Essays in Financial Econometrics

Advisor: Márcio Garcia

Co-advisor: Marcelo Medeiros

This thesis is composed of three essays on econometrics applied to finance. In the first chapter, I investigate the normality of standardized stock returns using a sample of nearly 11,000 U.S. equities spanning approximately 20 years, constructed with various realized measures of volatility. The results indicate that return distributions deviate from normality as illiquidity increases, and that the use of robust volatility measures is not sufficient to overcome this deviation. In the second chapter, I build on the findings from the previous chapter to study the problem of volatility forecasting across this broad set of stocks. I compare the performance of GARCH and HAR models in forecasting Value at Risk (VaR), a risk measure that does not directly depend on realized volatility estimates. The results show that the former model exhibits significantly higher accuracy than the latter in most cases, especially for less liquid stocks. The third chapter, in turn, presents a review of the most recent advances and challenges in natural language processing methodologies, with a focus on extracting information from text applied to economics. It explores a new emerging paradigm involving the use of LLMs in economic analysis. Still in this chapter, I conduct an inflation forecasting exercise aimed at exploring potential synergies between state-of-theart language models (LLMs) and other machine learning methods such as LASSO and Random Forest. The results suggest that, on their own, textual data may not outperform methodologies based on already well-established and influential datasets, such as FRED-MD, but can still marginally improve results when used in a complementary manner.