Forecasting sovereign CDS returns via deep learning

• Ano de defesa: 2021
• Autor(a) principal: Lima, Pedro Antonio Sá Barreto de
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Credit default swap soberano; Deep learning; Previsão para série de tempo; Recurrent neural networks; Redes neurais profundas; Redes neurais recorrentes; Sovereign credit default swap; Time series forecasting
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/96/96131/tde-24092021-154707/

This paper is an empirical exercise of forecasting for a class of credit derivatives known as sovereign credit default swaps. Utilizing non-linear and non-parametric machine learning techniques termed Deep Learning, this study was made utilizing daily emerging country data and a set of financial and macroeconomic indicators as features. The non-linear nature of the financial derivative studied here suggests that this novel technique can better capture data behavior compared to a baseline random walk model. A Grid Search Cross-validation is conducted to estimate the hyperparameters of the model. To evaluate the predictive out of sample forecasting is utilized deterministic and statistical metrics concluding that there is a predictive gain utilizing this technique.