What an allocator actually learns from four flat years of industrial output

Six annual readings of American industrial production, on a 2017 base of 100: 95.1, then 99.3, then 101, then 100.8, then 100.1, then 101.2. The first two are a recovery from 2020. The last four are a flat line. Somebody who looked away in 2022 and looked back in 2025 would find the industrial sector almost exactly where they left it, two tenths of a point higher, which on this scale is indistinguishable from unchanged.

A flat line is the most misread object in economic data. It invites two confident readings, and both of them are wrong in the same way. The first says nothing is happening, which is the reading of someone who wanted the number to be lower and is pleased. The second says the number is broken, which is the reading of someone who wanted it higher and is not. Both treat a single aggregate as a verdict on a country, and an aggregate is not a verdict. It is a sum, and a sum's most important property is what it throws away.

The position

My position is that the industrial production index is not a scoreboard for industrial policy, has never been one, and should stop being asked to serve as one by both sides of that argument. It reports one number for a sector that contains everything from chip fabrication to cement. Capacity added in one part of it and capacity retired in another are netted against each other before the reader sees anything, and the net has been close to zero since 2022. That is a true statement about the arithmetic. It is very nearly no statement at all about where in the country somebody could put money to work.

This matters because an allocator does not face the aggregate. An allocator faces a plant, in a place, making a specific thing, with a specific set of buyers and a specific labour market around it. The question is never whether the industrial sector grew. The question is whether this line of production is expanding, whether the people to run it exist within commuting distance, and whether anyone will still want the output in year eleven. The index cannot address any of those and does not claim to.

What sits inside the flat line

Consider one component. Domestic semiconductor production, published on the same 2017 base of 100, ran 121 in 2021, 131 in 2022, 135 in 2023, 144 in 2024 and 165 in 2025. The source line identifies it as part of the Federal Reserve's G.17 industrial production release, which is the same statistical family as the headline. Across those same five readings the headline index moved 1.9 points, from 99.3 to 101.2. The semiconductor line moved 44, and 21 of those arrived in the final year alone. Narrow the window to the flat stretch since 2022 and the headline moves two tenths of a point, from 101 to 101.2, while the component moves 34, from 131 to 165.

I am not going to tell you what the rest of the industrial sector did to accommodate that. The file I am working from gives me index levels and no component weights, and without weights the residual is not a calculation, it is a guess wearing a decimal point. What I can say is narrower and more useful than the guess would be. A total that moves 1.9 points over five readings, while a line published inside the same release moves 44 over the identical span, is not functioning as a summary of that line. Anyone who cites the headline to characterise what is happening in semiconductors, in either direction, is using a number that demonstrably does not track it.

The join I am not allowed to make

There is a second series that everyone reaches for at this point. Manufacturing capacity utilisation ran 72.9 percent in 2020, 77.3 in 2021, 77.6 in 2022, 76.6 in 2023, 75.4 in 2024 and 75.3 in 2025. It peaked in 2022 and has drifted down in each of the three readings since. The tempting move is one line of arithmetic: output is flat, the share of capacity in use is falling, therefore capacity itself must be rising, therefore the plants are in fact being built and the index simply has not caught up.

It is a clean story and I cannot support it with these two series. Manufacturing capacity utilisation covers manufacturing. The production index covers the industrial sector, which is wider. Dividing a manufacturing ratio into an industrial-sector index produces a number whose denominator and numerator describe different populations, and the answer it gives is not wrong so much as meaningless. The mistake is easy to make precisely because both series are Federal Reserve publications, both are annual, and percentages and index points sit together on a page as though they belonged to each other.

The discipline is to stop there rather than to round the difference away. What survives is smaller and still worth having. Within manufacturing alone, the share of capacity in use is 2.3 points below its 2022 level and has fallen three readings running. Something in the relationship between what manufacturers can make and what they are making has been loosening for three years. Which side of that ratio moved is not in this file, and I would rather leave the question open than close it with a division I would have to defend.

The strongest case against me

The best objection to everything above is short. If the net is flat, the net is flat. An economy that opens one fabrication plant and closes three older factories has not added industrial capacity, whatever the composition looks like, and an allocator who takes comfort in composition is telling himself a story about the part of the data he likes. Composition, on this view, is what people reach for when the level disappoints. It is the oldest move in economic commentary and it usually deserves the suspicion it gets.

I think that objection wins something real, and it should be given what it wins without argument. Anyone who promised that a decade of industrial programmes would show up as a step change in aggregate industrial output should be judged against 101.2 on a 2017 base, and by that test the promise has not been kept. The flat line is a legitimate answer to an aggregate claim. If the claim was aggregate, the rebuttal is aggregate, and there is nothing unfair about it.

What the objection does not win is the deployment question, because nobody deploys capital into an aggregate. The netting that makes the index useful as a national summary is exactly what makes it useless as a siting instrument. Two states can produce the same contribution to a flat national line while one is adding a supply chain and the other is losing one, and the index will never distinguish them. An allocator who reads the headline and concludes there is nothing to do has not been careful, and an allocator who reads it and concludes the programmes worked has not been careful either. The number does not license either conclusion.

What six numbers are actually worth

Held to what it can support, the flat index still earns its place in a file. It sets a ceiling on how large any composition story can be at the national level: whatever is growing, it has not yet been enough to move the total, and anyone describing a transformation in progress is describing something the aggregate has not registered. It also gives a reader a stable base against which component series can be read, which is the only reason the semiconductor comparison above says anything at all.

1. Read the index as a constraint on the size of a claim, not as a measure of activity.
2. Refuse arithmetic between series that cover different universes, however compatible the units look.
3. Ask which component moved before asking whether the sector moved, and accept that the answer is usually not in the same file.
4. Carry the retrieval date with the number, because a figure without a date is an assertion rather than an observation.

Two further series belong in the same file for context, neither of them the subject here. Output per hour in the nonfarm business sector stands at 118 on a 2017 base, against 109.1 in 2020, so the same sort of index that shows no movement in industrial output shows a good deal of movement in what an hour of work produces. Manufacturing employment ran 12.1 million workers in 2020, peaked at 12.9 million in 2023 and stood at 12.6 million in 2025. Neither of those settles the argument. Both of them make it harder to tell a simple version of it, which is usually the sign that they are worth keeping.

The most useful thing a flat series does is refuse to settle a question. Four years of readings clustered around a single point are not a dull result to be explained away or a scandal to be exposed. They are an instrument telling you, accurately, that it is the wrong instrument for what you are trying to decide. The allocator's response to that is not disappointment. It is to go and find the series that is actually about the decision, and to be honest that the national headline was never going to be it.