Identifying optimal investment strategies with deep reinforcement learning

Abstract

The rise of fully automated trading systems has transformed global financial markets, placing greater emphasis on intelligent data-driven decision making. This thesis explores the development of optimal investment strategies using Deep Reinforcement Learning (DRL), with a particular focus on the Proximal Policy Optimisation (PPO) algorithm. Historical closing price data from a diversified portfolio of seven technology stocks was collected, processed and combined with market indicators to form the model inputs. A supervised learning baseline was first established using a Multilayer Perceptron (MLP) to provide a performance benchmark. Subsequently, DRL agents that incorporate different neural network architectures, MLPs, Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN), were implemented within a custom PPO framework designed for multiple stock portfolio management. Each model was evaluated using two state representations: normalised closing prices and normalised engineered market features, enabling a comparison of model performance under varying input dimensions. The evaluation was carried out using Monte Carlo rollouts in a custom simulated trading environment using 2023 test data. The PPO agent with an MLP architecture and engineered features achieved the most stable returns, averaging a gain of 152%, while the CNN-based agent with closing price-only input reached a maximum return of 265% but with a higher volatility. These results suggest that, when the models are appropriately structured and trained, DRL agents can outperform both traditional supervised learning approaches and passive strategies in simulated markets, offering a promising foundation for further research into adaptive algorithmic portfolio optimisation.