Doug, thanks for your questions.

The technique is not really all that original. My inspiration was the
article explaining the birth/death model, "Impact of business births and
deaths in the payroll survey," in the May 2006 Monthly Labor Review. The
birth/death model does indeed "work pretty well" annually, as evaluated from
comparing the B/d adjusted estimate with the actual benchmark revision. But
if you look at the monthly numbers, there's a remarkable seasonal variation,
with April being the high point of the year for adding in "births" and
January and July actually recording "negative births" (or could they be firm
deaths in excess of reported deaths?)

My point being that the seasonal variation of the b/d imputation exceeds the
seasonal adjustment factor of the overall employment number. And, in a
sense, there's little the BLS can do about reconciling this because the data
they use for modeling is QUARTERLY, not monthly. Well, actually, the BLS
could go to a quarterly birth/death adjustment but then their reports
wouldn't be as timely. It's all a question of balancing the timeliness of
the data reporting with its quality.

Again, if you're evaluating the accuracy of the results on the basis of an
ANNUAL benchmark, it hardly matters that your QUARTERLY data are slightly
skewing your MONTHLY reports. Got it? It's all in the different time frames.
If I'm wrong here about how BLS evaluates accuracy, then I can only attrbute
it to my source, a BLS report titled "How the Business Birth/Death Model
Improves Payroll Employment Estimates". One can only assume that the BLS
evaluates accuracy the same way they report the evaluation.

My use of the word "prediction" is tongue in cheek. All I'm doing is a
linear trend projection based on the empirical observation that the January
to June seasonal variation is pretty regular. As you point out, January is a
month of heavy layoffs, followed by a pretty steady seasonal run-up in jobs.
>From year to year the slope of the January to June unadjusted numbers
(taking January as the seasonal trough and June as the seasonal peak) may be
different each year but within each year it's a fairly straight line. Such
steadiness makes sense if you think about employment being a relatively
stable process. A seasonally adjusted decline in employment sometimes refers
to an unadjusted increase in employment that was , however, less than the
usual increase for the month.

My prediction for May is thus based on two things: a linear projection from
the January to April seasonally-unadjusted trend (such as can be observed in
the 1992-2008 data) and the conclusion that "nothing happened in April" to
signal a turning point in employment, that is to say the apparent
deceleration of job loss in April to 539,000 from 699,000 in March and
681,000 in February was a statistical hiccup and not a turning point. And
some of that hiccup gets added to the May number.

Just to be perfectly clear, my 'prediction' for May is based on the
assumption that nothing has changed, fundamentally, and nothing will change
this month. If you take the average of the April report and my May
prediction, you get -657,000 to -687,000 -- pretty much in line with the
reported March and February numbers.



On Sun, May 10, 2009 at 12:56 PM, Doug Henwood <[email protected]> wrote:

>
> On May 10, 2009, at 3:25 PM, Sandwichman wrote:
>
>  May BLS Non-farm payroll employment decline: -775,000 to -835,000 jobs
>> lost.
>>
>> http://econospeak.blogspot.com/2009/05/prediction.html
>>
>
> Those are some original techniques you're using there, Comrade Sandwich.
>
> You say:.
>
>  Another thing I noticed is that seasonally unadjusted January to June
>> employment numbers consistently display a steady slope, with the
>> seasonally-adjusted employment figures intersecting that line at April. So
>> far this year, the January to April segment of that line has been flat, with
>> a minor dip down in February and March followed by a small bounce in April.
>> That small bounce might reflect the 62,000 census workers plus the residual
>> unadjusted seasonal variation from the birth/death model.
>>
>
> The BLS has been working on the birth/death model for a long time, and it
> works pretty well. If you think you've discovered a problem with it, then
> maybe you should let them know.
>
> But I don't get your point at all about the slope of the unadjusted
> numbers. January is normally a month of heavy layoffs, so the SA takes out
> almost 3 million jobs that month at current employment levels. Feb-June are
> normally months when SA adds jobs - a little over 3 million all together.
> Over the course of a year, the seasonal factors cancel out. But if you look
> at the pattern for the first half of the year alone, you might find some
> sort of slope.
>
> For example, if unadjusted employment were unchanged at April's NSA level
> of 132.295 million, you'd get the following changes after SA (in thousands):
>
> Jan           +2,870
> Feb             -649
> Mar             -692
> Apr             -734
> May             -751
> Jun             -402
> Jul           +1,218
> Aug             -123
> Sep             -423
> Oct             -724
> Nov               +8
> Dec             +269
>
> Actual monthly changes, so far this year:
>
>       SA       NSA
> Jan    -741    -3,615
> Feb    -681    -  164
> Mar    -699    -   84
> Apr    -539    +  241
>      -----    ------
> tot  -2,660    -3,622
>
> How does this suggest anything about what might happen in May?
>
> Doug
>
>
>
> _______________________________________________
> pen-l mailing list
> [email protected]
> https://lists.csuchico.edu/mailman/listinfo/pen-l
>



-- 
Sandwichman
_______________________________________________
pen-l mailing list
[email protected]
https://lists.csuchico.edu/mailman/listinfo/pen-l

Reply via email to