Portfolio optimization based on multi-task relationship learning

Among all the portfolio optimization studies that consider uncertainty, it has been shown that by combining different strategies, estimation error can be hedged and better out-of-sample performance can be achieved. In general, these kinds of combining strategies rely heavily on the distribution assumption to decide the combining weights, which may weaken their practical applicability. To improve this, this paper utilizes multi-task relationship learning (MTRL) framework to estimate global minimum variance (GMV) weight and mean-variance (MV) tangency weight simultaneously, by estimating the correlation matrix between these two weights, achieving a combining-like result. In terms of empirical tests, this paper uses daily data of A-share market spanning from 2000 to 2019 to construct two datasets: factor portfolios and stock portfolios randomly collected from HS300 index constituents. The empirical out-of-sample result based on factor portfolio dataset suggests that MTRL-MV can achieve better performance than original MV, equally-weighted (EW) and other combining strategies in terms of Sharpe ratio, standard deviation and turnover. In the end, this paper uses three datasets: The subsample of factor portfolios from Chinese A share and US stock market and individual stock portfolios, to conduct robustness test, which confirms MTRL-MV's improvement compared to other strategies. © 2021, Editorial Board of Journal of Systems Engineering Society of China. All right reserved.