We study machine learning models for stock return prediction in non-stationary environments
and identify a fundamental nonstationarity–complexity tradeoff: more complex models reduce
misspecification error but require longer training windows that exacerbate non-stationarity.
We address this tension with a novel tournament-based model selection procedure that jointly
optimizes the model class and training window that adaptively evaluates candidates on nonstationary validation data. Our theory shows that the method balances misspecification, estimation
variance, and non-stationarity, performing close to the best model in hindsight. Applied
to 17 industry portfolios, the approach improves out-of-sample R2 by 14–23% relative
to rolling benchmarks, delivers markedly stronger performance during recessions, and yields
trading strategies with 31% higher cumulative returns.