Reference 03 November 2026 · 6 min

A plain-English guide to data revisions: when the old number is the news, and how to write a plan that a revision cannot move.

Written Tuesday November 3, accompanying the October 27 analysis: consumer confidence rose to 91.7, "up 7.5 points," but September was revised from 81.9 to 84.2 in the same release (Conference Board); the plan's fixed thresholds of 74.9 and 88.9 were unaffected. Why data get revised, which series revise most, when the revision is a third of the headline move, and why pre-registered thresholds are written as levels rather than distances from a base that can change.

By Igor Komolov

Written Tuesday November 3, 2026, accompanying the Tuesday October 27 analysis note, which graded a pre-registered plan against the Conference Board's October consumer confidence index. The index came in at 91.7, "up 7.5 points" from September. But September had been first reported at 81.9 and was revised to 84.2 in the same release, so the rise was 9.8 points against the number everyone had been looking at, and 7.5 against the number nobody had seen. The plan's thresholds were fixed in advance at 74.9 and 88.9, so the revision did not change the grade. This guide explains why economic data get revised, which series revise most, how a revision can be the real news in a release, and how to write a plan so that a revised base cannot quietly move the goalposts.

Why data get revised

A first print is an estimate made with incomplete information. Survey responses arrive late; seasonal factors are re-estimated as more years accumulate; benchmarks (tax records, censuses, annual surveys) replace samples; definitions change. Each of these produces a revision, and they stack. The Conference Board's monthly revision is the simplest kind: the preliminary cutoff date leaves late responses out, and the final figure folds them in (Conference Board, consumer confidence). September's 81.9 became 84.2 because the late respondents were, on average, more confident than the early ones.

Which series revise most

Series	Revision pattern	Example from this thread
Nonfarm payrolls	Two monthly revisions, then an annual benchmark; revisions of tens of thousands are routine	See NFP anatomy
GDP	Advance, second, third estimate, then annual updates	Q2 2026 is "2.2 percent" today; it has been revised already and will be again (the advance estimate)
Jobless claims	Prior week revised every Thursday, usually upward by a few thousand	The week to October 10 was revised up to 200,000 in the following week's release (October 22 note)
Consumer confidence (Conference Board)	Prior month revised with late responses	September 81.9 to 84.2
Import and export prices	Prior months revised as late prices arrive	August 2026 went from 0.7 to 0.5 percent (import prices guide)
Michigan sentiment	Preliminary and final each month; no later revision	October preliminary 46.3, final 48.0 (October 23 note)
CPI	Seasonally adjusted figures revised once a year; unadjusted index is final
Treasury auction results	Not revised

Sources: the issuing agencies' technical notes, and the analysis notes linked. The direction of claims revisions (usually up) is a feature of late-filed claims being added; it is not a conspiracy and not a signal.

When the revision is the news

A release has two numbers that can surprise: the new month, and the old month's new value. Markets price the first and often ignore the second, which is a mistake when the revision is large relative to the series' typical monthly change. On October 27, the 2.3-point upward revision to September was a third of the size of the headline rise. A reader who compared 91.7 with the 81.9 they remembered would have seen a 9.8-point jump; one who read the release would have seen 7.5. Both are true. Which one "the market" reacted to is unknowable, and the 30-minute window around 10:00 AM moved the 10-year 0.4 basis point, so it hardly matters for that day. It matters a great deal for anyone trying to say whether confidence is rising: the revised path (84.2 then 91.7) is a steadier climb than the unrevised one (81.9 then 91.7).

The payrolls version of this is the classic case. A headline beat with a large downward revision to the prior two months can leave the three-month total lower than before the release. The negative NFP print guide has an example.

Revisions and pre-registered plans

This is where the Tuesday plan earned its design. It said: at 74.9 or lower, the 10-year closes below Monday's 4.986 percent; at 88.9 or higher, above it. Those lines were built as September's 81.9 plus or minus 7.0 points. Then September moved. Had the plan been written as "7.0 points above the prior month," the strong threshold would have shifted to 91.2 with the revision, and 91.7 would still have cleared it, but by half a point instead of 2.8. Had the revision been larger, a print that the plan's author expected to trigger could have been reclassified as no call after the fact, with the author choosing which base to use. Fixed numbers remove that choice.

The rule, then: thresholds are written as levels, not as distances from a number that can change. When a plan must reference a prior value (claims against last week's level, ECI against last quarter's), it should quote the value it used, so that a revision cannot be used either to rescue or to sink the grade. The Friday ECI plan did this: "against Q2's 0.9 percent," a number the release could in principle have revised. It did not, but the plan was protected either way. The same discipline applies to the open-ended conditional statements graded in these notes; event windows covers the companion problem of choosing the price series.

What revisions do not tell you

The direction of the next revision. An upward revision to September says nothing about where October's 91.7 will settle.
Bias in the agency. Systematic revision patterns (claims up, early GDP estimates off by tenths) are documented by the agencies themselves and are a property of the collection method.
Which number the market used. There is no way to observe it. Treat the reaction as a reaction to the release as a whole.

A checklist for any release

Read the prior period's revised value before the new headline; compute the change both ways.
Compare the revision's size with the series' typical monthly change; if it is a large fraction, it is news.
In a plan, write thresholds as fixed levels, and quote any base value used.
Note which series are never revised (auction results, market prices) and which are revised for years (GDP, payrolls).
When citing a historical figure, say which vintage it is, or that you do not know.