Mixed Copulas for Modeling Dependence in Pairs Trading

Summary

This article describes combining copula families to represent dependence patterns that a single family may miss. It outlines mixtures of Clayton, Frank, Gumbel, and Student-t copulas, whose differing tail behavior can capture joint extremes as well as dependence in the center of a distribution. The examples are motivated by pairs trading and include mixtures intended to represent asymmetric tail dependence or more symmetric heavy-tail behavior.

For estimating a mixture, the article presents a penalized likelihood approach using SCAD to shrink unimportant component weights toward zero while limiting bias in larger weights. Cross-validation selects penalty settings, and an Expectation-Maximization procedure alternates between estimating component membership weights and updating copula parameters. The article describes an MQL5 implementation and related model-selection and signal tools. It does not report quantified trading results or establish that a particular mixture is best; selecting candidate families, tuning the model, using representative training data, and recalibrating as market conditions change remain important limitations. It also notes the added computational cost and non-convex optimization involved.

Key ideas

• A mixed copula combines component families to model several forms of dependence in one distribution.
• Clayton and Gumbel components represent lower and upper tail dependence, while Frank emphasizes central dependence without tail dependence.
• SCAD penalization can shrink weak mixture weights while reducing the bias associated with shrinking large weights.
• Cross-validation tunes the penalty, and EM alternates between component weighting and parameter estimation.
• Model quality depends on candidate selection, representative data, recalibration, and computationally demanding optimization.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.