Latent factor model for asset pricing

Research output: Contribution to journalArticlepeer-review

10 Scopus citations


One of the fundamental questions in asset pricing is ‘Why different assets earn different average returns?’ In this paper, we designed an autoencoder based asset pricing model to explain the return difference among the stocks in an index. The trained autoencoder generates a set of latent representations that constitutes a combined -‘communal’- factor to better explains a large portion of the return differences among the stocks in an index. After analyzing all the stocks in S&P-500, Russel-3000, and NASDAQ-100, we found that our proposed latent factor model outperforms many other factor models in predicting the next day's return. Notably, the experiment results show that on average non-communal stocks earn 0.05% over communal stocks. However, the risk associated with this non-communal stock is also 0.8% higher than communal stocks. The experiments confirm that the superior performance comes from the compensation of high risk associated with these non-communal stocks. Investors will benefit from our latent factor model to identify these communal and non-communal stocks for a high return while diversifying their asset portfolio.

Original languageEnglish (US)
Article number100353
JournalJournal of Behavioral and Experimental Finance
StatePublished - Sep 2020

All Science Journal Classification (ASJC) codes

  • Finance


  • Asset pricing
  • Autoencoders
  • Fintech
  • Machine learning
  • Nonlinear factor model


Dive into the research topics of 'Latent factor model for asset pricing'. Together they form a unique fingerprint.

Cite this