Transmuting Unequally Spaced Data

Unequally spaced data poses a dilemma on how to aggregate high-frequency variables to model a low-frequency variable. To tackle this quandary, this work proposes to apply MI(xed) DA(ta) S(ampling) (MIDAS), which allows the independent and dependent variables to be sampled at various and different frequencies, to forecast the real GDP growth in Brazil using macroeconomic data. The results show that the restricted polynomial MIDAS specification can outperform the AR(1) and the unrestricted Midas for out of the sample recursively estimated nowcasts. Furthermore, this paper showcases the impact of different pooling schemas to enrich forecast combinations using the Midas framework: only the inverse MSE weighted forecast combinations beat the benchmark under the Augmented Diebold–Mariano test. Finally, the MSE cumulative ratio emerged as a compelling framework to uncover unforeseen swerves. Fortuitously, the cumulative MSE ratio revealed that between 2014Q3 until the end of 2015, the quotient for the monetary base MIDAS model continuously declined. While this behavior might not be related to the “fiscal pedaling”, its trend contributes to the economic policy narrative during those years.