Transformers-based models for portfolio management

Portfolio management, i.e. the management of a collection of investment assets, is the task of making investment decisions based on strategies that ensure maximum profit for each investment period. It is an optimization problem aiming at finding the best actions for selecting the most profiting assets for a time period [1]. This task is challenging because of the difficulties in representing assets price series, since these are non-stationary and exhibit noise and oscillations. Deep learning models have been used to tackle this problem. For example, Long Short-Term-Memory (LSTM) models have been proposed for stock price prediction to construct and optimize portfolios [2]. However, deep models application to portfolios is complicated because of two factors: first, deep learning representation of financial assets prices doesn’t always capture short-term and long-term dependencies in patterns of price series; and secondly, most deep learning models fail to capture correlations among assets. Thus, new models based on Transformers architectures have been proposed [4].  Transformers are networks based on attention mechanisms and represent the state-of-the-art in sequence modeling tasks like machine translation and language understanding [3]. Variants of the Transformers are nowadays employed for financial applications [4].

The goal of this thesis is to study the recent advances in the field of portfolio management and selection, with particular emphasis on the use of Transformers architectures, and to apply these models to a real problem.