Estudo e comparação das técnicas de validação cruzada desenvolvidas para séries temporais

Abstract

Testing the generalization capacity of an algorithm obtained is crucial for any prediction methodology, for which cross-validation methods were developed. However, when dealing with data that have dependency among observations, as in the case of time series, the usual cross-validation methodologies are not appropriate. Currently, there is no single and usual way to test the generalizability of a predictive model for time series. In this work we will study and compare four different variations of methods for cross-validation, this will be done through simulations of different time series.