Sergio Lago Alves
2017-11-13 - This paper presents a new algorithm, based on a two-part Gibbs sampler with FFBS method, to recover the joint distribution of missing observations in a mixed frequency dataset. The new algorithm relaxes most of the constraints usually presented in the literature, namely: (i) it does not require at least one time series to be observed every period; (ii) it provides an easy way to add linear restrictions based on the state space representation of the VAR; (iii) it does not require regularly-spaced time series at lower frequencies; and, (iv) it avoids degeneration problems arising when states, or linear combination of states, are actually observed. In addition, the algorithm is well suited for embedding high-frequency real-time information for improving nowcasts and forecasts of lower frequency time series. We evaluate the properties of the algorithm using simulated data. Moreover, as empirical applications, we simulate monthly Brazilian GDP, comparing our results to the Brazilian IBC-BR, and recover what would historical PNAD-C unemployment rates look like prior to 2012.