Motivation

Portfolio optimization is a core challenge in finance. Traditional models like Mean-Variance Optimization rely on assumptions that often break down in real markets. In contrast, Reinforcement Learning (RL) offers a flexible, adaptive approach to sequential decision-making under uncertainty — making it well-suited for dynamic financial environments. Recent advances in deep RL (e.g., DQN, PPO, SAC) present new opportunities for building data-driven investment strategies. However, applying RL in finance still poses key challenges in risk control, stability, and real-world applicability.

Thesis Objectives

This thesis aims to explore how Reinforcement Learning (RL) can be applied to portfolio optimization and evaluate its effectiveness compared to traditional methods. The work will focus on the following main objectives:

• Designing an RL Environment for Financial Markets:
  Define the states, actions, and reward functions needed to model portfolio optimization as an RL problem. This includes using historical asset prices, technical indicators, and other financial signals as inputs and defining rewards based on returns, risk, and transaction costs.
• Implementing and Testing RL Algorithms:
  Several state-of-the-art RL algorithms (such as DQN, PPO, and SAC) will be applied to the portfolio optimization problem. These methods will be trained and tested on historical data, and their performance will be compared to classical portfolio strategies like equal-weighting or mean-variance optimization.
• Using Real Financial Data for Evaluation:
  The models will be evaluated on real US, European, and/or cryptocurrency market data. Special attention will be given to how the models perform during market stress or unusual periods, such as the 2008 financial crisis, the COVID-19 pandemic (2020–2021), and the geopolitical and economic events starting in 2022.
• Analyzing Robustness and Interpretability:
  The thesis will also examine how stable and reliable the learned strategies are when market conditions change. In addition, simple interpretability tools will be used to understand the RL agents’ decisions and evaluate which input features are most important.