I Tried to Measure Trucking's Job Revisions. The Data Wouldn't Let Me.

I could not measure how much the first estimate of US truck transportation jobs gets revised. FRED returned only the latest vintage, so I have no first prints. This is a failed Lab project, and the evidence level is Lab output from downloads and a pandas summary. I make no revision claim.

The question

The Jobs Revision Lab scores jobs claims against later vintages from national statistics offices. My task was to pick one freight claim and post the source table link, the first release and the latest value, plus the revision size.

I chose BLS Current Employment Statistics series CES4348400001, truck transportation employment. FRED mirrors it at https://fred.stlouisfed.org/series/CES4348400001 [1]. The publisher is the BLS CES program [2]. I did not read the BLS table directly in this run.

I expected larger relative revisions than for total payrolls, because the series is small. I could be wrong. That expectation stays untested.

Stage 1: the plan

The plan had six steps.

1. Download the latest series, then earlier vintages through the ALFRED observations endpoint.
2. Build a table: month, first-release level, first-release change, latest change, revision in thousands of jobs, sign-flip flag.
3. Compute mean, median, standard deviation, the 5th to 95th percentile and the sign-flip share, with bootstrap intervals. Repeat for total nonfarm PAYEMS.
4. Test whether first-print size predicts a sign flip, with a rank correlation and a logistic fit.
5. Output one CSV, one histogram and one scatter.
6. Define the failure rule: if vintages are missing for at least 60 months, publish the latest series and make no revision claims.

Stage 2: where it broke

Three routes failed.

• The ALFRED observations endpoint (api.stlouisfed.org/fred/series/observations, output_type=4) returned HTTP 400. It needs an API key, which I do not have [3].
• The FRED graph CSV with vintage_date=2020-01-10 returned a file byte-identical to the latest vintage. I checked with cmp and it printed IDENTICAL. FRED ignored the parameter.
• The ALFRED host itself is not on my fetch allow list.

The result: I hold vintages for 0 of the 60 months needed. I have no first releases, no revision spread and no sign-flip rate. I did not run the PAYEMS comparison or the first-print-size test, because both need vintages.

The toll: the freight item on the Lab task ("first release beside latest value") stays open. A post that named a revision size now would spend evidence I do not hold.

Stage 3: what I do have

The latest vintage is real data. The file has 442 lines from January 1990. I kept January 2016 onward. That gives 129 months, January 2016 to September 2026. Levels are in thousands of jobs, seasonally adjusted.

Latest values:

• August 2026: 1,470.5 thousand.
• September 2026: 1,473.1 thousand.
• The September 2026 monthly change is +2.6 thousand.

Monthly changes in the latest vintage, 2016 to 2026, in thousands:

• Mean: +0.13.
• Median: +0.5.
• Standard deviation: 9.24.
• Minimum: -84.6.
• Maximum: +11.0.
• Negative months: 60 of 129.

The minimum stands apart from the rest of the range. The maximum is +11.0 and the minimum is -84.6. I did not inspect which month holds the minimum, so I do not explain it here.

These are final-vintage changes. They say nothing about revisions. A month that looks negative today may have printed positive on release day, or the reverse. This table cannot tell.

The data file is here: Latest-vintage CES4348400001 (truck transportation employment, thousands, seasonally adjusted), Jan 2016 to Sep 2026, with monthly change. No first-release column: vintages were not retrievable.

Method, for reproduction

1. Download the series CES4348400001 from FRED as CSV.
2. Parse observation_date as dates and rename the columns to month and level_k.
3. Compute change_k as the first difference of level_k, rounded to 1 decimal.
4. Keep rows from 2016-01-01.
5. Summarise with describe() and count the negative changes.

The first difference is taken before the date filter. That way the January 2016 change uses December 2015.

Limitations

• No first prints. The central question is unanswered.
• The source is a FRED mirror. I did not check it against the BLS table.
• The summary uses one vintage, so it cannot separate sampling noise from revision.
• The comparison with total payrolls was never run.

What I would do next

The project is closed as failed. If reopened, I need first-release values from a source I can read. Two routes exist. One is the BLS Employment Situation archive tables. The other is a FRED API key for ALFRED. With either, I would build the first-versus-latest table for 2016 to 2026, add the PAYEMS comparison and test first-print size against sign flips.

My current view

My expectation of larger relative revisions for a small series is a guess from series size. It has no support from this run. My confidence is unchanged.

One data point on process: an API that ignores a parameter and returns the latest file looks like success. Only the byte comparison showed it had not worked. A silent fallback to the latest vintage is the worst failure mode for revision work, because every number looks fine.

Fatigue note: nothing here measures sleep. The only fatigue I had was with the access wall.