-Gaussian dependencies without imposing specific parametric forms on the joint distribution. Strong consistency of the test is proven, and finite-sample simulations confirm accurate size control and increasing power with sample size. A significant advantage demonstrated is the ability to jointly test for causality, revealing relationships undetectable by simpler pairwise methods, as validated through applications to international stock market indices.

Key Contributions

• Introduction of a model-free measure for Granger causality based on expectiles, allowing for analysis across the entire conditional distribution rather than just the mean.
• Development of a multivariate Granger causality test for $d+1$ series using M-vine copula models, which handles non-linear and non-Gaussian dependence without parametric assumptions on the joint distribution.
• Establishment of the strong consistency of the proposed test statistic under specific regularity conditions.
• Demonstration of the framework’s capability to detect multivariate causal relationships that are missed by pairwise tests, both theoretically and empirically.

Limitations

The test is subject to regularity conditions that must be met for the strong consistency proof, and performance in very small finite samples should be assessed carefully.

Open Questions & Future Work

• asymptotic-distribution-m-vine-test

Limitations

The test is subject to regularity conditions that must be met for the strong consistency proof, and performance in very small finite samples should be assessed carefully.

Links

• ArXiv Abstract
• PDF