Deep Reinforcement Learning Applied to Investment Portfólio Management

• Grant number: 23/13484-8
• Support Opportunities: Research Grants - Innovative Research in Small Business - PIPE
• Start date: July 01, 2024
• End date: March 31, 2025
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computer Systems
• Principal Investigator: Roberto Fix Ventura
• Grantee: Roberto Fix Ventura
• Company: TEKTON INDUSTRIA, COMERCIO E SERVICOS LTDA
• CNAE: Outras atividades de serviços financeiros não especificadas anteriormente
• City: São Paulo

Abstract

This project aims to develop Artificial Intelligence (AI) models and algorithms, specifically a particular branch of AI, Deep Reinforcement Learning (DRL), for automated investment portfolio management, primarily focused on stocks and other assets traded on stock exchanges. This makes this service accessible to a broad spectrum of investors at a low cost. Investment management is complex, requiring simultaneous analysis of vast volumes of data and taking into account a variety of factors, including systemic and macroeconomic risks.DRL combines Deep Learning, which uses artificial neural networks to identify complex patterns, and Reinforcement Learning, where an agent improves its decisions through interaction with the environment and a system of rewards and punishments.DRL technology is particularly suited for investment management due to its ability to handle dynamic environments, learn from reward/punishment feedback, manage high-dimensional data, and make long-term decisions. Furthermore, DRL can adapt to ever-changing financial environments, learning through trial and error.The resulting system from this project will be divided into four main components: DRL models and infrastructure, Brokerage integration, Front-end for investor access, and Back-end for user management and other aspects. Phase I of the project will focus on the development of DRL models, which are the most complex part of the system. Once successful, the other components will be developed in Phase II.Constructing a DRL model offers numerous possibilities and variations, requiring experimentation with a large number of computational models. The models will be evaluated based on criteria such as returns (absolute, relative, CAGR), risk (volatility, beta, Sharpe and Sortino indices), diversification (asset allocation, sectoral, geographical), drawdown, and liquidity.The most successful DRL models will be ranked to be applied according to investors' profiles, considering their investment objectives, investment horizon, risk tolerance, and liquidity needs. This will ensure that the developed solutions meet the specific needs of different investors, offering a tailored approach to investment portfolio management. (AU)