COMPARING THE PREDICTIVE PERFORMANCE OF ARIMA, VAR AND VEC MODELS IN COINTEGRATED TIME SERIES
This work aims to compare the predictive power of the univariate ARIMA models with the multivariate VAR and VEC models under the cointegration hypothesis. The analysis was condutec on Monte Carlo simulations using statistical methods to evaluate the predictive power. A model was estimated using exchange rate (USD/BRL) and 12-month cumulative inflation (IPCA) to check the Monte Carlo simulation insights. The model's predictive power on out-of-sample data is similar to Monte Carlo simulations' 95% confidence intervals. The conclusion of the exchange rate and inflation data experiment was in line with simulations. The predictive power of the models was similar within the confidence intervals of the mean. However, the size of the confidence intervals demands more complex validations before choosing the modeling specification for prediction. The results are also in line with market best practices, where even knowing the appropriate theoretical model specification, we should pay attention to out-of-sample performance experiments for greater assertiveness in predictions. The work concluded that even in a set of cointegrated series, where there is a guide for modeling VEC in the econometric literature, to get more assertiveness of predictions, it is necessary to do out-of-sample experiments outside the sample to choose the model with greater predictive power.