Multivariate Pairs Trading Using Temporal Dependence Structures

Multivariate pairs trading is a strategy that tries to exploit inefficiencies in the relative value pricing between stocks and baskets of related assets. In this research we set up such strategies, where the baskets are identified by elastic net regularization. The elastic net rigorously combines Lasso and ridge regression resulting in compact and robust baskets. We model the spread dynamics of pairs using copula-based (semi-)parametric time series models. This type of model allows for a range of dependency structures and marginals that can be modelled separately. This affords great flexibility and opens up a new way on modelling the spread dynamics. We investigate the strategies’ performance for the Japanese universe using daily stock prices of the Nikkei 225 index constituents ranging from August 3rd 2001 to April 1st 2010. We find that the generalization to account for nonlinear associations in the spread dynamics does not necessarily lead to more trading opportunities. However, this generalization leads to ‘better’ trading signals, i.e. a higher rate of trades that converges. After optimizing the input parameters we find annualized Sharpe ratios up to 1.35 for the copula model and 0.69 for its linear counterpart. Hedging the positions by the identified baskets rather than only stocks’ sector members lowers the portfolios’ standard deviations, but does not lead to significant higher Sharpe ratios.