Autoencoder asset pricing models and economic restrictions — international evidence

We evaluate the performance of the Conditional Autoencoder (CAE) model by Gu et al. (2021) across U.S. and international datasets, considering economic constraints such as the exclusion of microcap and illiquid firms and the inclusion of transaction costs. The CAE model captures nonlinear relationships between returns and firm characteristics by jointly estimating latent factors and conditional betas while enforcing the no-arbitrage condition. The original study demonstrated significant reductions in out-of-sample pricing errors from both statistical and economic perspectives in the U.S. context. We validate these findings on the original U.S. dataset and show that the model generalises well to a U.S. dataset with a broader set of firm characteristics and to international markets. When economic constraints are introduced, portfolio profitability declines substantially. Profitability drops by 60%–85% when shifting from the full sample to the liquid sample before trading costs. However, after costs, only the liquid strategies remain profitable. In particular, long-only strategies on the liquid sample are the only ones to consistently outperform market benchmarks across all datasets, achieving Sharpe ratios between 0.65 and 0.78 for both equal- and value-weighted portfolios. Overall, the findings underscore both the limitations and the practical potential of the CAE model under realistic market frictions.