with Frederik Simon & Tom Zimmermann
Management Science, 2026
We consider parametric portfolio policies of any complexity using deep neural networks to optimize investor utility. Risk aversion acts as an economic regularization mechanism, with higher risk aversion constraining model complexity. Empirically, Deep Parametric Portfolio Policies generate 43-102 basis points higher monthly certainty equivalent returns compared to linear policies. Looking beyond expected returns, non-linear portfolio policies better capture the complex relationship between investor preferences and firm characteristics but the benefits of using complex models vary with investor preferences. Results hold across different utility functions and remain robust to transaction costs and short-selling restrictions. Overall, economic regularization constrains model complexity much like statistical regularization but emerges endogenously from investor preferences.
solo-authored
Investors with limited attention learn how information maps into value from the typical patterns they observe across firms, so a firm whose combination of information few others share is harder to process. Using unsupervised machine learning, we measure this atypicality (ATYP) as the part of a firm's information that these typical patterns cannot explain. Empirically, we document that atypical firms attract more dispersed and less accurate beliefs, and that their prices absorb the news in past returns and industry shocks more slowly. Where attention is scarce and short sales are costly, atypical firms are overpriced and earn lower subsequent returns.
with Dieter Hess & Frederik Simon
Journal of Accounting Research, Major Revision
Using the full set of Compustat financial statement variables, we find that flexible machine learning models significantly increase the out-of-sample accuracy of street earnings forecasts for horizons of up to five years, yielding economically meaningful improvements for valuation and asset pricing. Compared to traditional linear models, our approach reduces the mean squared forecast error one year ahead by 28% and generates annualized long-short return spreads of approximately 18% in implied cost of capital portfolios. Moreover, the disagreement between machine learning and traditional forecasts predicts both future earnings and abnormal returns, indicating that the machine learning model captures information not yet reflected in prices. Although income statement variables are the most important predictors at short horizons, balance sheet fundamentals gain substantial relevance as the forecast horizon extends. Finally, a novel decomposition reveals that these predictive improvements stem primarily from modeling non-linear relationships rather than from merely expanding the input set.
with Justus Boeker & Tom Zimmermann
This paper develops a framework for nonlinear parametric portfolio strategies that explicitly incorporates transaction costs. Building on recent advances in machine learning–based asset pricing, we propose a one-step approach that directly optimizes investor utility rather than relying on expected return forecasts as an intermediate target. We analyze both ex-post and ex-ante methods for reducing trading costs and show that simple ex-post adjustments can substantially improve net performance with only minimal efficiency losses.