This paper explores the application of deep sequence modeling techniques—such as RNNs, LSTMs, GRUs, and Transformers—to asset pricing, focusing on predicting returns and measuring risk premia. The authors first provide an overview of the development of these models, highlighting their ability to capture complex sequential dependencies in financial time series that traditional econometric models often fail to address. They then conduct a comparative empirical analysis using U.S. equity data, evaluating the out-of-sample performance of various sequence models against conventional benchmarks.

The results demonstrate that deep sequence models, particularly LSTM-based architectures, significantly outperform traditional time-series models in predicting asset returns and estimating risk premia. The study emphasizes the importance of incorporating historical path dependence, which sequence models naturally handle through their recurrent structure. The authors also discuss the advantages and limitations of these models, noting that while they offer superior predictive power, they require substantial data and careful hyperparameter tuning. Overall, the paper provides strong evidence that deep sequence modeling is a valuable tool for asset pricing, with LSTM models emerging as the most effective among the methods tested.