Straight Up Statistics: The Magic of Seasonal Adjustments
By AUSTIN NELSON
Have you ever wondered what the heck it means when you read that economic data is “seasonally adjusted?” How can non-seasonally adjusted data show one trend while seasonally adjusted data shows something completely different? Which dataset is the most reliable?
The in-depth answer to these questions requires a PhD in statistical analysis. For those of us who don’t know a kernel regression from a Henderson 13-term moving average filter, the short answer is that seasonal adjustment is a process by which consistent seasonal effects are removed from a time series of data. And yes, you can trust them. Well … sort of.
The effect of seasonal adjustment can be most easily explained through an example. Suppose you are looking at a series of data measuring gasoline consumption in the United States to identify trends related to the price of a gallon of gas. A logical hypothesis is that when gas gets more expensive, people drive less.
In examining this dataset, however, we would expect to see increased consumption in the summer months when everyone hits the road for their vacations. Gas prices often rise during the summer when that additional demand constricts supply, so if you were looking at data from a single year without considering seasonal effects, you might wrongly conclude that people actually consume more gasoline when prices rise.
In fact, much of the increase in fuel consumption during summer months has nothing to do with fuel prices, so the seasonal effects need to be removed from the series before any meaningful analysis of consumption versus price can be undertaken.
Looking at non-seasonally adjusted figures by themselves is a bit like saying pumpkin sales spike in October, without mentioning Halloween.
So how does one “remove” seasonal effects from a dataset? By examining several years of data, patterns in the movement of the data can be identified that happen over and over again in the same way each year. From these patterns, statisticians create (through a variety of near-magical statistical techniques) a “filter” that allows them to subtract the seasonal effects from the dataset of interest, theoretically leaving only non-seasonal effects, like that of price on gas consumption.
And VOILA! you have seasonally adjusted your data. The same techniques are applied all the time to financial and economic datasets, so much so that most people accept this “seasonal adjustment” without thinking twice about it.
Our advice is to think twice about it – especially with housing data.
One of the most common patterns in home buying is that sales tend to slow during the winter months. This makes sense, since moving in the winter sucks, and its easier to move kids from school district to school district over the summer. Now, housing economists — particularly our friends at the National Association of Realtors — are adept at spinning even the worst reports in a positive light.
Data released today showed abymsal existing home sales in November, which should come as no surprise to anyone who’s opened a newspaper in the past couple months. Nevertheless, the Realtors managed to find a silver lining. Chief economist Lawrence Yun “[hopes] the home sales impact from the stock market crash turns out to be short-lived, as was the case in 1987 and 2001,”. If data don’t improve this winter, look for Yun and his crew to start blaming bad weather, snow and a whole host of things that make conditions look better than they are.
The lesson: Never accept data or data analysis at face value.
As my grandfather always said, there are lies, damn lies, and statistics. Unless you can understand how a particular piece of data is derived and can trust the collection and analysis methods that went into its creation, it is as informative as a two-year-old’s fingerpainting.
Seasonal adjustment is no different. Even though almost none of us can understand the mathematical techniques and statistical assumptions that go into the production of official economic figures, you can still look critically at datasets to determine whether they make sense.
In many cases, non-seasonally adjusted data is available along with seasonally adjusted data. Compare the two. Do the changes make sense?
For instance, if all of a sudden non-seasonally adjusted home sales are on par with activity over the summer, one could logically conclude the efforts to unfreeze the mortgage and credit markets may be working. If data bumps along about the same as last year, well, they better get a bigger bailout.
Also think about the source of the data. Does the source have a reason to overly stress or even inappropriately apply a seasonal adjustment to suit their needs? If so, you probably shouldn’t be trusting any data that comes from that source, seasonally adjusted or no. The data source should also have citations for the methods used to complete the adjustment. Even if you don’t know what the citation means, there are those that do and the information should be available to those experts to review.
All this being said, in most cases seasonal adjustment is a completely legitimate analytic technique. Government data has standardized techniques for seasonal adjustment that are well accepted and continually scrutinized. And while many take issue with the government’s collection techniques and even the way they count, say, unemployment, rarely are seasonal adjustments accused of being used to fudge official numbers. Most institutions that put out data reports on a regular basis are also very open about their techniques: These are the ones that can be trusted.
To conclude, with all the sources of data that are available in today’s information age it is becoming increasingly important to develop a healthy skepticism for any particular piece of information. Data is only as reliable as its source and its application.
Tags: data, gas consumption, Henderson, Housing, kernal regression, NAR, Realtors, seasonal adjustment