Precificação de ativos utilizando cadeias de Markov e aprendizagem por reforço: uma possível abordagem para previsão financeira

Abstract

Markov chains and Reinforcement Learning methods have emerged as promising tools for modeling complex stochastic systems, especially in financial environments characterized by uncertainty and volatility. In this context, the present work proposes the joint application of these approaches to the pricing and forecasting of financial assets, using Petrobras (PETR4) and Vale (VALE3) stocks as case studies. The developed model integrates the probabilistic dynamics of discrete-time Markov chains with the adaptive capabilities of the Q-Learning algorithm through reinforcement learning, enabling the agent to adjust its buy and sell decisions based on the defined accumulated rewards. The methodology was implemented in the Python programming language and evaluated using the Root Mean Squared Error (RMSE) metric, comparing the simulated results with the real values of each asset over predefined periods and algorithm-defined parameters. The obtained results demonstrated that the model is capable of consistently representing the evolution of asset prices, showing better performance over shorter training horizons depending on the asset. This behavior reinforces the potential of integrating stochastic techniques and machine learning methods for financial forecasting in emerging markets.