Aprendizado de máquina na precificação de empresas: uma estratégia de investimento

Abstract

Traditional valuation models present significant challenges. Their limitations include bias, which compromises impartiality due to prior information and external pressures; uncertainty, common in startups and transitioning companies; and complexity, intensified by globalization and changes in accounting standards. Additionally, techniques like discounted cash flow may be overly complex and not easily generalizable, quickly becoming obsolete due to their de- pendence on company and time-specific parameters. This work proposes pricing companies listed on the B3 stock exchange and, based on this pricing method, selecting assets to compose an investment strategy. Using machine learning algorithms, the study adopts market capitali- zation (Market Cap) as the dependent variable and explores various models to determine the fair price of companies. The accuracy of the models was evaluated using the Mean Absolute Percentage Error (MAPE) metric, indicating the need for additional data to mitigate overfitting and underfitting. The investment strategy, which selects undervalued companies and rebalan- ces the portfolio quarterly, proved effective in outperforming the benchmark in almost all data groups and algorithms tested. Finally, statistical evidence indicated that there is no relationship between the MAPE metric and the returns achieved by the investment strategies.