Machine learning to establish proxies for investor attention: evidence of improved stock-return prediction

It is widely recognized that limited attention capacity of individual investors affects stock performance. We construct five aggregate investor attention indices for each stock by extracting common information components related to stock returns from various attention proxies using equal-weighted (EW), principal component analysis (PCA), partial least squares (PLS), gradient boosting decision tree (GBDT), and random forest (RF) methods. In a sample of all Shanghai Stock Exchange 50 constituent stocks, we identify two attention indices constructed by machine learning algorithms, RF and GBDT, that provide economically meaningful enhanced prediction of stock returns in both in-sample and out-of-sample periods. Moreover, these indices are negatively related to return volatility. Results suggest the utility of using machine-learning to form proxies of investor attention and reveal the excellent forecasting power of these proxies in asset pricing.