Projections are a model of the economy under stated assumptions, not a forecast. They assume full employment at the end of the horizon, no recession and no war, and every release rebuilds them with a new base year and revised inputs. The direction of a projection is far more dependable than its magnitude, and the further out the horizon the wider the real uncertainty around a printed figure.
What the model assumes before it starts
A number printed to one decimal place, ten years ahead, rebuilt from scratch every year. The precision is real and the certainty is not, and the gap between those two is worth understanding. Precision and accuracy are different properties of a number.
The projections are built on assumptions stated openly rather than hidden. That the economy reaches full employment by the target year, that no major recession or conflict intervenes, and that long-run patterns in participation, productivity and industry demand continue. Each assumption is published rather than buried in a footnote.
Those are reasonable assumptions and they are also why a shock reshapes everything. A projection is not a prediction of what will happen; it describes where current patterns lead if nothing interrupts them, and something usually interrupts them. A decade is long enough for at least one genuine surprise.
Three reasons the numbers move between releases
The first is a new base year. Each release starts from more recent actual employment, so the starting point itself shifts before anything else is considered. The revision begins before any judgment has been applied.
An occupation that grew faster than expected in the base period begins the next projection from a higher number. That mechanically changes its growth rate even when nothing about the outlook has changed at all. A moving starting point produces a moving percentage automatically.
The second is reclassification, since occupational definitions get revised periodically and an occupation split in two is not comparable with its earlier self. The third is changed assumptions about technology adoption, immigration, health spending and demographics, each of which is itself an estimate produced somewhere else. Estimates built on estimates inherit all the uncertainty beneath them.
What survives revision well
Direction and rank survive well. An occupation projected to decline is usually still projected to decline next time. The sign of the number is far sturdier than its size.
The ordering of the fastest-growing list changes far less than the specific percentages attached to it. That stability is what makes the release useful despite the revisions. Rankings hold up even when the individual figures move.
Demographically driven projections are the most stable of all, because the driver is already visible in today’s population. The people who will need more healthcare in 2034 exist now and their ages are known, which is why healthcare projections have held up better than anything else. Population arithmetic is the sturdiest input in the whole model.
What survives badly
Precise magnitudes survive badly, and so does anything resting on how quickly a technology gets adopted. Those are judgments about organizational behavior rather than arithmetic on a population. Organizations adopt tools at speeds nobody can reliably predict.
Adoption speed is the single input most likely to be revised, and it moves in both directions rather than one. That symmetry is worth remembering when reading a confident claim about decline. Overprediction and underprediction have each happened repeatedly here.
Some occupations declined faster than projected because software arrived sooner than expected. Others were written off entirely and did not decline, because adoption was slower or the work was harder to specify than it looked from outside. Specification difficulty is consistently underestimated from a distance.
The precision problem
A projection printed as 40.1 percent carries an implied accuracy the method cannot support. The decimal is real arithmetic performed on the model’s output. The decimal is honestly computed and easily misread.
What it is not is a claim that the answer is accurate to a tenth of a percentage point ten years out. Those are two very different assertions and the formatting does not distinguish them. A printed decimal implies a confidence nobody claimed.
Read it as growing considerably faster than most and you are using it correctly. Read it as a forecast of 448,800 nurse practitioners in 2034 and you are treating a ranking as a measurement. Those two readings lead to very different decisions.
What projections structurally cannot include
They cannot include occupations that do not exist yet. The model projects the current occupational classification forward, so genuinely new work appears only once it is large enough to be classified. Classification lags reality by years in a fast-moving field.
That systematically understates change in emerging fields, and it is a limitation of the method rather than an oversight. Nobody can count a job category that has not been defined. The limitation is logical rather than a failure of effort.
They also exclude anything the assumptions rule out by construction. A recession is not projected as a possibility with a probability attached; it is assumed not to happen, which means the numbers describe one specific scenario rather than a range. No probability is attached to any part of the outcome.
How to read one honestly
Read it as a ranking rather than as a measurement. That single reframing fixes most of the ways these figures get misused. Rankings are what the method genuinely supports well.
Saying this occupation is growing faster than most and that one is shrinking is a claim the data supports well. Saying there will be 448,800 of these jobs in 2034 is a claim the data prints and does not really support. Printing a figure is not the same as vouching for it.
Then check which assumption most affects the occupation you care about. If the driver is demographic, trust the figure more; if it rests on technology adoption, treat it as one plausible path among several. Then watch the next release to see which way it moved.
Why the revisions are a feature rather than an embarrassment
A model rebuilt annually against fresh data is doing exactly what it should. The alternative would be a figure that never updates and drifts quietly away from reality. A stale forecast is worse than a revised one every time.
Revision is the mechanism by which a projection stays connected to what is actually happening. Treating each release as a correction of the last misunderstands the exercise. Each release is a fresh model rather than an erratum.
What matters is watching which direction the revisions move over successive releases. An occupation revised downward three releases running is telling you something the single most recent figure cannot. Direction of revision is itself a signal worth reading.
Common questions
Are projections forecasts?
No. They are a model under stated assumptions, including full employment at the target year and no major recession or conflict.
Why do they change each release?
A new base year, occupational reclassification, and revised assumptions about technology, demographics and industry demand.
Which projections are most reliable?
Demographically driven ones. The people who will need healthcare in 2034 already exist, which makes those figures unusually stable.
Which are least reliable?
Anything resting on how fast a technology is adopted, because that is a judgment about behavior rather than arithmetic.
How should I read the numbers?
As a ranking rather than a measurement. Direction survives revision; the precise magnitude often does not.
Are employment projections forecasts?
No. They assume full employment at the target year and no major recession or conflict, so they describe where current patterns lead if nothing interrupts them.
Which projections are most reliable?
Demographically driven ones. The people who will need healthcare in 2034 already exist and their ages are known, which removes the guesswork.
Why do the numbers change each release?
A new base year shifts the starting point, occupational definitions get revised, and the assumptions about technology, immigration and demand are themselves estimates.