Revisions and why the first print is not the truth

Why the first number is an estimate 

A statistic such as gross domestic product is not counted. It is assembled. Business surveys, tax records, customs declarations, retail data and government accounts are combined into one figure describing an entire economy over three months. Those inputs arrive at very different speeds, and several of them do not exist at all until long after the period they describe has closed. The agency nevertheless has a publication date, fixed months ahead and printed in every calendar, and on that date it publishes what it holds.

What it holds on the first date is partial. A survey closes with a response rate well short of complete, so the missing responses are modelled from the ones received and from the same respondents in earlier periods. Components for which no source data exist yet are estimated from related indicators that do. Seasonal adjustment factors, which strip out the part of a movement that recurs every year, are themselves estimated from a history the new observation will go on to alter. The output is a considered estimate produced against a deadline, and the agency says so in the release. Nothing is concealed. The word preliminary is usually in the title.

The trade off is deliberate. An agency could wait until every input had landed and publish one figure, accurate and more than a year late, or publish early and correct itself as evidence arrives. Statistical agencies have almost universally chosen the second, because a figure that appears long after the period it describes is of no use to the central banks, finance ministries and businesses the statistic exists to serve. Timeliness is bought with accuracy, and the revision is what it costs.

One period, several numbers 

Because correction is designed in, a single period does not have one figure. It has a sequence of them, each called a vintage. The shape differs by statistic and by country, and every agency publishes its own revision policy alongside its calendar, but the pattern is recognisable everywhere.

1. An advance or flash estimate, published soon after the period closes, resting on the smallest share of the eventual source data and in some cases on a partial sample.
2. Intermediate estimates in the following months, each incorporating source data that has arrived since, and each published under the same series name with a different qualifier.
3. An annual revision, in which the series is recalculated against slower administrative sources such as tax filings, which count rather than sample and are available at no monthly or quarterly deadline.
4. A benchmark or comprehensive revision, at intervals of several years, updating definitions, base years, classifications and seasonal methods together. It can restate the level of an entire history at once, so figures quoted from either side of one are not comparable without restating.

Final, where it appears in the name of an estimate, means final for that round rather than final in any lasting sense. A quarter described as final is still subject to the next annual revision and to the following benchmark. Some series are still being amended years later, which is why anyone comparing across a long history works from a stated vintage rather than from whatever a database returns today.

Which releases are revised most, and why 

The size of a revision is not arbitrary and is not a measure of an agency's competence. It follows from how the figure is built, which makes it predictable in a limited but useful way: the more of a number that is estimated rather than counted on the day, and the slower its sources arrive, the larger the eventual amendment tends to be.

• Gross domestic product is the most heavily revised headline statistic almost everywhere, because it is the broadest. It aggregates output, income and expenditure across a whole economy, and some of its largest components, business investment, inventories and trade in services, are among the slowest things in an economy to measure.
• Employment counts drawn from business surveys are amended in each of the following two months as late responses arrive, then benchmarked once a year against administrative payroll records. The benchmark is the larger correction, and it is scheduled.
• Trade balances and inventories depend on customs and company filings that arrive on their own timetable rather than on the statistical calendar.
• Consumer price indices are, in most countries, revised rarely or not at all in their headline unadjusted form, because the prices were collected within the month in question and little is outstanding. Seasonally adjusted versions of the same index are recalculated when the seasonal factors are updated.
• Sentiment and purchasing manager surveys publish a flash reading from part of the sample and a final reading a week or two later from the whole of it. The gap between the two is a revision under another name, and it is scheduled.

A calendar shows none of this. Two releases can carry the same importance marking, land at the same minute and sit inside the same widening of quotes, while one is close to settled on the day and the other is a first draft that will read materially differently within the year. An importance marking measures the attention a release attracts, not the confidence attached to the number inside it.

A view built on a number that is later withdrawn 

The surprise arithmetic from the previous lesson makes the consequence compact. A surprise is the distance between the consensus expectation and the published figure. When the published figure is replaced, the surprise computed on the day is replaced with it, and the sign can change.

In both cases the figure the market had in front of it and the figure describing that same quarter a few months later fall on opposite sides of what was expected. A position opened on the day was sized against a quantity that no longer exists in the published record. What does still exist is the exposure, the cost of having opened it, and every price it has traded at since.

Why the correction rarely produces the original reaction 

A revised figure for the same period usually passes with a fraction of the movement the first print produced, and often with none that is separable from ordinary trading. That is not indifference. Four mechanisms work against it, and each is structural rather than a matter of what anybody thinks.

• It is not published as its own event. A revision arrives inside the next release of the same series, under a headline describing a new period, and sits in a subsidiary line of that document. The entry the calendar draws attention to is not the one the revision is in.
• There is nothing to be surprised against. The reaction to a scheduled release is driven by the distance between a compiled consensus and what arrives, and no consensus is compiled for a revision. Without an expectation, the mechanism that turns a release into a repricing has no input.
• Positioning has already turned over. Whatever was opened, closed, added to or stopped out around the first print was resolved over the following weeks for reasons unconnected to the pending correction, so there is rarely an accumulated position the correction forces anybody to unwind.
• The information has usually arrived by other routes. By the time a quarterly figure is amended, further monthly data covering the same period has been published, alongside company reporting, surveys and commentary from the central bank watching the same series.

Exceptions exist and their shape is consistent. A scheduled annual benchmark revision, announced well ahead and understood to be large, is treated by some participants as an event in its own right and appears on some calendars as one. A revision that changes the direction of a series rather than its magnitude is on occasion reported as news. Neither exception changes the general case, and neither can be read in advance from a calendar entry alone.

Where practitioners disagree 

The first argument is about which vintage is the right one to study, and it is genuine because the two answers serve different questions. One tradition works only with first prints, on the grounds that they are what was known at the time and what prices responded to. Real time databases exist to preserve each vintage as published, because an ordinary historical series silently overwrites the past with figures nobody had. Another tradition works with the fully revised series, on the grounds that it is the better description of what an economy actually did, and that describing an economy should not mean inheriting the errors of its early estimates. Both are right about the question they are answering, and conflating them produces studies that appear to find patterns which were never available to anyone at the time.

The second is about whether revisions carry information of their own. One position treats them as measurement error without direction, so the sequence of amendments is noise around a value that arrives eventually and the last revision says nothing about the next. Another holds that revisions cluster, because the models used to estimate missing inputs are calibrated on a recent past, and a period in which an economy is turning is exactly the period in which the recent past is the worst guide, so a run of amendments in one direction is watched as a possible signature of a turn. The second is a hypothesis about a statistical process rather than a rule, no figure is attached to it here, and the traditions holding each view have not settled it between them.