We overcome the low performance of standard portfolio optimization by modeling portfolio weights as functions of stock characteristics. Instead of predicting asset returns in a first step and subsequently modeling portfolio weights in a second step, we directly predict optimal portfolio weights. Building on the model proposed by Brandt, Santa-Clara Valkanov (RFS 2009), we find portfolio weights by optimizing on investor’s average utility.
We expand their one-step weight prediction approach by implementing deep neural networks. Our deep learning models utilize a customized loss function that directly attempts to maximize the utility of a power utility investor. The optimization regularizes input features and unveils relevant asset characteristics from a portfolio perspective that differ to their asset pricing counterpart. The resulting portfolios outperform both the benchmark equal weighted portfolio and the Brandt, Santa-Clara Valkanov (RFS 2009) portfolio by a considerable Sharpe ratio increase of up to 28% for the whole testing period. We unveil hidden interactions between asset characteristics from a portfolio perspective and are able to adjust the custom loss functions to account for turnover- , diversification- , volatility- and short sale constraints.