Optimizing Resource Allocation with Intelligent Agents

Abstract

Playing the stock market is one of the many frontiers of applied artificial intelligence. The problem of optimizing assets allocation within a portfolio to yield a return above the market is a hard problem. Current approaches include multi-agent systems in which the task of gathering data, predicting trend and choosing assets are divided between many specialized cooperative agents that simulates the environment of an investment fund. This paper builds upon existing models to create ORACLIA (Optimizing Resource Allocation and Capital Lucrativeness with Intelligent Agents), a multi-agent system for portfolio optimization in the Brazilian stock exchange, Bovespa. A multi-agent architecture was defined and implemented with the use of machine learning classification models to generate buy and sell signals based on the success or failure probability of a predefined strategy. The classification models are able to predict success or failure of a strategy with over 80% accuracy for certain assets. These models successfully replaces the traditional approach that uses regression to predict stock price trends. Backtesting results, conducted in an environment simulated with actual stock market data from 2015 to 2016, shows ORACLIA achieves net results 8x higher than single asset portfolios or general benchmarks such as Ibovespa, the main index for Bovespa.