Related Links

• How Data Revisions and Uncertainty Affect Monetary Policy
• Stocking Up: What Inventory Levels Say About the Economy

Transcript

Tim Sablik: My guest today is Thomas Lubik, a senior advisor in the Research department at the Richmond Fed. Thomas, welcome back to the show.

Thomas Lubik: Thanks so much, Tim. Thanks for having me back. Always a pleasure to be here.

Sablik: Fed officials rely on economic data to help guide their policy decisions. But these data don't always provide a complete or fully accurate picture of the economy. This was especially true when economic data reporting was disrupted by the federal government shutdown last fall.

But even outside of shutdowns, macroeconomic data have always been subject to revision, meaning the first picture of the economy that comes in may not be completely accurate. Thomas, what does that revision process look like for an indicator like GDP, for example?

Lubik: It may be surprising to the listeners, but a lot of the macroeconomic data are getting revised quite often and quite a lot. GDP is actually a prime example for this. Any initial GDP number that is reported tends to get revised four or five times.

To use an example, we are in the first quarter of 2026. So, as of now, we don't really have any picture of how the economy is performing at this point in time. We will get the first picture in terms of the GDP number that the Bureau of Economic Analysis produces only in late April. This is called the advance estimate of GDP — one month after the quarter has ended. Then, a month later — so, getting into late May — we will get the first official estimate. This gives a much more precise picture of where the economy is. Then, there's a final estimate, and this usually comes a few weeks later but more towards the end of Q2, so almost a full quarter later than the first quarter. So, our knowledge of first quarter GDP is quite delayed.

But it doesn't end there. Annual benchmark revisions are typically in September, where the Bureau of Economic Analysis looks at the methodology, gets better data, and updates the revisions. They tend to be smaller than the previous revisions. Then, about every five years, [there is] another benchmark revisions where the BEA — Bureau of Economic Analysis — either updates the methodology [or] brings in new data, better data. A prime example for this is that, almost two decades ago, the BEA started counting intellectual property as an investment good in the investment component.

All in all, the first GDP number that we see is not the final number, and the differences actually can be quite stark. A good example is the second quarter of 2025. The advance estimate was 3.0 percent [while] the final estimate was 3.8 percent. That's a quite substantial upward revision.

Sablik: Yeah, almost a percentage point bigger.

You already touched on this a little bit, but why are macro data revised so frequently?

Lubik: It's basically two reasons. First of all, macro data, as opposed to financial data, are trying to measure something that is actually very hard to measure. You can just observe the price data or financial data — we've got stock prices. But GDP — gross domestic product — is the total sum of the value added produced over a specific time period in an economy of all goods and services. So how do you count this? Ultimately, you have to count widgets. You have to count consumption. You have to count exports, imports, inventories — and we talked about inventories in our last conversation. All of this has to be collected. This is a substantial and enormous data collection effort.

It also depends on the frequency of the data reporting. For instance, if GDP were only collected once a year, then the statistical agencies have a year to get all the data together. Well, a year is a really long time to get the data right, but decisionmakers — whether monetary policy decisionmakers or businesses — want to have much higher frequency reports on GDP. Because of the difficult measurement problems, this clashes against our need to know more, hence all of these revisions.

Sablik: We'll come back to those policy questions in a moment. I want to talk a little bit more about this revision process. Are there any key macro indicators that don't undergo this revision?

Lubik: Yes. The classic example is unemployment. We receive monthly unemployment numbers and the U.S. unemployment data [are] based on a survey. The Bureau of Labor Statistics calls up participants in that survey and asks them, "During the last month, have you been employed and, if not, what have you been doing during this time?" If you've been looking for a job, then you are counted as unemployed. These responses will not be revised. They will report it as is because it doesn't really make much sense to go back to that specific respondent and ask them, "Do you remember what you were doing a few months ago?"

Similarly, the consumer price index is also considered final and will not be revised.

Sablik: You mentioned the size of the revision to the recent GDP numbers. You examined the size of these macro data revisions overall in a recent Economic Brief article that we'll link to. How big are these revisions on average?

Lubik: Some can be strikingly large. What we've computed in the Economic Brief is a measure of revisions. We took the final data number and compared it to the advance estimate.

Over the full data sample — from the mid-1960s up until now — the average size of the revision in GDP is 0.45 percent, so almost half a percentage point. And it's positive, which means that any initial number that is being reported is likely to be understated because it will be revised upward on average by half a percentage point. That's a pretty big number.

We also looked at revisions in the PCE Price Index. The revisions are much smaller there. They are on the order of, over the full sample, 0.08 percent. This is almost a rounding error.

In terms of the employment revisions — the number of jobs that are being created every month — the average revision is 17,000 jobs. So again, it's also a little bit understated, the initial estimate relative to the final estimate.

Sablik: So, unemployment is not revised, but the job numbers are revised?

Lubik: Yes, because they come from two different surveys. This is one of those examples of how the type of data collection affects whether there will be revisions or not.

Sablik: You mentioned that you looked at a sample going back to the 1960s. Has the size or frequency of revisions changed over time?

Lubik: Yes, that's also one of the more fascinating aspects of the data revisions. What we found in this Economic Brief is a break in how the data revisions have behaved in the early 1980s. So, typically, we found a break in 1983.

For GDP, the average size of the revisions, and also the volatility of the revisions, has changed considerably. Before 1983, the average revision was 0.7 percentage points, whereas for the second half of the sample it was 0.35.

We don't find any differences in samples in PCE inflation. That is always a fairly stable series. But the volatility has declined — the size of the revisions has not changed, we see fewer outliers.

Interestingly, the same pattern applies for revisions in employment, where the size of the revisions and the average size of the revisions have declined since 2000. The statistical agencies — Bureau of Labor Statistics, BEA — get better at doing things, and I think this is the underlying drive of why these revisions have become smaller.

Sablik: Did you find any patterns to the revisions? You mentioned with GDP, the average is a positive or revision up. But are there systemic biases up or down?

Lubik: I made the point that GDP revisions have been positive on average, so the initial estimate understates the strength of the economy. But I think this is just an artifact of the difficulty of data collection.

Ideally, the data revisions should be zero. In the case of GDP, they are not zero on average. [With] PCE inflation, they are. What we also want to avoid is that they are systematically, continuously going in one direction. What we found is that for GDP revisions, yes, revisions tend to be positive, on average, on the upside. But in the quarter afterwards, the revision tends to be negative and smaller. This is a good sign. It basically means that the initial data release will be overstated, but the error is likely to be corrected.

This is not the case for employment. If a revision is high in one quarter, it is very likely that it continues to be high in subsequent quarters.

Sablik: Now we come to the big policy question for us here at the Fed, which we alluded to earlier in the conversation. How should policymakers approach setting monetary policy in real time, knowing that the data that they're looking at will be revised?

Lubik: It's a bit of a frustrating question for a policymaker. Luckily, economists have thought about this topic for a really, really long time. More than 50 years ago, in 1967, Bill Brainard of Yale University — incidentally, one of my dissertation advisors — wrote the classic research paper on this and it established what is often called the Brainard principle.

Very loosely speaking, the Brainard principle says that if, as a policymaker, you don't quite know what is going on, just move very carefully. Don't be aggressive. If you don't move aggressively, the likelihood of making a mistake will be smaller. I think this is the basic principle that underlies monetary policymaking under uncertainty, and specifically under data uncertainty.

Sablik: Are there any risks or downsides for policymakers attempting to adjust for future data revisions by moving carefully like you mentioned?

Lubik: Yeah, this is where it gets tricky and the devil is in the details. Recent research has shown that the Brainard principle is not a universal principle. Things can go horribly wrong.

This boils down to the issue that this Brainard principle holds in the models that we use only in the case of parameter uncertainty. The policymaker knows the data, so there is no revision in the data process. It knows the structure of the economy, so how the economy generally works. But the policymaker does not know how specific actions like interest rate increases translate into the borrowing costs that consumers face. They don't know how responsive consumer behavior is to interest rate changes. In a scenario like this, then the Brainard principle applies. But this is technically not the scenario that we have to confront with data uncertainty, with revisions.

What could go wrong? Suppose the policymaker does not know the true model of the economy, how the economy works. In order to get a better understanding of the monetary transmission mechanism, economists estimate a model of the economy. If the data that the economists use to estimate their model are not the true data — and we know the data will be revised and they're likely not the true underlying data of the economy — then the estimates will be biased. Then, the policymaker has a wrong understanding of how the economy really works and may pursue wrong policies.

In a paper that I wrote with my longstanding co-author, Christian Matthes, we laid out the case for this. We also wrote a shorter article together with Tim about this. We basically showed that the 1970s Great Inflation can be precisely explained through this mechanism. The Fed did the best with what they could do, but the data were just bad. Their assessment of the economy, based on what they had to work with, just happened to be the wrong one.

Sablik: And, as you mentioned earlier, the revisions in that period were greater.

Lubik: It was just a very unfortunate period in terms of policymaking and also these enormous revisions.

There's also a second issue that can go wrong with data revisions. This is based on much more recent research, as economists have developed a better understanding of what uncertainty means. If the policymaker is aware that the data will be revised, one way to respond to this is use sophisticated statistical methods to separate the signals from the noise, to extract what the true value of these data are. You do have some indications — you know that, on average, GDP is likely to be revised upwards by half a percentage point. So, maybe you want to take this into account. We receive a GDP number of 3 percent, so you might assess that, well, maybe it's 3.5 percent.

Where it gets tricky is that in a situation like this, no one knows the truth — neither the policymaker nor the economists nor financial markets. So, all of the participants in the economy then try to make this assessment. This can create expectation spirals that don't have anything to do with the underlying fundamental data but is based on a dispersion of views of what the true state of the economy really is. This can lead to fluctuations based on beliefs.

The fascinating thing is that when market participants, policymakers, households, businesses, all try to extract their information from the data that they know are not the true data and will be revised, the optimal way of conducting policy is to respond to the data as they are. You don't run into the issue of creating these self-fulfilling expectations.

Sablik: Very interesting. So, responding to the data as they are, would that be then following the Brainard principle or something? It would depend ...

Lubik: It would not be the Brainard principle, so you would not try to attenuate. You just see GDP growth at 3 percent [and] respond as if it were 3 percent.

I should add a big caveat on this. Uncertainty has many facets. I'm uncertain about things. You are uncertain about other things. Policymakers are uncertain about yet another set of things. So, the way we think about optimal policymaking under uncertainty really is conditional on how we model this uncertainty. But on the plus side, this gives economists lot of work to do in the future.

Sablik: Good job security. [Laughs] Thomas, thanks so much for joining me today and a really interesting discussion about your work.