The Impact of Short Sale Costs on Machine Learning Portfolio Allocation Models

This thesis researches the impact of short selling costs on the performance of machine learning portfolio allocation models. It empirically synthesizes two opposing views in the financial literature. On the one hand, the practice of combining stock characteristics increases the gains of anomaly portfolios, while on the other hand, borrowing fees diminish returns from anomaly portfolios. Using a large set of stock characteristics, we create excess return predictions using machine learning techniques, including principal component regressions, random forests, neural networks, and long short-term memory networks. We find that machine learning models, particularly principal components regressions and neural networks, generate significant excess and abnormal returns to the investor. Nevertheless, while borrowing fees are significantly positive, their impact on machine learning portfolio profitability is not substantial enough to fully diminish returns. As such, combining anomalies through machine learning proves tobe an efficient and robust method to generate persistent returns for investors.