The use of Deep Reinforcement Learning for portfolio optimization in the Finnish stock market

Abstract

Portfolio optimization is a fundamental challenge in finance, aiming to allocate assets in a way that balances risk and return. Traditional approaches, such as Mean-Variance Optimization, rely on historical data and strong statistical assumptions, which may limit their adaptability to dynamic market conditions. Recent advancements in Machine Learning, specifically in Deep Reinforcement Learning offer a great alternative, enabling adaptive decision-making in complex financial environments by allowing models to learn optimal trading policies through trial and error. This study evaluates the Advantage Actor-Critic (A2C) model for portfolio optimization within the Finnish stock market. In addition to individual stock data, the model incorporates technical and macroeconomic indicators to enhance state representation, aiming to improve responsive- ness to short-term market fluctuations and broader economic trends. The model is trained on an eight-year period from January 2015 to December 2022 and then tested over a two-year period from January 2023 to December 2024. The performance during the testing period is evaluated relative to benchmark strategies based on cumulative returns and other chose evaluation metrics. While the model demonstrates potential by outperforming traditional benchmarks in a setting without trading fees, its instability during training and inability to adapt to transaction costs raise concerns about its robustness and real-world applicability. The findings highlight key challenges in applying DRL to portfolio management, particularly in terms of policy stability, state representation, and transaction cost sensitivity.