What do we really know about the cross-sectional relation between past and expected returns?