Announced investment is not a statistic and industrial policy should stop being scored on it

An announced investment figure is not produced by an agency. It has no collection method, no revision schedule and no retrieval date. It is a number a company or a department states on the day it suits them to state it, and nothing afterwards obliges anyone to restate it. A project that is cancelled, halved or quietly deferred leaves the announced figure exactly where it was.

Industrial policy in the United States is nonetheless scored on that number. It is the figure that appears in the programme summary, in the anniversary release and in the argument about whether any of it worked. It is treated as though it were a measurement, and it is not one.

The position

My position is that the announced figure should be abandoned outright as a measure of industrial policy, and that it should not be replaced by a better announcement, a verified announcement or an audited pipeline of announcements. It should be replaced by two series that already exist.

Nothing new needs collecting. One published series counts the money actually put in place on manufacturing plant. Another counts what factories made. Between them they cover the input and the result, they are produced by agencies on a schedule, and they are capable of going against the policy. That last property is the one the announcement lacks, and it is the reason the announcement is preferred.

The line that fell

Construction spending on manufacturing plant is published by the Census Bureau. On this site it carries six annual observations: 75 billion dollars in 2020, 82 in 2021, 125 in 2022, 202 in 2023, 244 in 2024 and 215 in 2025. Each figure is a seasonally adjusted annual rate, so it states a year's spending rather than a month's. The 2024 reading is more than three times the 2020 reading, which is a large movement by any standard and the strongest thing the input side has to say.

That series has one property an announcement does not. It falls. The step from 2024 to 2025 is a fall of 29 billion dollars a year, the only fall in the six readings, and it arrives without anyone withdrawing anything or issuing a correction. Spending put in place is counted as it happens, so a project that stalls simply stops contributing to it. An announcement made in 2023 is still a 2023 announcement in 2026, whatever became of the site.

The line that did not move

The Federal Reserve Board publishes manufacturing output as an index on a 2017 base of 100. It reads 93 in 2020, 97.6 in 2021, 98 in 2022, 97.3 in 2023, 96.4 in 2024 and 97.3 in 2025. Across the four years in which plant spending went from an annual rate of 82 billion dollars to 215, the index went from 97.6 to 97.3. Its highest reading anywhere in the window is 98, and the 2025 reading sits 2.7 points below the base year.

The objection

Lag is the strongest objection, and it deserves to be put at full strength. A plant whose construction was paid for in 2024 does not produce anything in 2024. It has to be finished, commissioned, staffed and run up to something like a normal rate before an output index carries a single unit of what it makes, and for a large fabrication plant that sequence takes years. On this objection the two lines are not in contradiction at all. They are the same event observed at two different points in its life, and comparing the 2025 readings is comparing a cause with an effect that has not happened yet.

I accept the mechanism and reject the conclusion usually drawn from it. A lag is a reason to withhold judgement, not a reason to keep the announcement in the meantime. An announced total is not an early reading of an output index; it is a statement of intent that will never be revised against events, and installing it as a placeholder guarantees that the placeholder becomes the only figure anyone remembers. Where the honest answer is that it is too early, the honest thing to publish is that sentence.

Where a result does show

One narrow index has moved. Computer and electronic product output comes from the same board and sits on the same 2017 base. It reads 105.7 in 2020, 110.3 in 2021, 117 in 2022, 116.6 in 2023, 118.3 in 2024 and 126.8 in 2025. The step into 2025 is 8.5 points, the largest single-year move in the series, and the 2025 reading stands 26.8 points above the base year.

Set the two indices beside each other on their shared base and the distance in 2025 is 29.5 points. That is the shape a delivered result makes in this kind of data. It is not a broad aggregate lifting; it is one narrow line separating from a flat one. Whether the separation was caused by any programme is a question these series cannot settle, and I will not pretend otherwise. What it does demonstrate is that the measure registers a result when there is one to register.

What to publish instead

The replacement needs no new collection and fits in three lines. Report the money put in place on manufacturing plant, stated as the annual rate it is. Report the output index for the sector the policy actually named, on its published base. Then report the distance between the latest reading and the base year, alongside how many years of the programme have elapsed.

None of that is satisfying, and the dissatisfaction is the point. A programme two years old will show spending and no output, and the correct published line is that it shows spending and no output. A programme six years old that still shows no output has told the reader something worth knowing. The announced total would have said the same thing at both moments, which is to say nothing.

There is a cost here and it should be stated rather than buried. Announced totals are large, immediate and comparable across countries, and the two series proposed in their place are none of those things. What those series are is published by an agency, revised when the facts change, and capable of falling. Plant spending fell by 29 billion dollars a year in 2025, and the fall required nobody's permission. That is the property worth having, and it is the property an announcement can never acquire.