Out-of-sample Predictability and Risk of Individual Stocks



We study the out-of-sample performance of portfolios of individual stocks based on panel data models designed to predict the cross-section of returns of individual stocks. Both expected returns as well as exposures to risk factors are modeled as functions of firm characteristics like size, various valuation ratios, momentum indicators, liquidity and industry dummies, with possibly time-varying coefficients based on macro-economic variables like the term spread, credit spread and dividend yield. The out-of-sample predictability is tested on a control sample that contains the three most recent years of returns, which have not been used in the model selection stage. We consider a number of portfolio strategies, which differ along two dimensions:

  1. the holding period
  2. the ex-ante neutrality to systematic risk factors 

Using linear programming we maximize expected portfolio returns while controlling diversification and the expected exposure to systematic risk factors. We find that the best performing stategies are industry neutral long-short portfolios.