Opinions and observations expressed on this blog reflect the authors' individual experiences and should not be construed to be financial advice. None of the members of this blog are licensed financial advisors. Please consult your own licensed financial advisor if you wish to act on any recommendations here.

Sunday, November 6, 2011

On Economic Troglodytes

One of the things I see on more "populist" forms of financial news commentary is a grave mistrusting of seasonally adjusted data.  For the uninitiated, a seasonal adjustment is a filtering process for volatile data series that have a clear seasonal pattern.  By correcting for the observed historical seasonal patterns, you can get a smoothed look at what the data are without the typical chop.  Good examples of data with significant seasonal oscillations are housing starts (with far more in the spring and summer than in the winter) and employment (with massive amounts of layoffs right after the holiday season and robust hiring in the summer).

However, there is a group of people who don't trust seasonally adjusted data, largely because I don't think that they understand it.  Take this post from Minyanville.  The author boldly dismisses seasonally adjusted jobless claims data as not being "real" and asks that readers look at the non-seasonally adjusted data.  Then, there is a choice quote later:
The actual weekly initial claims data exhibits week to week patterns each year that are consistent, as certain industries tend to add and subtract workers at the same time each year. Rather than smoothing the data to obscure what really happened last week, we can compare the numbers directly with prior years' performance during the same weeks to get an accurate reading of the current trend. Like an optometrist, we can look at small changes and ask whether they are better, worse, or about the same as last year. By carefully evaluating subtle changes, we gain clarity of vision.
As it so happens, that is precisely what the seasonally adjusted data does, however imperfectly.  The reason that, beyond a simply seasonal adjustment, economists like to look at a 4-week moving average for jobless claims is that, even after going through the rigors of a seasonal adjustment, large one-time events can occur like a major corporate bankruptcy, a strike, or a major weather event.  By taking a simple average, you can somewhat smooth this out.  Doing so is, by definition, somewhat backward looking, but it is a tool that you can use if you so choose.  One thing I've learned in my line of work is that there is no one right way to analyze data.  The circumstances may call for a few different ways of looking at it.

What some of the troglodytes like to do is use a 12-month moving average or something like that instead of a seasonally adjusted number, claiming that actually represents the real data.  A neat little trick here is that the 12-month moving averages of the non-seasonally adjusted data and the seasonally adjusted data come out almost exactly the same, which is what you would expect if the seasonal adjustment is worth its salt.  See this below with housing starts data:

However, 12-month moving averages don't capture "real time" changes in the data.  It's for much the same reason that year over year comparisons are useless.  If you had a huge run up in the early months of the 12 month period, the leveled off and are now starting on a downward trend, you won't catch it in the moving average or the year over year numbers for another few months.  Look at the three measures of potentially judging what housing starts were doing during the housing boom and bust of the last decade (CLICK ON PICTURE FOR LARGER IMAGE)

The seasonally adjusted data catches the inflection point earlier and more decisively than either of the other two measures.  In the other two, you can eventually see it, but the seasonally adjusted data provides the much clearer signal.  This is because, with the seasonal filter, you can look at a current month and judge what it means on a "real time" basis rather than being dependent on backward looking measures that take months to provide a signal.  Also, compared to non-seasonally adjusted numbers, which at best rely on a year on year comparison, you can much more easily detect the trend.

This is why, even though seasonally adjusted numbers are not "real" numbers, they do provide the best picture of what is going on of all the data that get presented.  Frankly, people who reject seasonal adjustments as being some statistical creation are nothing but troglodytes.

No comments:

Post a Comment