Abstract:
This thesis aims to study machine learning to improve the trading performance of pairs trading strategy. Pairs trading strategy is one of the most well-known algorithm trading developed in the 1980s by a group of scientists and mathematicians. The concept of the pairs trading strategy is to exploit the mispricing of two equities which their prices tend to move in the same manner. When the algorithm captures the mispricing behavior by equations or indicators, traders open a short position on the equity which its price is relatively higher than the equilibrium and open long position of the other equity. If the prices reach an equilibrium point, the trade positions are closed with realized profit/loss. However, there are many factors that influence the profitability of the algorithm. This thesis applies machine algorithms that consist of Artificial Neural Network, Logistic Regression, and XGBoost to predict the profitability from lagging indicators from the trading records. The methodology of the thesis aims to tune and maximize the performance of machine learning algorithms such as feature selection, standardization, GridSearchCV, etc. The result of using trained machine learning is quite satisfactory. The scores from implementing the machine learning on out-of-sample are mostly higher than 60%, meaning that the models are capable of predicting the profitability of signals from lagging indicators. The cumulative profit or the balance curve from using machine learning is significantly higher than the balance curve from normal pairs trading strategy.