Deep Parametric Portfolio Policies

We generalize the parametric portfolio policy framework to learning portfolio weights via deep neural networks. We find that network-based portfolio policies result in an increase of investor utility of between 30 and 100 percent over a comparable linear portfolio policy, depending on whether portfolio restrictions on individual stock weights, turnover, short-selling or transaction costs are imposed. Gains extend to different utility functions and network structures, and can be traced to a more efficient utilization of the non-linear relationship between firm characteristics and moments of the cross-sectional return distribution