What does Years Out actually show?

The job market for the year you graduate rather than the year you are looking. A student choosing a major in 2026 will not enter the market until 2030, but every hiring statistic they are shown describes today. Years Out separates the two: it measures where people with a given major actually ended up, how wide that major's fallback is, what the earnings spread looks like rather than just the median, and where the postings trend is heading — with the uncertainty band widening the further out you set the graduation-year slider.

What is fallback width and why is it the headline number?

Fallback width is the effective number of occupations a major actually leads to — the exponential of the Shannon entropy of its exit distribution, Miller-Madow corrected for small-sample bias. A major whose graduates concentrate in one destination has a narrow fallback; a major whose graduates scatter has a wide one. It became the headline metric for an unglamorous reason: the original headline, entry-level postings velocity, requires a commercial dataset priced between roughly twenty-two thousand and a hundred and thirty-six thousand dollars a year, while fallback width is computable from free public microdata and almost nobody publishes it. The data constraint reordered the product.

Does this tell me which major to choose?

No, and it is built so that it cannot. There are no rankings, no risk scores, no traffic lights and no trap-major labels, and it will never say do not study this. Individual variation — ability, school, network, luck — far exceeds the differences between majors shown here, and every figure on the page is a group statistic rather than a prediction about any one person. The restraint is deliberate: a tool that issues verdicts about eighteen-year-olds on this evidence would be overstating what the evidence supports.

What are the known weaknesses?

The page lists six, each labelled with whether it can be fixed. It covers bachelor's-degree holders only. It uses a one-year sample, so small majors are unreliable — flagged on screen and fixable with the five-year file. The postings line counts all seniority levels, not entry-level, which is the sharpest limitation because the tool's premise is that entry-level and senior hiring have decoupled; the clean entry-level series is paid-only. AI exposure is a prediction about tasks, not a record of replacement. It measures people employed now, not this year's graduates. Wages originally ignored cost of living — that one is fixed and toggleable.

Years Out — The Job Market You'll Graduate Into

This tool won't tell you whether to study it. It shows you the shape of the bet.

Fallback width Measured

effective occupations

Top-exit concentration Measured

%

AI exposure of exits Proxy

Exit fan — where people with this major actually went Measured

ACS 2023 PUMS, employed bachelor's holders aged 22–32. Left column is share of graduates; right column is that occupation's AI exposure (0 = none, 1 = the full job could be covered).

Occupation	Share of graduates		AI exposure ●

Method and known bias

Shares are PWGTP-weighted. Exposure comes from Eloundou et al. (2023) human-rated β scores, joined from Census occupation codes to SOC by token overlap; the match rate for occupations listed on this page is 97.2%.

Limits: covers bachelor's degree holders only. FOD1P records the bachelor's field even when a person's career is driven by a later graduate degree. It measures the stock of employed workers, not recent-graduate flow. Exposure is a predicted measure — not a record of jobs actually replaced.

What actually happened to them Measured

Three outcomes for recent graduates of this major, each shown against the median of the majors on this page. Blue dot is this major; the grey tick is the peer median.

How each is defined

Underemployment — the share of employed 22–32 year-old graduates working in occupations where fewer than half of all employed workers aged 25+ hold a bachelor's degree. The threshold is computed inside ACS PUMS itself rather than imported from an external crosswalk, so the definition is self-contained and auditable. It is a proxy for "a job that did not require the degree," not a judgement about the job.

Unemployment — share of 22–32 year-old graduates in the labour force who are unemployed (ESR = 3). Single-year sample, so treat differences under about 1 point as noise.

Further schooling — share of graduates aged 27–34 who hold a master's degree or higher. High values mean the bachelor's alone is often not the terminal credential in this field — which is a cost, in both years and tuition, that the headline earnings figure does not show.

Earnings spread, not just the median Measured

Full-time workers aged 22–32, 25th to 75th percentile of annual wages; the tick is the median. The width of the bar is the point. A high median with a wide spread is a different bet from a lower median with a narrow one.

Method and limits

ACS 2023 PUMS WAGP, adjusted by ADJINC, restricted to WKHP ≥ 35 and positive wages, PWGTP-weighted. Wages are self-reported and cluster on round numbers, so medians for different majors sometimes land on the same value — the percentile spread is the more informative reading. National figures; cost of living is not adjusted for, and the geographic mix differs by major.

Door velocity — new postings in this field, past and forward Measured + estimated

Indeed new-postings index (Feb 2020 = 100). Solid line is observed; dotted line extrapolates the last 24 months; the shaded area is the confidence interval. Drag the graduation-year slider and watch the interval widen.

Table view and method

The projection is a log-linear trend fit to the last 24 months. The confidence band is the standard deviation of historical 12-month changes × √(years ahead) — a random-walk-style expansion. Volatility is estimated excluding the 2020–21 COVID shock, which would otherwise inflate the band with a one-off event. The band widens as you push graduation later. That is deliberate: the further out you look, the less anyone can tell you.

This line measures postings at all seniority levels, not entry-level only. Occupation × seniority postings data is paid-only in the US (Lightcast, $22k–136k/yr) with no free substitute. The known national entry-level signal: Indeed, May 2026 — entry-level postings −7.5% YoY against senior postings +14.7% YoY.

Fallback width — where you sit Measured

Effective occupations = exp(entropy of the occupational distribution), with Miller–Madow small-sample bias correction applied. Higher means more viable alternative paths if the main exit closes. Click any row to switch.

Why this number deserves more weight than a forecast

Four-year labor market prediction is inherently unreliable. Under unreliable prediction, the rational move is not to maximize the chance of guessing right — it is to choose the option whose cost of being wrong is lower. The occupational flow structure of graduates changes slowly, which makes fallback width considerably more reliable at a four-year horizon than door velocity.

Bias warning: entropy is systematically underestimated at small samples (roughly −23% at n ≈ 1,000). Majors below n = 1,500 are flagged and should not be compared directly against large majors. The production version should use the ACS 5-year sample.

Route divergence — are recent graduates taking the same paths? Measured

The three largest destinations for mid-career graduates of this major

Share of each cohort working in those three roles

What this can and cannot tell you

The three occupations shown are the largest destinations for mid-career graduates of this major. The right column shows what share of recent graduates are in those same three.

This measure cannot separate two different causes, and it is important not to over-read it. Part of any gap is ordinary career progression — people move into management and senior roles with age, so mid-career destinations naturally include jobs a 24-year-old has not reached yet. Another part may be a genuinely changed entry market. A single cross-section of ACS cannot tell them apart; only a longitudinal panel could.

Read it as a question worth asking about a specific major, not as an answer.

Another explanation

The strongest refutation of this tool's own narrative, placed here rather than left for someone else to bring up.

It may not be AI. It may be interest rates. The Economic Innovation Group (Jan 2026) analyzed 238 million US job postings across 767 occupations and found that entry-level and senior postings fell in parallel starting in spring 2022 — six months before ChatGPT launched, and aligned with the Fed's first rate hike in March 2022. If that holds, the primary driver of shrinking entry-level hiring may be monetary policy rather than technology.

Stanford Digital Economy Lab addressed this interest-rate confound directly in a February 2026 research note, and their July 2026 finding on gender gaps also concluded that AI was not the main driver.

This tool presents an observable fact — the entrance is narrowing — alongside the competing explanations for it. It does not claim causation.

What this tool cannot tell you

Every limit below is load-bearing — each one changes how a number on this page should be read. They are listed here rather than hidden in a methodology footnote, with an honest status for each: some are fixable, some are not fixable with any public data that exists today.

Bachelor's degree holders only

ACS records field of degree only for people who hold a bachelor's. Anyone whose path ran through an associate degree, a certificate, an apprenticeship, or no degree at all is invisible in every number here — including in the underemployment denominator. This tool describes what happened to people who finished a four-year degree, and nothing else.

Not fixable

One-year sample, so small majors are unreliable

Built on the ACS 2023 1-year file. Entropy — the basis of fallback width — is systematically underestimated at small samples: roughly 23% low at n ≈ 1,000 and 68% low at n ≈ 100. Majors below n = 1,500 are flagged as sample-limited and should not be ranked against large ones. Miller–Madow bias correction is already applied, which reduces but does not remove the problem.

Fixable — 5-year file

The postings line counts all seniority levels, not entry-level

This is the sharpest gap between what the tool shows and what it is about. The door-velocity chart counts every opening in a field, senior roles included — while the entire premise of the project is that entry-level and senior hiring have decoupled. Occupation × seniority postings data is paid-only in the United States (Lightcast, $22k–136k/yr); every published study on this question licensed it. The free national signal is directional only: entry-level postings −7.5% YoY against senior +14.7% (Indeed, May 2026).

Not fixable free

AI exposure is a prediction, not a record of replacement

The exposure score rates what a language model could plausibly do with an occupation's tasks. It is not a measurement of jobs actually automated, and a high score is not evidence that anyone lost work. Observed-usage datasets exist — Anthropic's Economic Index and Microsoft's Copilot study both measure what people actually delegate to AI — and compositing them in is the right next step, with the disagreement between predicted and observed shown rather than averaged away.

Partly fixable

It measures people employed now, not this year's graduates

ACS is a cross-section. The 22–32 cohort is the closest available proxy for recent graduates, but it is still a stock of everyone in that age band, not a flow of one year's leavers. This matters most for route divergence, where ordinary career progression and a changed entry market look identical. Only a longitudinal panel could separate them.

Not fixable

Wages ignored cost of living

Majors do not distribute evenly across the country, so a national wage median silently compares graduates living in different price environments. This is now corrected: each graduate's wage is deflated by the BEA regional price parity of the state they live in, and the toggle above the earnings chart switches between as-reported and cost-adjusted. The geographic concentration behind each major is shown under the chart, so the size of the correction is traceable rather than asserted. Adjustment is at state level; within-state variation between a metro and a rural county is not captured.

Fixed

Where every number comes from

Fallback width and exit distribution — ACS 2023 1-year PUMS (public domain). 644,541 employed bachelor's holders processed in full.
AI exposure — Eloundou et al., “GPTs are GPTs” (MIT). Production will composite this with O*NET 30.3 (CC BY 4.0), Anthropic Economic Index (CC-BY), and Microsoft Working with AI (CC BY 4.0).
Postings trend — Indeed Hiring Lab Job Postings Index (CC-BY 4.0).
National entry-level signal — Indeed Hiring Lab, 23 Jul 2026; Stanford Canaries Dashboard (no license stated — cited, not redistributed).
Counter-explanation — EIG, “Looking for the Ladder” (Jan 2026); Stanford, Feb 2026.

Design red line — this tool never issues a verdict. No "trap major" labels, no risk scores, no rankings, and it will never say "don't study this." Individual variation — ability, school, network, luck — far exceeds the differences between majors shown here. Every number on this page is a group statistic, not a prediction about any one person.