Factor-Based Portfolio Optimization

We show that a parsimonious factor model can alleviate the problems of using raw historical data subject to large idiosyncratic noise in mean-variance portfolio optimization. Through the factor structure, we incorporate forward-looking information into the expected returns, exploiting a set of variables known to predict the aggregate factor from prior studies. Estimating parameters in a manner consistent with the theoretical assumption that the mean and variance correspond to future return distributions leads to better out-of-sample performance, accentuating the importance of feeding proper information