Banca de DEFESA: Felipe Gonçalves Pereira

Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.
STUDENT : Felipe Gonçalves Pereira
DATE: 20/05/2024
TIME: 19:00
LOCAL: https://us02web.zoom.us/j/2395833498?pwd=MWRmUVlDL0MyWWo4dzg3K0d5NEZVdz09&omn=81396258321
TITLE:

Forecasting Inflation in Brazil with Machine Learning Methods: Integrating Shrinkage Method for Variable Selection with Shapley Value Interpretation

KEY WORDS:

Inflation Forecasting, Machine Learning, Variable Selection, Explainable Artificial Intelligence (XAI), Shapley Value

PAGES: 98
BIG AREA: Ciências Exatas e da Terra
AREA: Ciência da Computação
SUBÁREA: Matemática da Computação
SPECIALTY: Modelos Analíticos e de Simulação
SUMMARY:

This dissertation investigates the effectiveness of non-linear machine learning (ML) models in predicting the Brazilian consumer price index (IPCA). By leveraging advanced ML techniques such as Random Forest, AdaBoost, Extreme Gradient Boosting (XGBoost), and Gradient Boosting, we aim to enhance the accuracy of inflation forecasting within an emerging market context. Spanning from August 2010 to January 2024, our analysis utilizes a comprehensive dataset comprising 156 predictors. To optimize model performance, we employ recursive feature elimination (RFE) with ElasticNet, isolating the 30 most influential predictors.

Among the models examined, Gradient Boosting emerges as the most effective, exhibiting superior accuracy indicators. Notably, we enhance the interpretability of this winning model through the Shapley Value, an Explainable Artificial Intelligence (XAI) technique. By elucidating the individual contributions of variables, Shapley Value mitigates the "black box" nature of ML predictions, offering valuable insights into the dynamic interactions between predictor variables essential for strategic decision-making within institutions.

This dissertation underscores the potential of integrating sophisticated ML methodologies with traditional macroeconomic tools to refine forecasting capabilities. It also points towards promising avenues for future research, such as exploring deep learning and hybrid models. Moreover, our findings extend beyond academia, offering practical insights for navigating complex economic landscapes.

COMMITTEE MEMBERS:
Interno - 3164711 - JOAO CARLOS FELIX SOUZA
Interno - ***.366.561-** - JOAO GABRIEL DE MORAES SOUZA - UnB
Interno - ***.535.421-** - PENG YAOHAO - UnB
Externo à Instituição - MATHIAS SCHNEID TESSMANN - IDP