It’s important to remember that GDP data has been revised, especially with some nowcasts pointing to negative second-quarter growth. The first version is called the “advance” version because most of the input is estimated. Below are the different years of GDP, starting after the trough to better highlight the most recent correction.
figure 1: GDP, various years, in billions Ch.2012$ SAAR. Source: BEA via AFRED.
The errors between the advanced and third editions are shown in Figure 2 (expressed as log differences; so 0.0002 means a revision of 0.02%).
figure 2: Revisions between Advanced and Version 3, in journal form. Source: BEA, author calculations.
There are always revisions.The previous discussion of the G-7 GDP revision was hereClearly, contrary to earlier studies, these corrections tend to be positive, although whether the difference is statistically significant I haven’t checked.
Note that in the second quarter of 2020, there was a huge positive revision from the premium version to the third version – more than half a percentage point on a logarithmic basis. This is understandable given the turmoil associated with this period. However, the fact that GDP fell sharply in the second quarter of 2020 raises the question of dealing with seasonality beyond that quarter.It’s not something statisticians don’t know – see Luca and Wright (2021) also U.S. Bureau of Labor Statistics (2022).
This brings us back to what the different seasonal adjustment methods show us. Figure 3 shows the official BEA series, the official seasonally unadjusted series (all latest years), and my special seasonally adjusted series, using the period 2002-2022Q1, using log-transformed Census X-13 and additional outliers for 2020Q2.
image 3: GDP, BEA-reported seasonally adjusted (bold black), not seasonally adjusted (tan), and seasonally adjusted using the X-13 log transform, allowing for additional outliers in Q2 2020 (red) . The NBER uses shades of grey to define the peak and trough dates of the recession. Source: BEA, NBER and author’s calculations.
Note that my seasonal adjustment process does not replicate BEA’s. BEA seasonally adjusts the entire sample of individual components before summing. I’ve applied a seasonally adjusted filter to the aggregated series from 2002-22, which is the period for which I have nsa data. So my point is not that GDP growth was more properly measured as positive in the first quarter. Rather, it emphasizes the flaws of over-interpreting early releases on an aq/q basis. (In fact, by July 28, the 2022Q1 estimate will be revised in the annual benchmark revision.)
