Department of Labor Logo United States Department of Labor
Dot gov

The .gov means it's official.
Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you're on a federal government site.

Https

The site is secure.
The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

Tag Archives: Seasonal adjustment

What’s Going on with the Large Revisions in Seasonally Adjusted Employment Estimates?

When BLS released The Employment Situation for January 2022 on February 4, we reported larger-than-normal revisions to the seasonally adjusted monthly nonfarm employment estimates from the Current Employment Statistics survey. Let’s take a closer look at the seasonal adjustment process and explain why the revisions were unusually large.

BLS seasonally adjusts data to account for recurring seasonal patterns in a time series. It allows our data users to analyze the underlying trends without the influence of seasonal movements. Seasonal movements might entail hiring extra retail trade workers around the December holidays. Seasonal movements also may entail workers leaving payrolls, such as school bus drivers when the schoolyear ends. These types of movements happen around the same time every year and affect about the same number of workers each year. Because the seasonal movements in nonfarm employment typically don’t change by much year to year, revisions to over-the-month changes from seasonal adjustment are typically small. However, data users may have a hard time seeing seasonal patterns when large events like strikes, hurricanes, or recessions occur. The unprecedented changes brought on by the COVID-19 pandemic make it even more difficult to determine the seasonal patterns.

Every year, during the employment benchmarking process, BLS staff takes a more comprehensive look back at the seasonal adjustment of time series to account for changes in seasonal patterns of nonfarm employment. Months in which large, irregular events occur are treated as outliers, which means that the abnormal change will not be treated as a new seasonal pattern. When we seasonally adjusted the data for 2020 after 2020 had ended, we treated several months individually as outliers. That works well for short-lived events, like strikes or a weather-related event. At that time, with very few observations after the start of the pandemic, that type of individual outlier detection was appropriate.

For example, the economy lost a massive number of jobs in March and April 2020 at the start of the pandemic. If we had not treated those months as outliers, or irregular events, the seasonal adjustment model would have expected massive employment losses as a new seasonal pattern for March and April 2021. That would not have been appropriate because those huge losses in March and April 2020 were not likely to recur. If the model had included the large March and April 2020 employment losses, when large job losses did not occur in March and April 2021, the model would have shown large job gains for those months on a seasonally adjusted basis. That would have distorted the underlying employment trend.

We now have more data observations after the start of the pandemic, and we benefit from the hindsight of our normal annual review of the data, which encompasses the last 5 years of data. We incorporated two additional outlier types that better isolated the initial losses in employment from the pandemic, while simultaneously accounting for new seasonal patterns that we have detected. One new type of outlier accounts for a temporary change in the level of the series. For example, drinking places that serve alcoholic beverages experienced a significant employment decline in April 2020 but have gained jobs since then, even if in fits and starts. In this case, we treated April 2020 as a temporary change outlier, which will adjust the months from April 2020 forward to account for both the large drop and subsequent recovery in employment this industry has experienced. This will better account for this temporary period of readjustment as employment levels recover towards what they were before the pandemic.

Employment in drinking places, alcoholic beverages, January 2020 to December 2021, seasonally adjusted

Editor’s note: Data for this chart are available in the table below.

The other new type of outlier accounts for a permanent shift in the level of a series. An example of this is nursing care facilities, which continued to experience employment declines after April 2020. In this case, April 2020 was treated as a level shift outlier, which resulted in subsequent months being adjusted to better reflect the continued lower level of employment.

Employment in nursing care facilities, January 2020 to December 2021, seasonally adjusted

Editor’s note: Data for this chart are available in the table below.

Incorporating these additional types of outliers stabilized the normal seasonal patterns and smoothed out the initial large swings in the seasonally adjusted data. These changes resulted in a more accurate payroll employment series and will allow users to better differentiate longer-term trends from seasonal movements.

These changes resulted in some large revisions to our monthly seasonally adjusted data for 2021, although the revisions partly offset each other. For example, total nonfarm employment in November and December 2021 combined is 709,000 higher than previously reported, while employment in June and July 2021 combined is 807,000 lower. The change for all of 2021 is just 217,000 higher than previously reported. Although these revisions are larger than usual, trends in the data emerge more clearly.

Want to really dig into the details of seasonal adjustment for the nonfarm employment data? See our specification files and technical documentation.

Employment in drinking places, alcoholic beverages, January 2020 to December 2021, seasonally adjusted
MonthPreviously publishedRevised

Jan 2020

417,200419,800

Feb 2020

423,700426,400

Mar 2020

381,800383,600

Apr 2020

85,60084,800

May 2020

145,400140,600

Jun 2020

233,800225,300

Jul 2020

238,700226,200

Aug 2020

250,800243,800

Sep 2020

266,200260,000

Oct 2020

289,700278,500

Nov 2020

280,900281,300

Dec 2020

243,000239,700

Jan 2021

248,300248,700

Feb 2021

282,200280,700

Mar 2021

303,900298,300

Apr 2021

311,700309,900

May 2021

318,300322,900

Jun 2021

322,000331,200

Jul 2021

329,700336,800

Aug 2021

329,500342,900

Sep 2021

343,900350,600

Oct 2021

345,500354,200

Nov 2021

347,900365,000

Dec 2021

344,500363,800
Employment in nursing care facilities, January 2020 to December 2021, seasonally adjusted
MonthPreviously publishedRevised

Jan 2020

1,586,5001,585,100

Feb 2020

1,585,7001,584,800

Mar 2020

1,581,6001,581,200

Apr 2020

1,534,3001,537,200

May 2020

1,502,6001,508,000

Jun 2020

1,487,3001,490,000

Jul 2020

1,469,7001,470,600

Aug 2020

1,459,2001,463,300

Sep 2020

1,456,6001,454,500

Oct 2020

1,451,5001,447,800

Nov 2020

1,439,0001,437,200

Dec 2020

1,433,4001,430,900

Jan 2021

1,415,0001,416,700

Feb 2021

1,404,2001,406,300

Mar 2021

1,400,5001,403,600

Apr 2021

1,382,2001,388,700

May 2021

1,375,0001,383,700

Jun 2021

1,371,1001,376,300

Jul 2021

1,371,0001,373,700

Aug 2021

1,365,3001,367,700

Sep 2021

1,349,8001,346,700

Oct 2021

1,359,1001,349,200

Nov 2021

1,350,9001,346,000

Dec 2021

1,345,7001,345,100

The Challenges of Seasonal Adjustment during the COVID-19 Pandemic

In a previous edition of Commissioner’s Corner, we described seasonal adjustment, the process BLS and many others use to smooth out increases and decreases in data series that occur around the same time each year. Seasonal adjustment allows us to focus on the underlying trends in the data. Seasonal adjustment works well when seasonal patterns are pretty consistent from year to year. But what about when there are large shocks to the economy, such as natural disasters and the massive effects of the COVID-19 pandemic and resulting business closures and stay-at-home orders? Today we’ll look at how BLS addressed this issue.

First, a little background on seasonal adjustment. Here’s an example similar to one we have used before, looking at employment in the construction industry. Construction employment varies throughout the year, mostly because of weather. As the chart shows in the “not seasonally adjusted” line, construction adds jobs in the spring and throughout the summer before it starts to lose jobs when the weather turns colder. The large seasonal fluctuations make it hard to see the overall employment trend in the industry. That makes it harder to study other factors that affect the trend, like changes in consumer demand or interest rates. After seasonal adjustment, the construction industry grew by 1.2 million jobs from the beginning of 2015 to the end of 2019.

Construction employment, 2015–19

Editor’s note: Data for this chart are available in the table below.

BLS seasonally adjusts data in several of its monthly and quarterly news releases.

Two Approaches to Seasonal Adjustment

BLS uses one of two approaches to seasonally adjust data in these releases—projected factors or concurrent seasonal adjustment. When we project seasonal adjustment factors, we only use historical data in the models. That means we calculate factors in advance, so they are not influenced by the most recent trends. Concurrent seasonal adjustment uses all the data available, including the most recent month or quarter. As a result, the factors are influenced by recent changes.

Regardless of whether the factors are projected or concurrent, the seasonal adjustment models can be additive or multiplicative. We’ll explain more about that below. The COVID-19 pandemic affected the seasonal adjustment process in different ways depending on how the seasonal factors are calculated.

Approach #1

The Consumer Price Index, Producer Price Indexes, and Employment Cost Index use the projected-factor approach and calculate seasonal factors once a year. BLS staff estimated the 2020 seasonal factors at the beginning of 2020 and have used them throughout the year. When new factors for 2021 and revised historical factors are calculated, BLS will examine the effects of the pandemic on the seasonal adjustment models.

Approach #2

We use a concurrent process to calculate the seasonal factors each month for nonfarm employment estimates for the nation, states, and metro areas, unemployment and labor force estimates for the nation, states, and metro areas, and job openings and labor turnover estimates. Each quarter, BLS also uses a similar concurrent process to calculate seasonal factors for productivity measures and business employment dynamics. This helps create the best seasonal factors when seasonality may shift over time. For example, think of schools letting out for summer a little earlier than they usually do each year, or the changing nature of delivery services because of online shopping. Using the most recent data to calculate seasonal factors helps pick up these changes to seasonality faster than the forecasted method. The risk of using the concurrent process is that it may attribute some of the movement in the estimates to a changing seasonal pattern when it really resulted from a nonseasonal event. BLS also annually examines and revises the historical seasonal factors even if the factors were originally calculated using concurrent adjustment. As the saying goes, hindsight is 20/20.

Before the COVID-19 pandemic, the concurrent seasonal adjustment models required limited real-time intervention. Examples of potential reasons for intervention include major events like hurricanes. The COVID-19 pandemic is unusual in its severity and duration, so significant intervention was needed.

BLS intervened in several ways to create the highest quality, real-time seasonal factors. The tool we use most often is called outlier detection. We consider outliers not to represent a normal or typical seasonal movement. When we label an observation as an outlier, we don’t use it to inform the seasonal adjustment model. Since economic activity is still being heavily influenced by COVID-19 and efforts to contain it, BLS has detected more outliers. When this happens, concurrent models behave more like projected-factor models because the most recent data are not used to create seasonal factors.

The Local Area Unemployment Statistics program uses another type of intervention, a technique call a level shift. It is used when there is a sudden change in the level of a data series. In this case, level shifts were used over a series of months.

Additive versus Multiplicative Models

As noted earlier, all BLS programs review their seasonal adjustment models each year. One of the steps during this process is to select a model—either additive or multiplicative. We use an additive model when seasonal movements are stable over time regardless of the level of the series. A multiplicative model is better to use when seasonal movements become larger as the series itself increases—that is, the seasonality is proportional to the level of the series. That means a sudden large change in the level of a series, such as the large increase in the number of unemployed people in April 2020, will be accompanied by a proportionally large seasonal effect. BLS did not want this to occur. When there are large shifts in a measure, multiplicative seasonal adjustment factors can result in adjusting too much or too little. In these cases, additive seasonal adjustment factors usually reflect seasonal movements more accurately and have smaller revisions.

Because of the unusual data patterns beginning in March 2020, both the Current Population Survey, which we use to measure unemployment and the labor force, and the Job Openings and Labor Turnover Survey switched from multiplicative to additive seasonal models for most series and did not wait until the typical yearend model review.

BLS does not produce the weekly data on unemployment insurance. We do, however, compute the seasonal adjustment factors used by the Department of Labor’s Employment and Training Administration for their Unemployment Insurance Weekly Claims data. As we recommended, the Employment and Training Administration recently switched from using multiplicative to additive seasonal adjustments.

Our quarterly Labor Productivity and Costs news release uses input data from the Bureau of Economic Analysis, the U.S. Census Bureau, and several BLS programs. Most of the input data are already seasonally adjusted by the source agencies or programs. The productivity program only seasonally adjusts monthly Current Population Survey data on employment and hours worked for about ten percent of workers, mostly the self-employed, who are not included in the monthly data from the Current Employment Statistics survey on nonfarm employment and hours. The productivity program detected outliers in some of the data beginning at the start of the COVID-19 pandemic in March 2020 and accounted for them in the estimates.

Science and Art

Seasonal adjustment of economic data is a scientific process that involves complex math. But seasonal adjustment also involves some art in addition to science. The art comes in when we use our judgment about outliers in the data or when we decide whether an additive or multiplicative model more closely reflects seasonal variation in economic measures. The art also comes in when we recognize how complicated the world is. During 2020 we have experienced not just a global pandemic but also massive wildfires in several western states, a historic number of hurricanes that made landfall, and other notable events that affect economic activity. Did our seasonal adjustment models properly account for all of these events? I can say we have tried our best with the information we have available. As we gather more data for 2020 and future years, we will continue to examine how we can improve our models to help us distinguish longer-term trends from the seasonal variation in economic activity.

Acknowledgment: Many BLS staff members helped make the technical details in this blog easier to understand, and they all have my gratitude. Three who were especially helpful were Richard Tiller, Thomas Evans, and Brian Monsell.

Construction employment, 2015–19
MonthSeasonally adjustedNot seasonally adjusted

Jan 2015

6,320,0005,953,000

Feb 2015

6,361,0005,962,000

Mar 2015

6,334,0006,051,000

Apr 2015

6,392,0006,300,000

May 2015

6,427,0006,491,000

Jun 2015

6,441,0006,633,000

Jul 2015

6,472,0006,718,000

Aug 2015

6,490,0006,754,000

Sep 2015

6,508,0006,704,000

Oct 2015

6,547,0006,740,000

Nov 2015

6,598,0006,685,000

Dec 2015

6,630,0006,542,000

Jan 2016

6,620,0006,252,000

Feb 2016

6,650,0006,256,000

Mar 2016

6,680,0006,402,000

Apr 2016

6,701,0006,614,000

May 2016

6,691,0006,758,000

Jun 2016

6,702,0006,913,000

Jul 2016

6,736,0006,989,000

Aug 2016

6,737,0006,997,000

Sep 2016

6,768,0006,971,000

Oct 2016

6,798,0006,981,000

Nov 2016

6,819,0006,903,000

Dec 2016

6,821,0006,700,000

Jan 2017

6,847,0006,459,000

Feb 2017

6,889,0006,527,000

Mar 2017

6,909,0006,634,000

Apr 2017

6,916,0006,820,000

May 2017

6,928,0006,998,000

Jun 2017

6,955,0007,169,000

Jul 2017

6,960,0007,212,000

Aug 2017

6,990,0007,248,000

Sep 2017

7,004,0007,201,000

Oct 2017

7,027,0007,208,000

Nov 2017

7,066,0007,147,000

Dec 2017

7,093,0007,004,000

Jan 2018

7,114,0006,729,000

Feb 2018

7,200,0006,840,000

Mar 2018

7,205,0006,933,000

Apr 2018

7,223,0007,129,000

May 2018

7,266,0007,336,000

Jun 2018

7,282,0007,497,000

Jul 2018

7,304,0007,554,000

Aug 2018

7,335,0007,586,000

Sep 2018

7,355,0007,535,000

Oct 2018

7,378,0007,557,000

Nov 2018

7,376,0007,454,000

Dec 2018

7,402,0007,311,000

Jan 2019

7,452,0007,069,000

Feb 2019

7,423,0007,062,000

Mar 2019

7,443,0007,170,000

Apr 2019

7,469,0007,377,000

May 2019

7,478,0007,540,000

Jun 2019

7,497,0007,699,000

Jul 2019

7,504,0007,753,000

Aug 2019

7,508,0007,760,000

Sep 2019

7,524,0007,700,000

Oct 2019

7,541,0007,720,000

Nov 2019

7,539,0007,609,000

Dec 2019

7,555,0007,447,000