“Negative capability … is when a man is capable of being in uncertainties, mysteries, doubts, without any irritable reaching after fact and reason…” John Keats, 1817.

Will growth be high in the future? Or, will it be low? Everyone is preoccupied with this question in one way or another. Individuals worry about whether their real incomes will grow. Fiscal policymakers wonder what will happen to tax revenues and required expenditures. And central bankers are trying to figure out whether they should ease or tighten monetary policy.

In the end, what everyone usually needs is a forecast of average growth over the next few years – trend growth. Unfortunately, it turns out that we are not very good at forecasting trend growth. And that uncertainty makes decision making – for households, businesses, and policymakers – full of risk. John Keats, the British romantic poet, thought that great artists exhibit special “negative capability” in the face of uncertainty. Evidently, we all need a touch of the artist.

To highlight the problem, we will examine the evolution of historical projections of what is called potential GDP. In some economic models, potential GDP is the level of output that would be achieved if all prices and wages were to adjust flexibly (as we expect they would over the long run). That output is consistent with normal levels of employment and resource utilization. Monetary policymakers expect that when actual GDP is equal to potential GDP, inflation will be stable, so they try to minimize the deviations of actual output from potential. To forecast their budgets for the next decade (something that the U.S. Congress must do) fiscal policymakers assume that the economy will grow at the rate of potential GDP.

So, estimating potential GDP is very important. It is so important that, since 1991, the Congressional Budget Office (CBO) has published multi-year projections of potential GDP. (You can find a description of the CBO methodology here and a comparison with other methods here.) Initially these forecasts were for the subsequent 5 years; since 1996, the projections reach out 10 years. Each of these estimates incorporates a forecast of trend growth – the rate of growth that the economy is expected to achieve after temporary fluctuations fade. So, by looking at the vintages of these projections made in “real time,” we can see what government officials thought trend growth was going to be for every year for the past quarter century. That is, we can look at what their 1991 estimate was for growth through 1996, at the 1992 estimate for growth through 1997, and so on.

To show how these projections evolved, we constructed the following chart using data from the St. Louis Fed’s real-time database: the Archival Federal Reserve Data (ALFRED). The solid black line is the most recent estimate of potential GDP published in January 2015. The red line is actual GDP (in chained 2009 dollars) also from last month. The other lines – the spokes coming off the black line – are past estimates of potential GDP. (See the technical note at the end of this post for details.) Vintages published in the 1990s are in shades of tan and orange, while those in the 2000s are in blue.

This picture is quite striking. During the 1990s, revisions were consistently upward. That is, the 1992 estimate of potential GDP was below the 1994 estimate, which was below the 1996 estimate, and so on. Over the past 15 years, that pattern reversed: revisions have been consistently downward. The 2001 estimate is above the 2006 estimate, which is above the 2008 estimates. (To keep the chart readable, we show only 11 of the 40 available vintages.)

Chart 1: Selected Vintages of U.S. Real Potential GDP

This pattern, both the earlier upward adjustments and the more recent downward revisions, is indicative of a gradual learning process adjusting to a persistent change of trend that is difficult to detect. Remember, the people constructing these estimates neither knew the underlying trend, nor did they get to see the estimated trend over the full 25 years in the picture that we have here. All they knew at the time of the projection was that actual growth differed from what they had anticipated – higher in the 1990s and lower in the 2000s. They also made real-time estimates of the inputs into GDP and of the response of inflation to the gap between measured GDP and estimated potential GDP. As they accumulated more information over time and extracted the trend signal from that noisy information, they slowly (and steadily) adjusted their beliefs about potential GDP.

It is interesting to note that the recent downward revisions have been larger than the earlier upward ones. We can see this in the second chart, where we focus on the potential GDP estimates from 1992 and 2001. In each case, we have extended the line at the average growth rate in the final three years for which the CBO published data. For the 1992 vintage that is a trend growth rate of 2.08%, and for the 2001 vintage, it is 3.15%. The current estimate of potential (the thick black line) is somewhat below the midpoint of the two estimates, while current output (the red line) is roughly twice as far below the trend implied by the 2001 estimates of potential than it is above the one constructed from the 1992 estimates.

Chart 2: Selected Vintages of U.S. Real Potential GDP with forecasts

As the pictures indicate, our uncertainty about trend growth means that estimates of the gap between actual and potential output are subject to sizable error and substantial revision. To give some sense of how big these errors can be, consider the period around 1998. Real-time estimates suggested output was more than three percentage points above potential. Today we believe the gap was closer to zero. There are two reasons that this estimate has changed: revisions to actual GDP and revisions to the estimate of potential GDP. (We discuss the first of these here.) From 1999 to today, actual GDP has been revised up by nearly 3%. Over the same period, potential GDP was revised up by almost 6%. The result is that, over time, an *output gap* (between actual and potential GDP) initially thought to be 3% has disappeared.

The case of the mid-2000s is the opposite, albeit less dramatic. Today we believe that output was roughly equal to potential at the end of 2005. In the real-time estimate, however, output was thought to be at least 1.5% below potential. (And the revision is due almost entirely to downward revisions in potential GDP.)

We draw two lessons from this pattern of revision and learning. First, we should all be wary of anyone who claims to be able to forecast trend growth accurately and reliably. Even after the fact, it takes some time to discern the underlying trend. As a result, we need to build decision frameworks – for businesses and for policymakers – that are robust to the sorts of forecast errors we have seen in the past. Consider that approach the economist’s version of Keats’ negative capability.

Second, our inability to get a precise fix on the output gap presents significant challenges for monetary policy, as this is commonly used as a prime indicator of inflationary pressures in the economy. If central bankers are unsure of the size of the output gap (or even its sign), then the likelihood of policy errors rises substantially. That reinforces the view of monetary policy setting as a problem of risk management in which policymakers must balance the hazards and costs associated with potentially large errors.

*Technical note*: The vintages of potential GDP that are available from ALFRED have different base years. The earlier ones are in 1982 dollars, the most recent in chained 2009 dollars, and those in between use a variety of base years. A proper comparison of the series over time requires that they have the same base year. We cannot make this adjustment precisely, so we approximate it by assuming that the series in a given year has the same value as the current series in the first quarter in which the older series was published.

For example, for the 1992 series published in January of 1992, set the value of potential GDP for the first element in that series equal to the latest 2015 estimate. To be specific, in 1992, the CBO estimated that potential GDP for 1992Q1 was $5,126.6 billion in 1987 dollars. The 2015 estimate for this same quarter is $9328.2 billion in chained 2009 dollars. We multiply the 1992 series by the ratio of these two values, (9328.2/5126.6)= 1.82, and then plot the result starting in 1992.