GDP nowcasts: NY vs Atlanta

In July 2014, economists at the Federal Reserve Bank of Atlanta promoted what would be a popular "nowcast" research product, used to measure real-time growth in the economy.  The nowcast incorporates vital and varied economic Big Data measures daily, mostly calibrated to predict the first (advanced) quarterly GDP reading.  Then in April 2016, statisticians at the Federal Reserve Bank of New York copied their colleagues, producing a rival nowcast metric, leading to curiosity as to why the Federal Reserve consumes resources and our attention, to produce two national nowcasts from the same set of "facts".  As it turns out, the public benefits from seeing probability and uncertainty in action, as these two competing measures provide (sometimes uncontrollably) differing signals throughout a quarterly cycle.  The New York statistical product is more nascent (only 8 irregular readings so far), but already we have some initial empirical insights into these products precision when the markets use them jointly.  With just one of the bank's current Q2 GDP nowcasts, we would have expected the 90% confidence interval of the relatively buoyant 2.45% nowcast to extend on the low-side to 1.0%.  However when we combine the insights from both banks' nowcasts (here 57 days prior to report release), this 90% confidence interval only extends on the low-side to 1.8%.  This low-side on the confidence would have been even lower at 1.5%, if both nowcasts were perfectly identical, correlated random variables!  Which they are not.  At some point still, the accuracy of the unconventional analytical approach of these nowcasts will produce an error larger than would be expected generally between the advance estimate, and the "final" annual revision.  And the Federal Reserve themselves do not consider the nowcasts to be by themselves a superior model most of the time.

To start, see this graphic below which includes all of the paired nowcasts between the two banks, since April 8 (21 days prior).  While Atlanta provides nowcasts daily, in the blue shaded region we see all 3 nowcasts paired between the two banks for Q1.  We notice that the Atlanta nowcast increased through April 29 (0 days prior), while the New York nowcast dropped.  Both converged to ~0.7%.  This was higher than the 0.5% first (advance) reading shown as an "x", and lower than the 0.8% preliminary (second) reading shown as a triangle.  The typical initial error on the day or the advance GDP release is (given on the link above) a lot worse, at ~0.9%.

 

New York started providing nowcasts for Q2, on April 8 (113 days prior).  Atlanta however didn't join until April 29 (92 days prior).  We see this data in the red shaded region, along with the 5 (8-3) later nowcasts from both banks.  For April 29, there was a 1.0% difference between the 0.8% New York nowcast and the 1.8% Atlanta nowcast.  By the most recent nowcast of June 3 (falling only slightly due to the May labor report disaster of 38k with 59k downward revisions to prior months), we see that both nowcasts are near 2.45%.  Over Q2 this time, the New York nowcast has continuously increased.  It is worth noting that on this day New York provided a Q3 "nowcast" as well.  Doesn't forecasting Q3, in Q2, violate the definition of a "now"cast?

 

Given that the time to Q2 release is still fairly high at 57 days, the joint-confidence interval tends to still be large (though a little smaller than at 92 days).  Given only two quarters of joint nowcasts, it is difficult to gauge the shape of the confidence, yet it should (centered as shown below) around the most recent nowcasts. 

 

 

We can imagine that as we progress, towards the Q2 release date, that the low-side of the 90% confidence would fall more along the New York nowcast versus the Atlanta nowcast.  We can see something near 1.8% as this low-side of the growth nowcast, and a risk to market participants who might be expecting a >2% GDP growth.  Which would be an achievement we have not had in a year.

 

And nonetheless, the typical revision from these advanced estimates to the final revision (the annual revisions) is more than 1%, implying just focusing on matching to the advance reading is partly a meaningless given that all the Big Data in the world, between two banks, still can't provide a more accurate read on what is really happening.  And in fact even as they tinker and "evolve" the models it is certainly going to lead to a wild miss at some point.  To offset the celebrated accuracy the Atlanta nowcast had a year ago in 2015 Q1.  Where the release was 0.2%, Wall Street economists expected 1.0% and Atlanta nowcasted 0.1%.