Zivot, EricLiu, Yuming2023-01-212023-01-212023-01-212022Liu_washington_0250E_24952.pdfhttp://hdl.handle.net/1773/49659Thesis (Ph.D.)--University of Washington, 2022In this dissertation, I aim to forecast U.S. stock returns via state-space and probabilistic deep-learning models. In the first chapter, I propose new state-space models with stock and accounting variables to estimate the expected market returns. These approaches uncover the information existing in unobserved state variables through the predictive updating system based on the Kalman filter technique. The one-step-forward in-sample prediction for the state-space model with stock variables has R-squared as 13%, where the expected market return has a persistent component. I further improve the performance of forecasting market returns by incorporating accounting variables with a state-space model, having a higher R-squared, 18%. Both expected market returns and expected returns on equity have persistent components, but expected returns on equity are more persistent than expected market returns. Results from out-of-sample predictions further reinforce the forecastability of market returns based on proposed models, especially for short-range predictions. In the second chapter, I conduct a comparative analysis of advanced deep-learning models for forecasting U.S. market returns. These approaches uncover more information existing in the same dataset from the first chapter. I present a higher out-of-sample R-squared, ranging from 42.33% to 65.80%, compared with classical time series models. By introducing the family of probabilistic deep-learning models, I reinforce the argument that accounting variables are more informative than stock variables for predicting market returns. Also, I build a link between traditional metrics and advanced neural nets by having an extension of state-space models, the deep state-space model. Innovative deep-learning models simplify estimations of multiple time series and highlight the value of neural nets without losing interpretability. In the third chapter, I conduct an analysis of forecasting the U.S. stock returns based on the probabilistic deep learning methods described in the second chapter. By estimating aggregate-level stock returns, I find that DF-RNN and DeepVAR provide the most accurate results, with out-of-sample R-squared ranging from 63.14% to 76.51%. DSSM precisely estimates firm-level annual stock returns, with a 4.52% R-squared on the testing dataset. Also, by building a zero-net-investment trading strategy, I find that DeepAR and DSSM can help to construct profitable portfolios with cumulative returns ranging from 4.26% to 5.13% on out-of-sample periods. As a result, probabilistic deep learning models can generate state-of-the-art predictions of U.S. stock returns at both aggregate and firm levels.application/pdfen-USnoneBig DataDeep LearningHigh-dimensional Time SeriesProbabilistic Deep Learning ModelsReturn ForecastingState-space ModelEconomicsFinanceEconomicsThesis