The Pandemic Recession as a Giant Outlier

When I earlier blogged on Frank Schorfheide and Dongho Song (2020), I was focusing on exact methods for mixed-frequency data in Bayes vs. frequentist forecasting and nowcasting.

Quite apart from that, Schorfheide-Song provides eye-opening discussion of a key issue in "forecasting through" the Pandemic Recession (PR), namely how to treat the PR data in estimation. They find that "... forecasts based on a pre-crisis estimate of the VAR using data up until the end of 2019 appear to be more stable and reasonable than forecasts based on a sequence of recursive estimates that include the most recent observations."

The point is that the PR is in many respects a massive outlier, so that one has to think hard about what to do with it in estimation. That is, as always one wants to fit signal, not noise, and the PR is in certain respects a massive burst of noise, capable of severely distorting parameter estimates and hence forecasts and nowcasts.

Michele Lenza and Giorgio Primiceri address the same issue in another fine new paper, “How to Estimate a VAR after March 2020”. Their focus differs in lots of interesting ways, but the message is the same: One way or another, we need to heavily downweight the PR data.

The emerging message is a big deal: One should be careful before attempting to re-estimate forecasting and nowcasting models with data spanning the PR. Of course at this point there remain many open questions, but it's great to see the issues raised.