Characteristics and the Cross-Section of Covariances

We propose a new, minimally parameterized way of modeling stock-level covariances as a function of firm characteristics. Our model uses a large number of indicator functions to approximate the surface mapping two firms' characteristics to the correlation of their returns. We show that the method performs better than existing methods both in and out of our sample period. In truly out of sample tests, we show that using the information in firm characteristics, we can predict future covariances better than when using common asset pricing models. Correspondingly, we discuss how one can use our model as an avenue for understanding how characteristics relate to stock returns in the context of linear asset pricing models