Research

We investigate machine learning models for stock return prediction in non-stationary environments, revealing a fundamental nonstationarity-complexity tradeoff: complex models reduce misspecification error but require longer training windows that introduce stronger non-stationarity. We resolve this tension with a novel model selection method that jointly optimizes model class and training window size using a tournament procedure that adaptively evaluates candidates on non-stationary validation data. Our theoretical analysis demonstrates that this approach balances misspecification error, estimation variance, and non-stationarity, performing close to the best model in hindsight.

Applying our method to 17 industry portfolio returns, we consistently outperform standard rolling-window benchmarks, improving out-of-sample $R^2$ by 14--23% on average. During NBER-designated recessions, improvements are substantial: our method achieves positive $R^2$ during the Gulf War recession while benchmarks are negative, and improves $R^2$ in absolute terms by at least 80bps during the 2001 recession as well as superior performance during the 2008 Financial Crisis. Economically, a trading strategy based on our selected model generates 31% higher cumulative returns averaged across the industries.

We develop a nonparametric method to aggregate firm characteristics across a large supply chain network to explain cross-sectional expected returns. Each firm receives a pricing signal, nonlinearly constructed from the characteristics of neighboring firms within d-hops on the network. We find that $d = 3$ – encompassing network effects up to the third order – balances bias reduction from higher-order relations against variance from added complexity. Our model leads to a portfolio sorted by ML-driven firm-level estimated returns that condition on both historical supply chain data and firm characteristics. We achieve over a 16% out-of-sample Sharpe gain vs direct-link models, and outperform the Fama–French five-factor and PCA benchmarks. We find that the ML-managed portfolio improves mean-variance efficiency, measured by Sharpe ratio. Lastly, we show that the conditional mean return estimation of more central firms is 55% more sensitive to missingness of supply chain links compared to that of peripheral firms in the supply chain graph.

We propose Panel Multiple Testing that allows us to select covariates that explain a large cross-section with false discovery control. In our empirical asset pricing study, we select sparse risk factors from a factor zoo of 114, to explain 243 doubly-sorted portfolio excess returns.