A Deep Learning-Aided Approach to Portfolio Design for Financial Index Tracking

This paper considers the index tracking portfolio (ITP) design problem in financial markets, which aims at reproducing the performance of a financial index by investing in a subset of the assets constituting it. From a regression-based point of view, the ITP design problem is formulated as a mixed-integer programming (MIP). Leveraging the graph convolutional network (GCN), a calibrated GCN is proposed for asset selection followed by a lightweight MIP problem to realize asset allocation. Numerical simulations show that compared to existing methods the proposed learning-aided approach can generate comparable ITP design results and significantly accelerate the computation which is favorable for practical index tracking targets in finance.