On Monday, February 7th you might have heard that the unemployment rate for January was 4.0%. This isn’t technically true. The actual unemployment rate was estimated by the Bureau of Labor Statistics (BLS) at 4.4%. Meanwhile, the seasonally adjusted estimate of the unemployment rate was 4.0%. Seasonally adjusted data is the best data to use for analyzing and projecting economic trends. However, seasonally adjusted data from the BLS is unreliable for the years of 2016-2022; and future seasonally adjusted data that will be released by the BLS for the years of 2022, 2023 and 2024 will also be unreliable.
Seasonally Adjusted Data is More Useful
In many instances, seasonally adjusted data can give a better reading of economic trends. Data without seasonal adjustments goes up and down throughout a year in a way that changes slightly over time but is largely repeating. Statisticians at the BLS which reports CPI and employment data (and other government agencies) use statistically sound methods to account for these trends. The BLS methodology involves estimating seasonal trends for a year by looking at data from the previous 4 years and the next 4 years or as many as are available. The BLS engages in backward revisions to previously reported seasonally adjusted monthly data every year. So, a data point can be revised 5 times, first when its year ends and then the next 4 times as the next 5 years end.
It’s necessary to realize that Presidents would, if given the ability, change seasonal adjustments in a way that exaggerates economic performance while in office. Hence, a method of seasonally adjusting data which uses intuition would immediately be suspect. Therefore, a pre-set series of statistical methods to seasonally adjust raw data that the government publishes and then consistently sticks to is the best way to account for seasonal fluctuations.
Why Seasonally Adjusted Data Will be Inaccurate
An unfortunate consequence of the BLS adjustments is that statistically significant effects in a certain month in a certain year impact seasonal adjustments for the current year, the prior 4 years and the next 4 years. In particular, the weird effects of the virus will be significant. If a number goes up a lot in May 2020, the statistical analysis will conclude that that number is likely to go up in May. Therefore in all neighboring years, the seasonally adjusted data will account by showing a lower (or more negative) increase in that number then what one can see in the not seasonally adjusted data.
Evidence of Inaccurate Employment Data
From February to April 2020, the labor force participation rate decreased and then it increased from April to June. According to seasonally adjusted data as of the January 2022 employment report (released on February 7th, 2022), the labor force participation rate has always ticked down or stayed constant from April to June for the years 2016-2019 and 2021. The chart below shows this. The reason for this is that the statistical methods used to do seasonal adjustments overestimate the seasonal upward pressure on labor force participation from April to June because of the specific situation in 2020 which involved Republican-controlled states ending lockdowns.
Meanwhile, if one looks at pre-covid data as of the January 2020 employment report (February 7th, 2020), you find that (prior to backward revisions) the seasonally adjusted labor force participation rate was estimated as increasing in 2018 and 2019.
Hence, one can find some substantial evidence that seasonally adjusted employment data is (while better than any plausible alternatives) inaccurate going forward until the end of 2024.
Evidence of Inaccurate Inflation Data
From February to April in 2020 the rate of monthly CPI declined. Then in April to June it increased. From July to October 2020, the rate of monthly CPI inflation declined. The plot below which uses CPI Data as of the December 2021 report (January 12th, 2022) shows that the monthly CPI inflation trends in surrounding years move in the opposite direction than in 2020. This means inflation increases from February to April, decreases from April to June and increases from July to October for the years 2016-2019 and 2021. The one exception to this rule is the April to June period in 2021. Monthly CPI inflation is in this graph measured in log-% change in CPI from one month to the next.
However, using data as of the December 2019 report (January 14th, 2020), one sees a less consistent pattern. Monthly inflation data trends moved in the opposite direction in 2016-2019 than it did in 2020 according to the last dataset. The February-April inflation change was only positive 3/4s of the time. The April-June inflation change was only negative 3/4s of the time. The July-October inflation change was positive only 3/4s of the time.
Hence, one can find some substantial evidence that seasonally adjusted CPI data is (while better than any plausible alternatives) inaccurate going forward until the end of 2024.
Can Seasonal Adjustments be Fixed?
Short answer: no.
It’s hard to say how much can be done to fix seasonal adjustments. In hindsight, it may have made sense to seasonally adjust data by more than the previous 4 years and then backwardly revise the data by adding 5 more years (including the current one) at a time. Currently, ¼ of the dataset for future seasonal adjustments is severely messed up because of the outbreak in 2020. However, I don’t have a statistics PhD to understand the actual methods of seasonal adjustments done. Those who do don’t seem to have complained.
Any attempt to change officially reported seasonal numbers will likely invite accusations of statistical manipulation by partisan elites (I’m guessing the numbers will make Biden’s economy look slightly better or slightly worse). The best move may be for the general public to look at annual data or rates of change over a year between non-seasonallyt adjusted figures (for instance the unadjusted CPI index grew 7.2 log-% from January 2021 to January 2022). Meanwhile important institutions (like the FED or the IMF) which need to really understand the state of the economy, should probably have statisticians on staff who can do their own seasonal adjustments in a way where data from 2020 does not affect seasonal weightings.
Conclusion
In conclusion, it’s always important to not read too much into 1 months' worth of data, but this is especially important right now.
Correction:
In a previous version of this post, I wrote that the BLS revises non-seasonally adjusted unemployment rate estimates for two times after the initial publication of the estimate. This is not true. The BLS does do this for nonfarm payroll employment and other data in its monthly employment situation report. However, the unemployment rate is estimated only once by the BLS on a non-seasonally adjusted basis, and it sticks to its first estimate.