PAIRS TRADING: COMPARISON BETWEEN MACHINE LEARNING AND TRADITIONAL APPROACHES ON INVESTOR RETURNS

Abstract

Pairs Trading is a statistical arbitrage trading strategy that matches a long position in an asset with a short position in another asset. This strategy aims to profit by taking advantage of the price deviations between two assets. It is based on the idea that the spread of related assets will eventually revert to their historical trends. In recent years, as the strategy gained popularity, the arbitrage profits that can be made using traditional methods, such as Cointegration, Distance and Stochastic Spread have been declining. As such, in light of the larger role of Machine Learning in the industry, it will be apt to examine its ability to optimise pairs trading. In this paper, DBSCAN, K-Medoids and Fuzzy C-Means were applied to optimise pairs formation among NASDAQ stocks. The performance of each model was compared against the Baseline clustering using industries. It was found that when trade volume, historical prices and firm fundamentals were used as input features, DBSCAN performed the best with an average Fully Invested Return of 3.9%, a Maximum Drawdown of -22.3% and a Sharpe Ratio of 0.253. It was also found that the better performing pairs tend to be formed from between related industries.