Reinforcement learning for sequential decision-making: a data driven approach for finance

This work presents a variety of reinforcement learning applications to the domain of nance. It composes of two-part. The rst one represents a technical overview of the basic concepts in machine learning, which are required to understand and work with the reinforcement learning paradigm and are shared among the domains of applications. Chapter 1 outlines the fundamental principle of machine learning reasoning before introducing the neural network model as a central component of every algorithm presented in this work. Chapter 2 introduces the idea of reinforcement learning from its roots, focusing on the mathematical formalism generally employed in every application. We focus on integrating the reinforcement learning framework with the neural network, and we explain their critical role in the eld's development. After the technical part, we present our original contribution, articulated in three di erent essays. The narrative line follows the idea of introducing the use of varying reinforcement learning algorithms through a trading application (Brini and Tantari, 2021) in Chapter 3. Then in Chapter 4 we focus on one of the presented reinforcement learning algorithms and aim at improving its performance and scalability in solving the trading problem by leveraging prior knowledge of the setting. In Chapter 5 of the second part, we use the same reinforcement learning algorithm to solve the problem of exchanging liquidity in a system of banks that can borrow and lend money, highlighting the exibility and the e ectiveness of the reinforcement learning paradigm in the broad nancial domain. We conclude with some remarks and ideas for further research in reinforcement learning applied to nance.