The trajectories of the BLS vs. ADP private nonfarm employment series differ, even though for much of the year, the ADP series was above that of BLS.
Figure 1: Private nonfarm payroll employment from BLS (blue), and from ADP (brown), and CES March preliminary benchmark revision (dark blue triangle), all on log scale. Light green shading denotes data to be benchmark revised. Source: BLS, ADP via FRED, BLS, and author’s calculations.
The divergence in changes is even more marked if one zooms in on 2019.
Figure 2: Change in private nonfarm payroll employment from BLS (blue), and from ADP (brown). Source: BLS, ADP via FRED, and author’s calculations.
Keep in mind the 90% confidence interval for private NFP sampling error is about +/- 100K.
What can the latest divergence tell us? From Elad Pashtan, “The ADP Employment Report: Pay Attention to Large Surprises,” Goldman Sachs Economic Research, September 2016.
In sum, our findings suggest that the ADP employment report has only limited value in forecasting the CES, as much of its marginal information content appears to come from other publicly available data such as jobless claims that are incorporated by ADP into their official figures. However, large ADP forecast errors have more often than not been followed by CES forecast errors in the same direction, suggesting that forecasters might be able to improve their performance by partially adjusting their payroll expectations when such a large ADP surprise occurs.
Bloomberg consensus on the ADP release was for 140K, vs. reported 67K, so a surprise on the downside; for the BLS release, it was 175K vs. reported 254K, so the surprises went in opposite directions this time.
The 2019M11 figures are actually quite an outlier (in growth rates, not surprises). This is shown in a scatterplot over the 2013-19 period (a revamp of the ADP series was undertaken in 2012).
Figure 3: Log differences in BLS series against ADP series (blue circle), and regression line (red). Source: ADP, BLS via FRED, and author’s calculations.
It’s important to note that the ADP series is a composite. Once again from GS:
To produce their final estimate of payroll growth, ADP uses not only their raw proprietary data but also lagged information from the official BLS report and the Philadelphia Federal Reserve’s Aruoba-Diebold-Scotti Business Conditions Index (ADSBCI), a high frequency indicator that blends information from seven economic releases, including initial jobless claims and industrial production, among others.
Some recent work by Cajner et al. (2019) suggests ADP microdata can provide additional information that can be used to reduce the measurement error. This point is illustrated in this graph of different series and vintages of series around the last recession, from the paper.
Note that the series they examine — they dub it ADP-FRB — differs in nature substantially from the ADP series discussed above.