An Evaluation of Self-supervised Learning for Portfolio Diversification

Recently self-supervised learning (SSL) has achieved impressive performance in computer vision (CV) and natural language processing (NLP) tasks, and some early attempts are made in the area of finance. In this paper, we apply SSL to extract features from financial time series data, and use those features to measure the similarities between assets in the market. As similarity measurement is the key to portfolio diversification, we consider two portfolio optimisation problems: index tracking (IT) and minimum variance portfolio (MVP), with the additional diversification terms linked to different similarity measurements, which are sourced from different SSL algorithms. Both IT and MVP are both convex optimisation problems with deterministic solutions, therefore the performance difference is traced back to SSL algorithms, rather than other factors. Extensive experiments are conducted with eight SSL algorithms, and the analysis of the results of the experiments demonstrates the advantages of SSL over non-SSL alternatives.