Portfolio optimization: from markowitz to machine learning

Project Work presented as the partial requirement for obtaining a Master's degree in Statistics and Information Management, specialization in Risk Analysis and ManagementIn the past few decades, substantial progress has been made in portfolio optimization, especially with the emergence of machine learning. Therefore, it is essential to find the models that not only achieve the best results but also simplify the process. This project aims to demonstrate that to achieve optimal portfolios cannot be based only on traditional statistical methods. Therefore the Random Forest regression model, a machine learning model, was chosen to predict stock prices to complement the Markowitz model, a classical portfolio selection model. To evaluate the efficacy of the modified model compared to the classical model the following methodology was adopted: data was collected (from 2012 to 2019 from 10 companies and it was divided in 15 periods) and treated; some common technical indicators were extracted; one stock price was predicted per period; expected returns and partially estimated volatility were derived from the predictions and introduced in the classical model; 15 portfolios were constructed by each model; and finally, a performance analysis was conducted. The results obtained show that the 1-day predictions were quite accurate, almost 90%, and the modified model’s portfolios’ outperformed the classical model’s portfolios for most periods analyzed