• RSS Feed

The LinkUp Blog The Industry's Best-Kept Secret

December 3, 2011 / Toby Dayton

Learning to Use An Alethiometer

There were a number of interesting aspects of today’s jobs report from the Department of Labor related to each of the following:

1) LinkUp’s forecasting model

2) Flaws in the government’s numbers

3) The state of the jobs picture in the U.S.

In reverse order, I’ll start with today’s report. The Employment Situation Report issued today reported that 120,000 nonfarm jobs were created in the U.S. in November, a number that includes a gain of 140,000 private sector jobs and a loss of 20,000 government jobs. With the decline of total unemployed Americans falling from 13.9 million to 13.3 million, combined with a drop in the labor-force participation rate and an increase of 133,000 ‘Marginally Attached, Discouraged’ people who have stopped looking for work and are therefore not counted as unemployed, the unemployment rate dropped from 9% to 8.6%. And finally, the BLS revised their jobs numbers for both September and October. The initial September jobs report claimed that 103,000 jobs were created in September. In October, BLS revised this number to 158,000, and this morning, it was revised again to a gain of 210,000 jobs in September, a 104% increase from the original number. Similarly, the BLS initially reported that 80,000 jobs were added to the U.S. economy in October, a number that today was revised up to 100,000.

It is the magnitude and frequency those revisions to the government’s numbers that point to serious flaws in the Department of Labor’s monthly jobs reports. Over the past 11 months, there have been 17 such revisions to the government’s numbers. I have no idea what happens between the initial report and the subsequent reports 30 and 60 days later, but ┬áthe initial release of the ‘Employment Situation’ report is now so consistently revised in future months as to have rendered itself virtually meaningless when it is first issued. Not only do those revisions make it difficult to accurately assess what is going on with the nation’s jobs picture, but they have wreaked havoc on our forecasting model at LinkUp.

The forecasting model we have built around the jobs data from our job search engine is essentially based on 2 factors: the change from one month to the next in new and total job openings indexed from 22,000 company websites throughout the country (which we obtain from our own job search engine), and the job gains or losses reported by the BLS for a given month. Without question, a job opening is the best leading indicator of a future hire, and we can use our data from a given month to predict how many jobs will be added or lost in the following month. And because the LinkUp search engine only indexes jobs from company websites, we have the cleanest data set of any jobs data set in the country. (There are no old jobs, no duplicate listings, and no job scams, phishing jobs, work-at-home garbage, or fraudulent posts. And certainly most appealing, we don’t have any job listings posted by mass-murderers). We then use the ups and downs of new and total job listings on LinkUp for a given month to predict the jobs gained or lost for the following month based on the reported job gains or losses reported by the BLS for the prior month. This is not as complicated as it might appear, and I’ll use the past few months to better explain how our model works.

In June 2011, new and total job listings on LinkUp rose by an average of 5.7% from May. For the past 6 months, the number of days that companies have kept a job listing on their corporate websites has averaged 27 days, so it makes sense that June’s increase in job openings indexed by our search engine would translate into job gains in July as reported by the Labor Department. And indeed, the BLS report for July reported that 127,000 jobs were added to the U.S. economy in July. But to forecast the positive 127,000 jobs in July, our model relies on the relative differences between the job gains or losses reported by the BLS from one month to the next. To predict what would happen in July, for example, we started with the 20,000 jobs that were gained in June (as reported by BLS). Based on our model, we estimated that given LinkUp’s 5.7% increase in job listings in June, the +20,000 in June would go up somewhere between 100,000 to 150,000 jobs, and it actually went up 107,000 to the 127,000 jobs reported by BLS. (That’s the delta BLS in the table above, and I wish I knew how to find/use the symbols font in WordPress so I didn’t have to spell out ‘delta’).

In July, new and total job listings on LinkUp declined by an average of 0.6%, so our model predicted that in August there would be a decrease from the +127,000 from July of somewhere between 0 to 100,000 jobs. The actual decrease was 23,000, as the BLS reported that in August the U.S. economy added 104,000 jobs. Job gains for August were positive, but size of the increase was 23,000 less than the size of the previous month’s increase, right in line with what our model predicted.

In August, new and total job listings on LinkUp rose by an average of 4%, so we predicted that job gains in September would rise by at least 100,000 from the 104,000 jobs gained in August. And again, our model worked perfectly as the BLS reported that the U.S. added 210,000 jobs in September (an increase of 106,000 from the 104,000 increase in August). Unfortunately, the 210,000 jobs gained in September were not reported until this morning, 60 days after the initial September jobs report issued on the first Friday of October. The initial jobs report for September claimed that the U.S. added only 103,000 jobs that month, a decrease of 1,000 jobs from the 104,000 jobs gained the prior month. So in that first week of October, it looked like our model was dead wrong. But then the September number was raised to 158,000 in early November, and our model looked a little better. This morning’s reported increase in job gains during September to 210,000 has proven that our prediction for September was flawless.

The 60-day revision for September by the BLS also proved that our October forecast was accurate as well. In September, new and total job listings on LinkUp declined by an average of 10.9%, so our model indicated that the 210,000 jobs gained in September would drop to a gain of only 70,000 jobs in October. The initial BLS report for October reported a gain of 80,000, but that number was revised to a gain of 100,000 jobs in this morning’s report. My guess is that that number will be revised down in next month’s report, but in any event, we were pretty close.

All of this is a long, and perhaps terribly muddy way of saying that the initial numbers reported by the BLS are horrendously unreliable of late and that our forecasting model is fantastically accurate. (And I can assure you that neither of those statements are made cavalierly). It is also an overly wordy preamble to the following statement: There is absolutely no way that the numbers reported this morning for November are accurate. The 10.1% decline we saw in LinkUp’s index in October indicates that the rate of increase in job gains should have declined in November by at least 100,000, meaning that even if one assumes that the +100,000 number for October will remain unchanged next month, there were zero jobs created in the U.S. in November. And because we are henceforth abandoning our reliance on the ‘noisy’ initial numbers reported by the BLS, that is going to be the starting point for our next forecast, not the flawed +120,000 reported by the BLS this morning.

Like Lyra’s brilliant but gradual mastery of the alethiometer, so too are we improving our use of the powerful jobs forecasting tool we’ve built.