The annual revision does not land on the same month.
Two US statistical agencies benchmark comparable monthly series on two different calendars. One has drifted across eight months of the year in ten years; the other has not moved. Neither publishes the dates.
data note · 2026-08-15 · ALFRED, Federal Reserve Bank of St Louis
Three monthly series sit on this site's record: US semiconductor production, US computer and electronics production, and US new orders for computers and electronic products. Each is held as 127 monthly vintages, sampled on the 15th of every month from 15 January 2016 to 15 July 2026. A vintage is the series as it stood on that day: not the current numbers, the numbers that were on the wire at the time.
Holding them all makes one thing visible that no publisher prints. Most months almost nothing moves. Then a single release rewrites nearly the whole history at once, and the month it happens in is not fixed.
What the vintages show
For the two Federal Reserve production series, nine of the 127 vintages restated 100 periods or more. Everything else is small: on all the other vintages, the median release restated five periods.
| Vintage | Semiconductor production | Computer and electronics production | Months since the last one |
|---|---|---|---|
| 15 Apr 2016 | 529 | 529 | |
| 15 Apr 2017 | 502 | 530 | 12 |
| 15 Apr 2018 | 363 | 553 | 12 |
| 15 Apr 2019 | 407 | 514 | 12 |
| 15 Jun 2021 | 592 | 592 | 26 |
| 15 Jul 2022 | 604 | 604 | 13 |
| 15 Apr 2023 | 543 | 613 | 9 |
| 15 Jul 2024 | 511 | 628 | 15 |
| 15 Dec 2025 | 601 | 643 | 17 |
Figures are the number of monthly periods whose value changed in that vintage.
Four consecutive Aprils, then a 26-month interval with no such release in the sampled vintages, then June, July, April, July, December. The gap between one large restatement and the next ranges from 9 months to 26.
The Census series behaves differently. It also has one large restatement a year, ten of them across the same window, and they do not drift:
| Vintage | New orders, computers and electronics | Months since the last one |
|---|---|---|
| 15 Jun 2016 | 230 | |
| 15 Jun 2017 | 182 | 12 |
| 15 Jun 2018 | 195 | 12 |
| 15 Jun 2019 | 204 | 12 |
| 15 Jun 2020 | 160 | 12 |
| 15 May 2021 | 170 | 11 |
| 15 May 2022 | 122 | 12 |
| 15 May 2023 | 133 | 12 |
| 15 May 2024 | 146 | 12 |
| 15 Jun 2025 | 158 | 13 |
Ten annual revisions, every one of them in May or June, none skipped. Two statistical agencies, two rhythms, in the same file.
What this is and is not
This is a statement about the instruments, not about the numbers they carry and not about any company. Annual benchmark revision is normal, documented practice at both agencies: a statistical series is revised when better source data arrives, and a revision that rewrites the back history is the system working rather than failing. Nothing here says any figure was wrong.
What is not published anywhere is the timing. Each agency documents that it revises annually. Neither publishes a list of the dates on which it actually landed, because a statistical release describes the current series, not the one it replaced. That list only exists if somebody kept every earlier version, which is what this record does.
The practical consequence is narrow and real. Anyone comparing a figure they noted in, say, February against the same figure today is comparing two series, not two dates, if a benchmark revision fell between them. On the Fed series, knowing whether one fell between them means knowing the actual date, and the actual date has moved by as much as eight months across this window.
How the numbers here were produced
Vintages are sampled on the 15th of each month, so a restatement is located to within a month: "15 April 2018" means the restatement was present in the 15 April vintage and absent from the 15 March one. It does not mean the release was published on the 15th.
The 100-period threshold is a reading aid, not a definition used by either publisher. It separates the annual rewrites from routine monthly revision cleanly here, because in this window nothing sits between 5 periods and 122 periods restated.
The absence of a large restatement for the two production series between April 2019 and June 2021 is what the sampled vintages hold. This note does not offer a reason for it, because the record has no sourced one. It is recorded as an absence, in the same way the record's own gap register records what it does not have.
A correction this made visible
The third series was attributed on this site to the Federal Reserve Board. It is not the Fed's: A34SNO is produced by the U.S. Census Bureau, from its Manufacturers' Shipments, Inventories and Orders survey, and reaches this record through ALFRED, which archives other agencies' series as well as the Fed's own. One publisher name had been applied to all three series in the collector.
The wrong label was found by asking why this series revises on a calendar the other two do not. It revises differently because a different agency benchmarks it. Corrected 15 August 2026 in the collector and in the collected file; no value changed, only the attribution.
Sources
- Industrial Production: Semiconductor and Other Electronic Component (IPG3344S), Board of Governors of the Federal Reserve System, via ALFRED: https://alfred.stlouisfed.org/series?seid=IPG3344S
- Industrial Production: Computer and Electronic Product (IPG334S), Board of Governors of the Federal Reserve System, via ALFRED: https://alfred.stlouisfed.org/series?seid=IPG334S
- Manufacturers' New Orders: Computers and Electronic Products (A34SNO), U.S. Census Bureau, via ALFRED: https://alfred.stlouisfed.org/series?seid=A34SNO
Every vintage counted above is held on this site's record and can be read back there, one release at a time.