Tom Heintjes: Hello, and welcome back to another episode of the Economy Matters podcast. I'm Tom Heintjes, managing editor of the Atlanta Fed's Economy Matters magazine. Today we're visited by Pat Higgins, an economist here at the Atlanta Fed. Pat is the mastermind, I'll say, behind GDPNow, which is a tool that provides a real-time snapshot of gross domestic product. With all the attention on GDP—well, there's always a lot of attention on GDP, but especially these days—I asked Pat to join us on the podcast to discuss GDP and, more specifically, GDPNow's snapshots of GDP growth, or lack of growth, whatever the case may be. Thanks for being here today, Pat—it's good to have you back on.
Pat Higgins: Thank you, Tom.
The Atlanta Fed's Pat Higgins. Photo by Ted Pio Roda
Heintjes: Pat, so far I've avoided using the word "forecast" because I know GDPNow, as impressive as it is, is not a crystal ball. Am I right to avoid saying "forecast"? And can you distinguish between what GDPNow depicts and a forecast?
Higgins: Thank you, Tom. I would say the terminology around economic prognostications is probably loose enough to allow the numbers generated by GDPNow to be called either "forecasts" or something similar, like "economic projections." I guess one might quibble with calling GDPNow's estimates for a quarter that has recently ended a "forecast," since it is making a prediction about something that has already occurred in the recent past. But the more important distinction is that unlike economist surveys, GDPNow does not incorporate any subjective adjustments based on developments or factors that fall outside of the scope of the model. So it probably does make sense to call the headline numbers from GDPNow either a "model" or a "model-based forecast," or something similar to that. Whichever terminology you prefer, you are right that GDPNow is not a crystal ball. It tends to become more accurate as it approaches the official GDP release date, but that is definitely not always the case.
Heintjes: That's a great distinction—thank you for making it. Pat, can you briefly discuss the mechanics of GDPNow, keeping in mind that you're not talking to a fellow economist? How does it work over the course of a quarter, and what do we typically see over the course of a quarter, in terms of the run-up to the official, final GDP number the Bureau of Economic Analysis puts out?
Higgins: Perhaps the best way—or the easiest way—to convey how GDPNow works is with an example, and so I'll describe how it handles consumer spending—or the official name, "personal consumption expenditures"—which is the largest GDP subcomponent in the sense that it comprised nearly 70 percent of last year's $23 trillion in US gross domestic product. After the BEA releases June data on consumer spending on the last business day of what's currently July, GDPNow will initiate its forecast for third-quarter consumer spending growth. However, the model won't actually have any data at all on consumer spending for July through September of this quarter, and its forecast is going to depend on how it projects the data to come in for each of these three months based on recent macroeconomic data trends and its statistical forecasting equations. Before the BEA releases its estimate of July consumer spending, which will come out on August 26, there will be released data related to that on auto sales at the beginning of the month, and on retail sales, which will come out in the middle of the month. GDPNow will update its forecast of consumer spending based on each of those two data points. Although it won't exactly be able to get those parts of personal consumption expenditures exactly right, it will get a lot closer, typically, than it would without having that data. And that same pattern will apply exactly the same way to August and September—it will get some of the data, except by the end of the quarter it won't actually have services spending or consumer spending when it makes its final forecast in late October. It will have a lot of the data that will come in that's related to consumer spending, but not all of it—and in that sense, it's not surprising that GDPNow gets more accurate on that subcomponent than it does on some others—at the end of the quarter than it does the beginning, certainly.
Heintjes: I should note for listeners that we're having this conversation in very late July, which is the data we're seeing right now. And I want to ask you about the other subcomponents of GDPNow.
Higgins: For the other subcomponents of GDP, there actually isn't monthly data exactly like there is for consumer spending. So those subcomponents are probably a little more indirectly related to the source data than consumer spending. But I would say the methodology is fairly similar, for the most part. There are some examples—like nonresidential investment, and research and development, and other intellectual property products—where the first estimate is actually going to be based on judgmental trends from the BEA that GDPNow is never going to have any data on. For that subcomponent, it's not really surprising that that initial forecast it makes is generally not too dissimilar from the final forecast it makes, since it's really not getting a lot of intermediate data over the next three months. But for the other subcomponents, it's something sort of in between those—the consumption and intellectual property products cases—where it gets a lot of related data but not as closely related as consumer spending, but the numbers definitely do change. So essentially what happens is, GDPNow makes a forecast for the data that's related to GDP subcomponents. If the data comes in stronger than the model anticipated, the forecast goes up, and if it comes in weaker, then the forecast goes down, essentially. That always happens every quarter, essentially.
Heintjes: Pat, in 2022, GDPNow missed by nearly 2 percentage points in the first quarter, forecasting a slightly positive growth when actual output declined by more than 1 percentage point. Why was the error so large? And has GDPNow been less accurate during the pandemic than the five to eight years before that? And a follow-up question: Will we continue to see a larger...I hate to call it "margin of error," but I will for lack of a better term.
Higgins: I think "margin of error" is actually a pretty good way to think about it, since typically you want to say that there's a confidence interval, and maybe 90 percent of the time the data is going to come inside that confidence interval. But 10 percent of the time, it comes outside of that. And so you can characterize, based on historical patterns, how often forecasts have fallen inside a particular range—that's probably a good way to think about it. So for the first-quarter miss you referred to—a bit more than 1 percentage point—typically that would have been a larger error than the model might have expected prior to the pandemic, but I would say that's not an atypical error during the pandemic. Most of that error was concentrated in net exports—I think maybe a little over 1 percentage point—so nearly two-thirds of the miss was in net exports. For the second quarter—the number that actually just came out today, as we're speaking—net exports was also a relatively large miss, almost a percentage point on that. But just by chance, other subcomponents canceled those out in some sense, so GDPNow is overly optimistic in consumer and government spending forecasts. And it just happened to be the case that those offset each other. So sometimes that happens and GDPNow is reasonably close, but other times—like in the first quarter—it isn't, and you have relatively large misses.
Heintjes: In general, how would you characterize GDPNow's accuracy for a single given quarter? I imagine achieving great accuracy within a window that small is really challenging.
Higgins: Yes, I would say prior to the pandemic, something like half a percentage point to a percentage point might have been a typical miss. The subcomponent forecasters were, I would say, generally a little bit smaller than that. And then since we've entered in the pandemic, essentially all of the subcomponent forecast errors have gotten larger, and aggregate GDPNow forecast margin of error has gone up as well. The uncertainty has definitely gone up. I would say the forecast error—we've got this quarter for 0.3 percentage points—is probably a low. That's a pretty good forecast. Maybe something like a percentage point is probably more typical of what you would see.
Heintjes: Let's talk about the other models that take snapshots of GDP. Did you observe other measures similarly missing the mark? I'm not sure how much solace there would be, but was there a feeling of strength in numbers among those who forecast GDP growth?
Higgins: In the first quarter I would say GDPNow was overly optimistic. Some professional forecasters were as well. They probably missed by close to 2 percentage points as well on that first quarter number. So there probably was some consolation that if other forecasters are having a harder time being accurate, then it's not something just specific to the GDPNow model. Essentially, it's a harder environment to forecast in. I would say it's probably more often the case that GDPNow differs to some extent from other forecasters, particularly the consensus assessment from professional forecasters. And so I think it's important in some sense to compare GDPNow's forecast with professional forecasters, to give the model some sort of reality check—to emphasize this is not necessarily the best or only forecast of GDP there is.
Heintjes: We touched on the pandemic, and I want to follow up on that. How did the early months of the pandemic, going back to 2020, affect GDPNow specifically and, more broadly, GDP?
Higgins: I would say late in March of 2020, when we were all aware of the pandemic, and businesses had shut down and things like that—there really wasn't any March data released, where the macroeconomic impact was recognized by the data. And so, at that point GDPNow was still pretty optimistic about growth. But once that March data started coming in, it came down pretty steeply. By the end of the quarter, it was essentially forecasting a small contraction in GDP. The actual contraction was much larger, because services spending—which was where a lot of the decline was concentrated in—that data actually wasn't released before the first reports, so GDPNow's entire miss, 4 percentage points, was essentially concentrated in that one subcomponent.
Heintjes: And as 2020 went on, it was a most unusual year, we all recall. And what about the rest of the year?
Higgins: Essentially, through April and May the forecast kept coming down. The pandemic had an extremely large—unprecedentedly large—impact on macroeconomic measures like consumer spending and employment. At the nadir of its forecast, it essentially forecasted even a bigger contraction in GDP than actually happened—about twice as large. It did not anticipate the big rebound in activity that occurred in May and June of that year, and essentially as it got data points coming in that indicated stronger growth and a big rebound, the forecast essentially got revised up, both for the second quarter and the third quarter. By the end of the forecast cycle for each of those quarters, it had ended up reasonably close to what actually happened—although, 30 to 60 days before that, it was off by quite a lot.
Heintjes: In response to all this—these fluctuations and everything—did you make any modifications to the model? And why did you make any modifications, if you did?
Higgins: Probably the biggest modifications we made were in April 2021, when the most extreme months of the pandemic—in terms of its effects on macroeconomic data—were already apparent. And so, what happened from March to May is you had this huge decline in economic activity, and this really rapid rebound, in some sense. And those were so large, relative to what had happened in previous recessions—where you had essentially six to 12 months of weak economic activity, followed by some rebound gradually occurring—we didn't want to have the model essentially forecast something like that occurring again, where if you have a huge decline in economic activity you would expect a rapid rebound, since essentially that only happened because of the pandemic, and that really isn't a new normal for recessions or the way we think about them. So essentially what the model did was introduce what statisticians call "dummy variables," to say what occurred in this period is essentially specific to those two months or three months during the pandemic, and that we shouldn't expect those things to repeat themselves going forward. So I would say essentially we're trying to get the GDPNow model to work in the same way it did prior the pandemic, when it had only data prior to the pandemic. I would characterize it as more of a recalibration of the model than a structural overhaul in some sense.
Heintjes: I don't want to get too in the weeds here, because I'll quickly get over my head, but how do you see these improvements enhancing the accuracy of the GDPNow model?
Higgins: I would say you would not want to see this again, but if you saw something like another pandemic occur again, the model is going to not be impacted two or three quarters later by what happened in the pandemic, essentially. It's essentially like the big impacts of the second-quarter decline in economic activity, and the big rebound in the third quarter, influencing the model forecasts that were made in the first quarter of 2021 in a way that didn't make the model more accurate. It's essentially trying to take the distortions that are put into the model by the pandemic and make those go away as quickly as possible, in some sense.
Heintjes: Pat, let's turn from talking about the GDPNow model and discuss actual GDP. In the first quarter of the year real GDP declined, and GDPNow is projecting a decline in the second quarter as well. Does this mean the economy is or will be in a recession?
Higgins: You're right—GDP declined in the first quarter, and actually in the second quarter. So, according to some technical definitions that are cited in the news media by some, it could be characterized as a recession—although it's not a recession until the National Bureau of Economic Research, or the NBER, classifies it as such. The indicators they typically look at—things like personal consumption expenditures, payroll, employment, industrial production, etc.—those indicators are actually all smushed together in something called the Composite Index of Coincident Indicators from the Conference Board. That indicator actually, in the first half of the year, grew as rapidly as it did on average during the entire decade of the 2010s—a period where there weren't any recessions. So based on the way the NBER has classified recessions in the past, it's unlikely—I would say, in my own opinion—that they would characterize what happened in the first of the year as a recession, in spite of the decline in GDP in both of the first two quarters.
Heintjes: Well, you talk about the NBER's yardstick for a recession, which is two consecutive quarters of contraction. Is that the only yardstick used to measure a recession, or...?
Higgins: No, so they actually don't use that at all. I think in the financial press, the "two-quarter" definition—or even for some other countries—that might be used. But the National Bureau of Economic Research, I don't think they have a fixed methodology. But it's definitely not the two-quarter decline in real GDP. For example, in the 2001 recession, there were never two quarters of consecutive decline but that was still characterized as a recession in some sense. There are other indicators, like the unemployment rate. Every recession, that's gone up by at least half a percentage point. It's called the "Sahm rule," after Claudia Sahm, and that's still at the lowest rate it's been during the pandemic—it hasn't gone up at all. So GDP is really the one indicator that's behaved kind of like what you would see in a recession. Almost every other one has not.
Heintjes: Pat, if I can put you on the spot here: how do you interpret the signals we're seeing? We've talked about a lot of data, I guess a lot of it's noisy. How do you interpret those signals?
Higgins: I would say the signals from GDP are probably not as bleak as the actual headline numbers showing something like a 1.5 percentage point decline in the first quarter and about 1 percent in the second. In terms of its forecasting ability, there's a subcomponent of GDP called "real final sales to private domestic purchasers." That's kind of a long-winded name just for consumer spending and investment—except for everything in inventories, essentially—when they're actually making fixed investment rather than replacing their inventories. If you combine that with consumer spending, that actually forecasts GDP growth itself one to four quarters out more accurately—historically, at least—than GDP does. And so that indicator increased 3 percent in the first quarter and was flat in the second quarter—zero percent growth—but it did not decline. I would say, overall, that number's essentially indicating that the economy has definitely slowed somewhat, I would say, in terms of forecasting ability, that the numbers still really aren't giving a negative signal—the headline GDP numbers are.
Heintjes: As we've noted, GDP growth has been weak, or even negative, recently, but jobs growth has continued to be quite strong. Is this sort of divergence or disconnect unusual, historically speaking? And what are we to make of this sort of dynamic?
Higgins: Yes, I would say it's pretty unusual. The typical way I would say most analysts or economists compare the labor market with GDP, or the business sector output, is with productivity measurement. And so that productivity, which is closely related to GDP over employment, in the first quarter that decline in productivity was the second largest over the past 75 years. And based on the data we have for business sector output this quarter, what we're likely to see in the productivity report to be released early next month would be probably a two-quarter decline larger than anything we've seen in the postwar era, by probably quite a bit. So even though there are conceptual distinctions between productivity and GDP over employment, I would say those are minor enough to say that, really, these things have been quite at odds with each other.
Heintjes: And we have to allow for the revisions, which frequently happen as new data come in.
Higgins: Yes, that's right. Sometimes those revisions can be pretty substantial, and so we have seen negative productivity growth rates get revised up to some extent. Productivity data itself is actually quite noisy, so you can get negative readings definitely quite often on a one- or two-quarter basis—and even on a four-quarter basis, you'll see a negative productivity reading. I would just say the magnitude of what you're seeing currently just gives an indication of...we've really never seen anything where like 300,000 jobs are added a month for six straight months, and in both those quarters you had negative growth rates. That's definitely never happened before.
Heintjes: Pat, I can't let this episode end without touching on inflation. I know that GDPNow isn't an inflation tool, and there are many such measures out there that look at inflation, including some by the Atlanta Fed. But I wanted to ask you if the current accelerated rate of inflation has an impact on how GDP functions. Can you discuss that?
Higgins: Yes. Essentially, much of the source data that GDPNow receives—that's nominal dollar-value spending, released very often by the Census Bureau. What GDPNow does is take price indices from other reports—often released from the Bureau of Labor Statistics, like the consumer price index and the producer price index—to form a price index that's essentially related to the nominal spending indicator from the Census Bureau, and create its own real or inflation-adjusted series, essentially, and uses that to forecast the subcomponents of GDP. So that's going to work, essentially, whether or not inflation is high or low. We're still going to get data on those price indices, and GDPNow will be able to forecast a real growth rate. The one caution I would add is, essentially, once inflation gets higher and higher, the technical things that GDPNow has to do—like some seasonal adjustment of those indicators or some interpolation of indicators—that gets probably somewhat more difficult and less accurate when the inflation rates are rising into higher levels.
Heintjes: So those more technical issues like you just mentioned—seasonal adjustments, or seasonally adjusted price changes—they don't really impede GDPNow in performing its snapshots?
Higgins: I would say that's largely the case. The one caution I would give is when...so typically, if you had a retail sales report that showed something like 1 percent growth in a month, that would be a quite strong report. And we did get that a month ago, but we also had a consumer price index reading that went up over 1 percent that same month—and real retail sales growth was probably very minimal, in some sense. So it does make it more difficult to interpret some of those spending reports, like retail sales, but GDPNow does do the price adjustments necessary to forecast real GDP.
Heintjes: Well, Pat, this has been a really enlightening conversation. I've looked forward to it for some time, and you've given us all a lot to think about today. I want to thank you for taking the time to talk through this all with us today.
Higgins: Thank you, Tom—I very much enjoyed our conversation.
Heintjes: And that brings us to the end of another episode of the Economy Matters podcast. Again, I'm Tom Heintjes, managing editor of the Atlanta Fed's Economy Matters magazine, and I appreciate your spending some time with us today. I encourage you to visit Economy Matters at atlantafed.org and read the many interesting features we have for you there, as well as on the Atlanta Fed website, which is where you'll find the GDPNow tool we've been discussing this episode. I encourage you to follow us on social media as well, so you can get real-time GDPNow updates as we receive them, which happens quite often. Thanks again for listening, and let's meet again next month.