The Atlanta Fed's macroblog provides commentary and analysis on economic topics including monetary policy, macroeconomic developments, inflation, labor economics, and financial issues.
Comments are moderated and will not appear until the moderator has approved them.
Please submit appropriate comments. Inappropriate comments include content that is abusive, harassing, or threatening; obscene, vulgar, or profane; an attack of a personal nature; or overtly political.
In addition, no off-topic remarks or spam is permitted.
May 31, 2018
Learning about an ML-Driven Economy
Developments in artificial intelligence (AI) and machine learning (ML) have drawn considerable attention from both the real and financial sides of the economy. The Atlanta Fed's recent Financial Markets Conference, Machines Learning Finance: Will They Change the Game?, explored the implications of AI/ML for the financial system and public policy. The conference also included two macroeconomics-related sessions. A presentation of an academic paper, and the subsequent discussion, looked at why AI/ML has not (yet) shown up in the productivity statistics. Also, a policy panel on the implications of AI/ML developments for monetary policy was part of the conference. This post summarizes the policy panel discussion.
Vincent Reinhart, chief economist at Standish Mellon Asset Management, opened the panel discussion with the observation that developments in AI/ML could affect the performance of the overall economy in a variety of ways. For example, advancing technology could better match workers with jobs and, as a result, boost employment. On the other hand, it could also complicate job matching by forcing jobs and workers to become more specialized.
A combination of three factors is driving the recent growth in AI/ML, explained Carolyn Evans, head economist and senior data scientist at Intel Corporation: increased data availability, faster computers, and improved algorithms for analyzing the data. Like Reinhart, she noted that AI/ML could have various effects on the economy. For example, AI/ML is helping to reduce cost and boost supply. On the demand side, AI/ML is increasing the efficiency of product searches by buyers. However, as some online sellers become better than others at using AI/ML to help customers find the products they want, customer relationships may become stickier. In addition, firms may come to value interactions with customers more highly because these interactions could provide them with valuable data to use with AI/ML to better serve current and future customers. Evans raised the question of whether these developments could change the nature of pricing.
Dallas Fed president Rob Kaplan said he believes AI/ML is causing a structural change. It is not the first new technology to affect the economy, but the economic effects of this technology are more pervasive. For instance, business pricing power is already more constrained than it used to be, but even businesses that seemingly have some power currently worry that they make themselves more vulnerable to AI/ML-enabled disruption if they raise prices. Kaplan also emphasized the importance of skills training and building human capital to alleviate what he views as the inevitable loss of jobs to AI/ML.
The issue of how monetary policymakers should think about AI/ML was the focus of a presentation by Chicago Fed president Charles Evans. He observed that the "sign, magnitude, and timing" of any resulting structural change are all uncertain. This uncertainty, he said, argues against the use of fixed policy rules such as the Taylor Rule. He suggested that the Federal Reserve should instead follow an "outcome-based policy," adjusting policy based on the evolution of expected inflation and unemployment relative to the policy objectives of stable prices and full employment.
You can download the available presentations from the 2018 Financial Markets Conference web pages. The videos will be posted as they become available. Read Notes from the Vault for a summary of sessions on the strengths and weaknesses of ML, some financial regulatory and broader ethical issues, and the use of ML by investors.
February 13, 2018
GDPNow's Forecast: Why Did It Spike Recently?
If you felt whipsawed by GDPNow recently, it's understandable. On February 1, the Atlanta Fed's GDPNow model estimate of first-quarter real gross domestic product (GDP) growth surged from 4.2 percent to 5.4 percent (annualized rates) after a manufacturing report from the Institute for Supply Management. GDPNow's estimate then fell to 4.0 percent on February 2 after the employment report from the U.S. Bureau of Labor Statistics. GDPNow displayed a similar undulating pattern early in the forecast cycle for fourth-quarter GDP growth.
What accounted for these sawtooth patterns? The answer lies in the treatment of the ISM manufacturing release. To forecast the yet-to-be released monthly GDP source data apart from inventories, GDPNow uses an indicator of growth in economic activity from a statistical model called a dynamic factor model. The factor is estimated from 127 monthly macroeconomic indicators, many of which are used to estimate the Chicago Fed National Activity Index (CFNAI). Indices like these can be helpful for forecasting macroeconomic data, as demonstrated here and here.
Perhaps not surprisingly, the CFNAI and the GDPNow factor are highly correlated, as the red and blue lines in the chart below indicate. Both indices, which are normalized to have an average of 0 and a standard deviation of 1, are usually lower in recessions than expansions.
A major difference in the indices is how yet-to-be-released values are handled for months in the recent past that have reported values for some, but not all, of the source data. For example, on February 2, January 2018 values had been released for data from the ISM manufacturing and employment reports but not from the industrial production or retail sales reports. The CFNAI is released around the end of each month when about two-thirds of the 85 indicators used to construct it have reported values for the previous month. For the remaining indicators, the Chicago Fed fills in statistical model forecasts for unreported values. In contrast, the GDPNow factor is updated continuously and extended a month after each ISM manufacturing release. On the dates of the ISM releases, around 17 of the 127 indicators GDPNow uses have reported values for the previous month, with six coming from the ISM manufacturing report.
[ Enlarge ]
For months with partially missing data, GDPNow updates its factor with an approach similar to the one used in a 2008 paper by economists Domenico Giannone, Lucrezia Reichlin and David Small. That paper describes a dynamic factor model used to nowcast GDP growth similar to the one that generates the New York Fed's staff nowcast of GDP growth. In the Atlanta Fed's GDPNow factor model, the last month of ISM manufacturing data have large weights when calculating the terminal factor value right after the ISM report. These ISM weights decrease significantly after the employment report, when about 50 of the indicators have reported values for the last month of data.
In the above figure, we see that the January 2018 GDPNow factor reading was 1.37 after the February 1 ISM release, the strongest reading since 1994 and well above either its forecasted value of 0.42 prior to the ISM release or its estimated value of 0.43 after the February 2 employment release. The aforementioned rise and decline in the GDPNow forecast of first-quarter growth is largely a function of the rise and decline in the January 2018 estimates of the dynamic factor.
Although the January 2018 reading of 59.2 for the composite ISM purchasing managers index (PMI) was higher than any reading from 2005 to 2016, it was little different than either a consensus forecast from professional economists (58.8) or the forecast from a simple model (58.9) that uses the strong reading in December 2017 (59.3). Moreover, it was well above the reading the GDPNow dynamic factor model was expecting (54.5).
A possible shortcoming of the GDPNow factor model is that it does not account for the previous month's forecast errors when forecasting the 127 indicators. For example, the predicted composite ISM PMI reading of 54.4 in December 2017 was nearly 5 points lower than the actual value. For this discussion, let's adjust GDPNow's factor model to account for these forecast errors and consider a forecast evaluation period with revised current vintage data after 1999. Then, the average absolute error of the 85–90 day-ahead adjusted model forecasts of GDP growth after ISM manufacturing releases (1.40 percentage points) is lower than the average absolute forecast error on those same dates for the standard version of GDPNow (1.49 percentage points). Moreover, the forecasts using the adjusted factor model are significantly more accurate than the GDPNow forecasts, according to a standard statistical test . If we decide to incorporate adjustments to GDPNow's factor model, we will do so at an initial forecast of quarterly GDP growth and note the change here .
Would the adjustment have made a big difference in the initial first-quarter GDP forecast? The February 1 GDP growth forecast of GDPNow with the adjusted factor model was "only" 4.7 percent. Its current (February 9) forecast of first-quarter GDP growth was the same as the standard version of GDPNow: 4.0 percent. These estimates are still much higher than both the recent trend in GDP growth and the median forecast of 3.0 percent from the Philadelphia Fed's Survey of Professional Forecasters (SPF).
Most of the difference between the GDPNow and SPF forecasts of GDP growth is the result of inventories. GDPNow anticipates inventories will contribute 1.2 percentage points to first-quarter growth, and the median SPF projection implies an inventory contribution of only 0.4 percentage points. It's not unusual to see some disagreement between these inventory forecasts and it wouldn't be surprising if one—or both—of them turn out to be off the mark.
April 6, 2015
Is Measurement Error a Likely Explanation for the Lack of Productivity Growth in 2014?
Over the past three years nonfarm business sector labor productivity growth has averaged only around 0.75 percent—well below historical norms. In 2014 it was negative, as can be seen in chart 1.
The previous macroblog post by Atlanta Fed economist John Robertson looked at possible economic explanations for why the labor productivity data, taken at face value, have been relatively weak in recent years. In this post I look at the extent to which “measurement error” can account for the weakness we have seen in the data. By measurement error, I mean incomplete data and/or sampling errors that are reduced when more comprehensive data are available several years later. I do not mean the inherent difficulties in measuring productivity in sectors such as health care or information technology.
As seen in chart 1, negative four-quarter productivity growth rates have been quite infrequent in nonrecessionary periods since 1948. In S. Borağan Aruoba's 2008 Journal of Money, Credit and Banking article “Data Revisions Are Not Well Behaved,” he found that initial estimates of annual productivity growth are negatively correlated with subsequent revisions. That is, low productivity growth rates tend to be revised up while high rates tend to be revised down. This is illustrated in chart 2.
In each of the panels, points in the scatterplot represent an initial estimate of fourth-quarter over fourth-quarter productivity growth together with a revised estimate published either one or three years later. For example, the green points in each plot show estimates of productivity growth over the four quarters ending in the fourth quarter of 2011. In each plot, the x-coordinate shows the March 7, 2012, estimate of this growth rate (0.3 percent). The y-coordinate of the green dot in chart 2a shows the March 7, 2013, estimate of fourth-quarter 2011/fourth-quarter 2010 productivity growth (0.4 percent) while the y-coordinate of the green dot in chart 2b shows the March 5, 2015, estimate (0.0 percent).
In each chart, the red dashed line shows the predicted revised value of productivity growth as a function of the early estimate (using a simple linear regression). Chart 2a shows that, on average, we would expect almost no revision to the most recent estimate of four-quarter productivity growth one year later. Chart 2b, however, shows that low initial estimates of productivity growth tend to be revised up three years later while high estimates tend to be revised down. Based on this regression line, the current estimate of -0.1 percent fourth-quarter 2014/fourth-quarter 2013 productivity growth is expected to be revised up to 0.3 percent by April 2018.
The intuition for this is fairly straightforward. Low productivity growth could come about from either underestimating output growth, overestimating growth in hours worked, or a combination of the two. Which of these is most likely to occur, according to historical revisions? This is shown in chart 3, which plots the predicted revisions to four-quarter nonfarm employment growth and four-quarter nominal gross domestic product (GDP) growth conditional on two assumed values for the initial estimate of four-quarter productivity growth: 0 percent (low) and 4 percent (high).
Nominal GDP is used instead of real GDP as methodological changes to the latter (e.g., the introduction of chain-weighting starting in 1996) make an apples-to-apples comparison of pre- and post-revised values difficult. Using fourth-quarter over fourth-quarter growth rates since 1981, the diamonds on the solid lines in chart 3 show that an initial estimate of 0 percent productivity growth would, on average, be associated with a three-year upward revision of 0.39 percentage point to four-quarter nominal GDP growth and a three-year downward revision of 0.10 percentage point to four-quarter nonfarm payroll employment.
With 4 percent productivity growth, the diamonds on the dashed lines show predicted three-year revisions to nominal GDP growth and employment growth of -0.40 percentage point and 0.14 percentage point, respectively. As the chart shows, these estimates are sensitive to the sample period used to predict the revisions. Using only data since 1989 (not shown), the regression would not predict a downward revision to employment growth conditional on an initial estimate of 0 percent productivity growth. Overall, however, the plot suggests that revisions to output growth are more sensitive to initial estimates of productivity growth than revisions to payroll employment growth are. This is consistent with the sentiments expressed by Federal Reserve Vice Chairman Stanley Fischer and Atlanta Fed President Dennis Lockhart at the March 30–April 1 Financial Markets Conference that employment or unemployment data may be more reliably measured than GDP.
Nevertheless, according to charts 2 and 3, the importance of measurement error in productivity growth is fairly modest. Ex-ante, we should not expect last year's puzzlingly low productivity growth simply to be revised away.
Editor's note: Upon request, the programming code and data for charts used in this macroblog post is available from the author.
April 2, 2015
What Seems to Be Holding Back Labor Productivity Growth, and Why It Matters
The Atlanta Fed recently released its online Annual Report. In his video introduction to the report, President Dennis Lockhart explained that the economic growth we have experienced in recent years has been driven much more by growth in hours worked (primarily due to employment growth) than by growth in the output produced per hour worked (so-called average labor productivity). For example, over the past three years, business sector output growth averaged close to 3 percent a year. Labor productivity growth accounted for only about 0.75 percentage point of these output gains. The rest was due primarily to growth in employment.
The recent performance of labor productivity stands in stark contrast to historical experience. Business sector labor productivity growth averaged 1.4 percent over the past 10 years. This is well below the labor productivity gains of 3 percent a year experienced during the information technology productivity boom from the mid-1990s through the mid-2000s.
John Fernald and collaborators at the San Francisco Fed have decomposed labor productivity growth into some economically relevant components. The decomposition can be used to provide some insight into why labor productivity growth has been so low recently. The four factors in the decomposition are:
- Changes in the composition of the workforce (labor quality), weighted by labor's share of income
- Changes in the amount and type of capital per hour that workers have to use (capital deepening), weighted by capital's share of income
- Changes in the cyclical intensity of utilization of labor and capital resources (utilization)
- Everything else—all the drivers of labor productivity growth that are not embodied in the other factors. This component is often called total factor productivity.
The chart below displays the decomposition of labor productivity for various time periods. The bar at the far right is for the last three years (the next bar is for the past 10 years). The colored segments in each bar sum to average annual labor productivity growth for each time period.
Taken at face value, the chart suggests that a primary reason for the sluggish average labor productivity growth we have seen over the past three years is that capital spending growth has not kept up with growth in hours worked—a reduction in capital deepening. Declining capital deepening is highly unusual.
Do we think this sluggishness will persist? No. In our medium-term outlook, we at the Atlanta Fed expect that factors that have held down labor productivity growth (particularly relatively weak capital spending) will dissipate as confidence in the economy improves further and firms increase the pace of investment spending, including on various types of equipment and intellectual capital. We currently anticipate that the trend in business sector labor productivity growth will improve to a level of about 2 percent a year, midway between the current pace and the pace experienced during the 1995–2004 period of strong productivity gains. That is, we are not productivity pessimists. Time will tell, of course.
Clearly, this optimistic labor productivity outlook is not without risk. For one thing, we have been somewhat surprised that labor productivity has remained so low for so long during the economic recovery. Moreover, the first quarter data don't suggest that a turning point has occurred. Gross domestic product (GDP) in the first quarter is likely to come in on the weak side (the latest GDPNow tracking estimate here is currently signaling essentially no GDP growth in the first quarter), whereas employment growth is likely to be quite robust (for example, the ADP employment report suggested solid employment gains). As a result, we anticipate another weak reading for labor productivity in the first quarter. We are not taking this as refutation of our medium-term outlook.
Continued weakness in labor productivity would raise many important questions about the outlook for both economic growth and wage and price inflation. For example, our forecast of stronger productivity gains also implies a similarly sized pickup in hourly wage growth. To see this, note that unit labor cost (the wage bill per unit of output) is thought to be an important factor in business pricing decisions. The following chart shows a decomposition of average growth in business sector unit labor costs into the part due to nominal hourly wage growth and the part offset by labor productivity growth:
The 1975–84 period experienced high unit labor costs because labor productivity growth didn't keep up with wage growth. In contrast, the relatively low and stable average unit labor cost growth we have experienced since the 1980s has been due to wage growth largely offset by gains in labor productivity. Our forecast of stronger labor productivity growth implies faster wage growth as well. That said, a rise in wage growth absent a pickup in labor productivity growth poses an upside risk to our inflation outlook.
Of course, the data on productivity and its components are estimates. It is possible that the data are not accurately reflecting reality in real time. For example, colleagues at the Board of Governors suggest that measurement issues associated with the price of high-tech equipment may be causing business investment to be somewhat understated. That is, capital deepening may not be as weak as the current data indicate. In a follow-up blog to this one, my Atlanta Fed colleague Patrick Higgins will explore the possibility that the weak labor productivity we have recently experienced is likely to be revised away with subsequent revisions to GDP and hours data.
- Business Cycles
- Business Inflation Expectations
- Capital and Investment
- Capital Markets
- Data Releases
- Economic conditions
- Economic Growth and Development
- Exchange Rates and the Dollar
- Fed Funds Futures
- Federal Debt and Deficits
- Federal Reserve and Monetary Policy
- Financial System
- Fiscal Policy
- Health Care
- Inflation Expectations
- Interest Rates
- Labor Markets
- Latin AmericaSouth America
- Monetary Policy
- Money Markets
- Real Estate
- Saving Capital and Investment
- Small Business
- Social Security
- This That and the Other
- Trade Deficit
- Wage Growth