WISE: Wavelet based Interpretable Stock Embedding for Risk-Averse Portfolio Management

Abstract

Markowitz's portfolio theory is the cornerstone of the risk-averse portfolio selection (RPS) problem, the core of which lies in minimizing the risk, i.e., a value calculated based on a portfolio risk matrix. Because the real risk matrix is unobservable, usual practices compromise to utilize the covariance matrix of all stocks in the portfolio based on their historical prices to estimate the risk matrix, which, however, lack the interpretability of the computed risk degree. In this paper, we propose a novel RPS method named WISE based on wavelet decomposition, which not only fully exploits stock time series from the perspectives of the time domain and frequency domain, but also has the advantage of providing interpretability on the portfolio decision from different frequency angles. In addition, in WISE, we design a theoretically guaranteed wavelet basis selection mechanism and three auxiliary enhancement tasks to adaptively find the suitable wavelet parameters and improve the representation ability of the stock embeddings respectively. Extensive experiments conducted on three real-world datasets demonstrate WISE's superiority over the state-of-the-art portfolio selection methods in terms of return and risk. In addition, we introduce a qualitative analysis of the computed risk matrices of portfolios to indicate the interpretability of WISE on the computed risk degree from different frequency angles.