Intervalos de predição para o mercado de ação

Abstract

In the stock market, high-risk decision-making processes require both good forecasts and a rigorous quantification of uncertainty. To handle the uncertainty inherent in forecasts, we can use prediction intervals. To address such uncertainty, this thesis uses both parametric prediction intervals and those obtained through conformal prediction methods. Conformal prediction allows us to transform point forecasts from any model into prediction intervals with non-asymptotic guarantees and without strong assumptions about the data distribution. In classical conformal prediction, uncertainty of a classifier is typically quantified under the assumption of exchangeability (e.g., i.i.d. data). However, this assumption is not satisfied for financial time series samples. In this case, it is necessary to use conformal prediction methods that account for the dependency nature inherent in time series data. In this work, the conformal prediction methods considered were those proposed by Xu e Xie (2021) and Gibbs e Candes (2021). We used both the width of the conformal intervals and those obtained from parametric models to construct stock portfolios. The results indicated that, in contexts where the assumptions of the parametric model are met, the parametric method outperformed the conformal method in terms of performance, leading to portfolios with higher returns. Simulated examples where such assumptions are not satisfied are presented to illustrate the interval prediction capability of the conformal methods compared to the parametric intervals.