Deep learning for enhanced index tracking

Abstract

We develop a novel deep learning method for the enhanced index tracking problem, which aims to outperform an index while effectively controlling the tracking error. We generate a dynamic trading policy from a neural network that accepts a set of features as inputs. We design four blocks in the neural network architecture to handle different types of features, including regimes of the index and stocks, their short-term characteristics, and the current allocation. Outputs from the blocks are integrated into the final output that changes the portfolio allocation. We test our model on several indexes in empirical studies based on real market data. Out-of-sample results reveal the importance of different features and demonstrate the ability of our method in obtaining excess returns while effectively controlling the tracking error, downside risk, and transaction costs.