State-dependent stock selection in index tracking: a machine learning approach

We focus on the stock selection step of the index tracking problem in passive investment management and incorporate constant changes in the dynamics of markets into the decision. We propose an approach, using machine learning techniques, which analyses the performance of the selection methods used in previous market states and identifies the one that gives the optimal tracking portfolio in each period. We apply the proposed procedure using the popular cointegration technique in index tracking and show that it tracks the S&P 500 with a very high level of accuracy. The empirical evidence shows that our proposed approach outperforms cointegration techniques that use a single criterion (e.g., stocks with the maximum market capitalization) in the asset selection.