Statistical Arbitrage Trading using an unsupervised machine learning approach: is liquidity a predictor of profitability?

Abstract

We test a statistical arbitrage trading strategy, pairs trading, using daily closing prices covering the period 2000 – 2019. Stocks are clustered using an unsupervised machine learning approach and cointegrated stocks from each cluster are then paired. The strategy does not prove to be profitable on S&P500 stocks once adjusted for transaction costs. Conversely, the strategy appears to be profitable on the OSE obtaining annualized excess returns of 22% and a Sharpe Ratio of 0.84 after adjusting for both explicit and implicit transaction costs. We investigate whether a difference in the liquidity can explain why the strategy is more profitable on OSE, and provide evidence suggesting that pairs trading profits are closely related to the liquidity of the stocks traded.