Portfolio Optimization with Artificial Intelligence: A Neural Network Approach to Maximizing Risk-Adjusted Returns

This master thesis focuses on developing an AI model designed to generate profitable investment strategies that outperform market benchmarks. Specifically, the model constructs a portfolio of 500 US equities, determining optimal positions and allocations for each stock. The goal is to maximize returns while minimizing risk and transaction costs. The decision to focus on portfolio optimization, rather than individual assets, stems from the ability to reduce risk through diversification. This is further enhanced by introducing an additional input parameter, λ, which accounts for the investor's risk aversion. Building on the work of Zhang, Zohren, and Roberts (2020), the model leverages multiple data sources to create meaningful features. A careful process of feature selection and hyperparameter tuning was undertaken to identify the most effective configuration. The model itself is a custom neural network implemented in TensorFlow, with four input types: stock returns time series, financials time series, risk aversion λ, and previous day allocation. It minimizes a custom loss function inspired by Modern Portfolio Theory (MPT), balancing the trade-offs between maximizing returns, minimizing risk based on λ, and keeping transaction costs low. The strategies generated by the model were benchmarked against the market and validated using statistical significance tests. Ultimately, the model successfully reduces systemic market risk and tailors investment strategies to suit the unique needs of individual investors.