|
Abstract:
|
Nowcasters often hedge against uncertainty around the choice of model through forecast combinations. However, it is not always sufficient for a practitioner to simply hedge against uncertainty, it can also be useful to characterize uncertainty around a prediction. A decision maker does not just need a point estimate of the current state of the economy, but also the precision of that estimate and the range of possible outcomes and their probabilities. In pursuit of that goal this paper extends the framework of Chernis and Sekkel (2018) to produce density nowcasts for Canadian real GDP growth. We uses a medium-size dataset to generate model level predictive densities from four major classes of nowcasting models. The predictions are weighted by simple average, by their performance (Log Score or CRPS), or through calculating optimal combination weights (Conflitti et al 2015). We study the performance of the models and combinations in a pseudo-real-time out-of-sample exercise. We show that the combined densities can generate useful features for decision makers such as, changes in forecast uncertainty, and asymmetric predictive distributions.
|
