Quantitative stock portfolio optimization by multi-task learning risk and return

Abstract

Selecting profitable stocks for investments is a challenging task. Recent research has made significant progress on stock ranking prediction to select top-ranked stocks for portfolio optimization. However, the stocks are only ranked by predicted stock return, ignoring the stock price volatility risk — a critical aspect for stock selection and investments. Moreover, they preliminarily attempted to capture the effects of related stocks from a singular relation, disregarding the rich information regarding multiple spillover effects from related stocks and the distinctions in effects among various relations. Thus, we propose a risk and return multi-task learning model with a heterogeneous graph attention network (HGA-MT) to predict stock ranking for portfolio optimization. First, to aggregate the multiple spillover effects of related stocks, we introduce graph convolutional networks to fuse the effects of related stocks in each relation and design an attention network to allocate varying weights to different types of relationships. Second, we use a multi-task learning paradigm to learn stock return and volatility risks jointly. The stock ranking results are calculated by simultaneously considering the risk and return. Thus, Top-K ranked stocks are recommended in the portfolio for the next trading day to achieve higher and more stable profits. Extensive experiments prove that HGA-MT outperforms previous state-of-the-art methods in stock ranking and backtesting trading evaluation tasks.