We here at Wonkblog aren't the only ones obsessed with the monthly jobs reports. There's a whole cottage industry around predicting what the monthly numbers are going to say, and a major resource for prognosticators is the National Employment Report from the payroll processing firm ADP.

Usually released a day or two before the Bureau of Labor Statistics report, the ADP surveys 270,000 clients and asks how many people they have on payroll. That's slightly more businesses than the Current Employment Statistics survey, upon which the BLS number is based, contacts for its initial estimate (although the businesses in BLS's sample are larger than the ones in ADP's). Because of the size of the sample, it's taken seriously as an early indication of how the jobs market is doing.

That said, ADP and BLS have diverged of late. ADP predicted a gain of 162,000 jobs in September, well over the 114,000 estimate that BLS released. ADP's August report predicted 201,000 new jobs, compared to BLS' initial estimate of 96,000. Then again, BLS is often wrong, as I explained Monday. And a look at the numbers suggests that ADP has, since the recession started, done a better job.

I compared ADP and BLS's initial estimates to the final "benchmark" estimates of the BLS. BLS uses tax data to put together a final estimate of jobs gained, one which is generally more reliable than survey data.

From January 2001 (the first month for which ADP has numbers) to November 2011 (the last month for which benchmark numbers are available), BLS generally did a better job. It was off by an average of 68,229 jobs, compared to 71,405 for ADP. BLS generally undershot, predicting on average 10,290 fewer jobs would be gained than actually were, while ADP overshot by 11,145 jobs on average.

So far so good. But since the recession started in December 2007, ADP has had the upper hand. It's been off by 69,857 jobs on average, while the BLS has been off by 78,694. Much of that is because ADP was closer to the real scale of the jobs crisis during the recession, whereas BLS kept underestimating its scale:

 When the lines are above the X axis, they're too optimistic. Their jobs numbers were higher than the actual ones. If they're below the X axis, they're too pessimistic. Throughout 2008, BLS kept underestimating the number of jobs that were lost, while ADP, while not perfect, got much closer. This isn't to say it's perfect - it whiffed big-time in June 2010, getting the real number wrong by close to 400,000. But it's done better overall.

This doesn't mean you should ignore the BLS estimates every month. But it does suggest that we should be paying at least as much attention to ADP.