A major problem in any discussion about long-term trends is that after just a few years, even the most sophisticated projections are really not much better than guesses. For making this simple observation — which is obvious to anyone who has given it any though — I have been given a lot of grief and accused of “insulting” the fine work done by economists at places like CBO and elsewhere. The CBO itself includes a whole section of spreadsheets explaining its error problems.

Within 5 years, the CBO’s can only be 90% confident of predicting our budget deficit/surplus within a range of +/- four percent of GDP.  That is a huge error.  Just massive in the space of five years. Virtually any projection system is going to be able to get within that range.  Their projections of GDP are only slightly better.  Why?

Well, CBO, in particular labors under specific limitations.  CBOs projections are not self-sufficient predictions.  They are legislation scoring tools.  So, even though anyone paying attention to tax policy over the past decade knew we’d have annual adjustment to the Alternate Minimum Tax (AMT), CBO would never assume that because their job was to score existing legislation, not make independent predictions of the future.

But, beyond the CBO-specific problems, what most people don’t understand is that the vast majority of the “data” used by macroeconomists is not “data” at all in the sense of observation of incidents.  It is, instead, almost wholly inferential.  What is “labor productivity”?  It is a calculated measure, but there is no way to directly measure it.  Even in cases where you can actually conceivable measure something, such as gross domestic product, you can’t actually measure it in practice.  Yes, there is, indeed, a single, observable answer to the question, “what is the total value of all domestic transactions of goods and services,” but we don’t actually have a table of all such transactions which is then totalled up.  Macroeconomic statistics are not hard data, but rather are inferred or estimated data.

So, you have measurement error.  But beyond the measurement error, you have estimation and inferential errors.  Let’s think of labor productivity — which a key element in the projections of the Social Security Trustees in assessing the status of that program.  To come up with an estimate of future labor productivity growth, you need to predict, inter alia, output and work hours for given products and industries, estimates of the distribution of those products within industries and industries within the broader economy.  You need to adjust for qualitative and quantitative variances in outputs.  And  you cannot measure any of those things directly.  You can do some survey work and hope you can build a model for the rest of the economy.

The best hope is that all the errors you make are random and that you can there apply some sort of error estimate and create a fan diagram of likely outcomes.  The problem, of course, is that this approach masks the basic uncertainty rather than highlighting it, because the casual response to such a range of predictions is to average them and use that as a basis of policy recommendations.  But the average was just the initial uncorrected estimate.  So, think about the Social Security trustees report (which I will discuss in more detail shortly), which has a low, intermediate, and high cost estimate.  Everyone uses the intermediate…. because… well, it is in the middle… and well… if you average the high, low, and intermediate, you come up with, surprise, surprise the intermediate outcome.  But that point of the three predictions is not to average them, it is that given the predictable errors margins in the model, the three outcomes are all just as likely.

But unfortunately, errors are not random.  There are often chained together in difficult to predict ways.  There are all sorts of hidden correlations and just as bad, false correlations.  When I went to college, existing macroeconomic dogma held that the frictional unemployment rate was 6%.  So, any model that considered the possibility of less than 6% unemployment also assumed that such a condition would result in strong inflationary pressures.  The result, of course, was lower-than-actual estimate of economic growth, because if you modeled the economy growing fast, your assumptions forced you to modify your price deflator to adjust nominal economic growth down.

So, in short, economists working on long-term projections are smart, dedicated, technically proficient individuals, but they are working under conditions of tremendous uncertainty — data uncertainty and modelling uncertainty.

Which is why, for anything more than 3-5 years, you can make predictions of broad trends using relatively simple methods that will turn out to be just as accurate as any other.  Indeed, to the extent that your model makes predictions that are at odds with simple linear trendlines, you need to defend the model, not just assume the trendline is irrelevant.