Pairs Trading Using Machine Learning: An Empirical Study

Pairs trading is a quantitative trading strategy that exploits financial markets that are out of equilibrium. By identifying a pair of stocks that historically move together, and assuming that their price difference is mean-reverting, an investor can profit from deviations from the mean by taking a long-short position in the chosen pair. Throughout the years, several trading frameworks and methods have been established in order to optimize this strategy. These methods, in particular the stochastic (residual) spread method, are mainly based on the more traditional estimation techniques, such as the Expectation Maximization algorithm. Since machine learning techniques are becoming more popular in finance, we propose to develop a framework for pairs trading using neural networks. This thesis analyzes the performance of neural networks in pairs trading applied to Exchange Traded Funds (ETFs) both statistically and economically, and compares the performance with the more traditional methods. The results show that recurrent neural network is superior compared to the other methods, since it generates the largest returns, around 11%, as well as the highest Sharpe and Sortino ratios.