Empirical asset pricing and ensemble machine learning

Many of the sophisticated models for stock return forecasting and portfolio optimisation cannot beat naive equal-weighted models. The challenge is that, even in the age of big data, there are usually more potential variables than is appropriate for estimation. This thesis is dedicated to improving asset pricing models via ensemble machine learning methods without requiring more data. By introducing two ensemble methods, first, several representative sophisticated models of stock return forecasting are compared based on standard economic variables in the literature. The results show that both of the two ensemble methods could significantly improve these sophisticated models and found that these models can significantly outperform the equal-weighted combination of individual predictors. Their forecast gains stem from better performance during periods of market uncertainty and crises, and increased diversity and built-in shrinkage. Then, I introduce a general boosting framework for high-dimensional portfolio optimisation, where the classical mean-variance portfolios cannot work properly. The results indicate the effectiveness of these boosting methods in both low- and high-dimensional settings and they can outperform the 1/N portfolio in terms of several popular performance metrics.