Post hoc, ergo propter hoc?

On the occasion of the Fed’s Jackson Hole Symposium, the New York Sun published an editorial attacking central banking and fiat money. Let’s get this out of the way at the start: we are big fans of both. In our view, the world is a more stable and prosperous place with central banks than it was without them. And fiat money allows a central bank to stabilize the price of goods and services that would be quite volatile if, instead, we chose to steady the price of gold (the Sun’s apparent favorite). The result is higher growth from which we all benefit.

We also like tabloids. They’re fun.

Our main problem with the Sun’s piece is its all-too-common mode of argument. While essentially stating that correlation does not establish causality, the editorial goes on to do just that. In this case, the facts marshaled are that the U.S. unemployment rate was lower on average during the Bretton Woods fixed-exchange rate regime (from 1947 to 1971) than since.

It isn’t much of a challenge to find phenomena that are correlated with either secular or cyclical movements in economic activity. For example, sunspots appeared far more frequently during the Bretton Woods period (85 per month) than after (63 per month). Sunspots are those seemingly darker areas on the sun that are actually cooler regions of intense magnetic activity (for details, see here). Did the fixed-exchange rate regime have such cosmic influence?

How about spurious cyclical patterns? The following picture is our entry into the post hoc, ergo propter hoc U.S. business-cycle sweepstakes. [The Latin term is a common logical fallacy (y followed x in time, so x caused y) that is a frequent issue in economics.] The gray bars denote recessions as dated by the NBER. The left edge of each bar is the peak of a cycle, while the right edge is the trough. We have plotted this along with the monthly sunspot count. (You can read about sunspot cycles here. Ironically, economists use the term sunspot to refer to a phenomenon that is both random and completely arbitrary.)

Number of Sunspots

 Note: Shaded areas denote U.S. recessions. Sources:    NASA     and     NBER   .

Note: Shaded areas denote U.S. recessions. Sources: NASA and NBER.

From 1950 to 2014 there have been 10 recessions – that is, 10 instances of business-cycle peaks followed by troughs.  As it happens, in six of these 10 cases, economic slowdowns coincided with high levels of sunspot activity. Granted, a 60% success rate isn’t so great, but it isn’t bad, either. Anyone who can get recession timing right that frequently would be hailed as an economic oracle.

To continue, we can compare sunspot activity as a business cycle predictor to the predictive power of the stock market – something that truly is connected to the real economy. We find that 12-month declines in the S&P500 index anticipate more than 1.5 recessions for each one that occurred over this 65-year period (see chart). That is, sunspots have a problem with false negatives – they miss nearly half the recessions – while the stock market’s problem is false positives (it anticipates recessions that didn’t occur).

S&P500 Index (Percent change from a year ago)

  Note: Shaded areas denote recessions. Sources: Shiller    website    and    NBER   .

Note: Shaded areas denote recessions. Sources: Shiller website and NBER.

What are we to make out of all of this? People are programmed to look for patterns. It’s crucial to our survival that we have the ability to find patterns and deduce cause and effect in our surroundings. But the hardwiring that serves us well when we are managing the mundane daily tasks of eating or crossing the street, has an important shortcoming. We tend to err on the side of seeing patterns where none exist. And, with the advent of fields like data analytics – the study of enormous, ostensibly disorganized masses of numbers – we now have more places to look and bigger, faster computers to help us do the looking.

In a post earlier this year, we expressed doubt that big data would materially improve our measurement of GDP. Here, our concern is with the mechanical hunt for correlations. Given the sheer amount of data and computing power that we now have at our finger tips, finding things that move around together will be even easier than it used to be (when it wasn’t much of a challenge). The trick is to find correlations that are durable because they reflect economic fundamentals like technology, physical endowments, and preferences. Doing that means having a theory for why various data should be related.

Perhaps a few of the correlations revealed by big data will result in new and useful theories. That is the hope. The danger is that people will end up relying on spurious relationships – like thinking that they need to consult astronomers to forecast business cycle downturns. Some astronomers may be good economic forecasters (one of us is an avid stargazer), but if they are, it’s not because they are looking through a telescope to count sunspots.