Stock Portfolio Selection with Deep RankNet

Abstract

A new paradigm of learning-to-rank algorithms is introduced to investigate the statistical relationships between the short-term stock returns and the past return series. The authors construct the stock portfolios in terms of the predicted stock ranking on the cross section and implement the trading strategy on CSI 300 constituent stocks in the Chinese market. Significant outperformance is found for the RankNet algorithm, compared to the corresponding regression and classification methods using neural networks. RankNet methods with deep neural net structures are also proposed and are found to yield promising strategy profitability and comparable ranking precision in relative order predictions. Further studies show that the RankNet strategy returns partly load on common sources of systematic risk, suggesting investment behavior that partially incorporates return-based capital market anomalies.

TOPICS: Security analysis and valuation, portfolio construction, quantitative methods, big data/machine learning, emerging markets

Key Findings

• ▪ Learning-to-rank algorithms can effectively help improve portfolio construction compared to regression and classification methods.

• ▪ The deep rank neural network can robustly capture the statistical relationships between the stock ranking and input features, giving rise to promising profitability in strategy performance.

• ▪ The RankNet algorithm can partially extract several return-based anomalous patterns from the market, as revealed from the factor loadings in Fama–French regression analysis.