Apreçamento de debêntures sem liquidez usando aprendizado por máquina

Resumo

The objective of this study is to propose an alternative methodology for pricing illiquid debentures using machine learning techniques. The research compares the predictive performance of linear regression models, neural networks (RN), and gradient boosting decision trees (GBDT). For the empirical application, we employ cross-sectional regressions in order to more accurately capture the partial effects of the regressors at each point in time. Additionally, we implement the Upper Confidence Bound (UCB) reinforcement learning algorithm to enhance computational efficiency. The analysis is based on debentures indexed to the IPCA and DI, issued by companies with publicly available financial statements and with indicative rates disclosed by ANBIMA, over the sample period from March 2021 to September 2024. The results show that the GBDT model outperforms the others, with out-of-sample R² ranging from 68.9% to 99.2%, and an average of 92.0%. We find that machine learning methods can complement traditional mark to market approaches by providing greater speed and precision in the pricing of illiquid assets. Furthermore, the proposed approach enables large-scale implementation, reducing both operational costs and the subjectivity inherent in conventional valuation methods.

Ficha do documento

Tipo

Dissertação

Ano

2025

Instituição

Fundação Getulio Vargas

Fonte

Repositório da FGV

Idioma

Português

Acesso

Acesso aberto

Identificador

oai:repositorio.fgv.br:10438/37284

Temas

Economia

Palavras-chave

Debêntures IlíquidasRede neuralÁrvore de decisão com gradient boostingAprendizado por reforçoUpper Confidence BoundIlliquid debenturesNeural networkGradient boosting decision treeReinforcement learningEconomiaDebênturesPreços - DeterminaçãoRedes neurais (Computação)Arvores de decisãoAprendizado do computador