| Abstract [eng] |
Pairs trading is an old strategy used in the stock market. Being market neutral, it is especially appealing during periods of high volatility. In the last decade, many researchers in their works explored various methods for optimisation of the simple pairs trading strategy both in selecting the assets fit for the strategy and for initiating trading positions. In this paper pairs selection methods are not analysed focusing on the generation of trading signals instead. The traditional method, in this work referred to as simple pairs trading strategy, is based on preconceived limits for the pairs spread. As the spread reaches one of the limits, trading is started and when the spread returns to its equilibrium – trading is stopped, thus turning a profit. Several methods based on genetic algorithm for optimisation of these limits are compared. Reinforced learning algorithms are used not for the optimisation of simple pairs trading strategy limits but as autonomous formation of trading signals. Q-learning with a deep learning neural network and a double deep learning neural network methods are used for this purpose. Pairs are selected from the S&P 500 index using a popular cointegration approach. During an uncertain period in the testing range, pairs trading strategy profits remained higher than that of the S&P 500 index. Algorithms of reinforced learning return the best profits. |
