Scalable Financial Index Tracking with Graph Neural Networks

Zepeng Zhang, Ziping Zhao

Abstract: As a prevailing passive investment strategy in the financial world, index tracking aims at replicating or surpassing the performance of a financial index. The core part of an index tracking strategy is to design a sparse index tracking portfolio (ITP) from a basket of candidate financial assets. In this paper, a scalable two-stage approach is developed for ITP design under the minimax criterion, which consists of an asset selection stage (i.e., to select a subset of the assets from the index constituent stocks) and a capital allocation stage (i.e., to allocate the capital among the selected assets). The asset selection problem is tackled via a well-calibrated graph neural network (GNN), followed by a light-weight linear programming problem for capital allocation resolved via a standard solver. The idea proposed in this paper is novel for the area of ITP design in that it is especially scalable for tracking large-scale and dynamic-updating financial indices. Numerical simulations validate the scalability and high-efficiency of the proposed GNN-based approach with comparisons to the standard solver-based approach.