What will the average unemployment rate for recent college graduates in the US be in 2035?

TL;DR

My point forecast is 5.3%, and the distribution's median is 5.0%. The central 80% range is 3.4% to 7.4%, with a 7% chance of exceeding 8%. The main driver is whether the post-2021 jump in recent-graduate unemployment relative to the wider labor market persists through 2035.

Context

The target is the national unemployment rate for people aged 22–27 with a bachelor's degree or higher who are not enrolled in school. The New York Fed derives it from CPS/IPUMS data, seasonally adjusts it, smooths it with a three-month moving average, and normally updates it each February, May, August, and November. The latest vintage available on July 23, 2026 was published on May 5, 2026 and reported 2026:Q1 at about 5.7% (New York Fed series and metadata).

That 5.7% reading is high even though the broader labor market is not in recession: the national unemployment rate was 4.2% in June 2026 (Federal Reserve Monetary Policy Report). The resolution is less volatile than a monthly reading because it averages all four 2035 quarterly values.

Evidence

The historical backbone is the official monthly download and official metadata, May 5, 2026 vintage, retrieved July 23, 2026. Coverage is January 1990 through March 2026: 435 monthly observations, 145 reconstructed calendar quarters, and 36 complete calendar years. Units are percentage points of the labor force. The complete-year target history has a mean of 4.53%, median of 4.21%, standard deviation of 1.21 points, minimum of 2.77%, and maximum of 7.98%. The table gives the full annual history used in the model; the comparator is the New York Fed's internally consistent all-workers series, not headline U-3.

Simple mean reversion is the lower anchor. A quarterly AR(1) fitted over 1990:Q1–2026:Q1 has a coefficient of 0.84, a long-run mean of 4.64%, and a shock half-life near four quarters; its 2035 median is 4.44%. A no-break regression gives recent-graduate unemployment $=1.142+0.619\times$ the all-workers rate. Adding a post-2021 level shift raises fit from $R^2=0.61$ to $R^2=0.78$, with a +1.63-point shift. The raw gap changed from -1.29 points on average in 1990–2019 to +0.81 in 2021–2026:Q1, reaching +1.47 in 2026:Q1. Those calculations use the same New York Fed download.

Backtests argue against a narrow interval. At four nine-year forecast origins, the AR model's median versus the later realized annual value was 3.48% versus 6.56% for 2000→2009, 4.07% versus 5.27% for 2005→2014, 6.99% versus 3.92% for 2010→2019, and 4.53% versus 4.79% for 2015→2024. The model catches ordinary mean reversion but misses the timing and aftermath of large shocks.

The macro center is near 4.2% for aggregate unemployment. The June 2026 FOMC longer-run median was 4.2%, with a 3.8%–4.5% participant range, while CBO projected the rate returning to 4.2% from 2032 onward (FOMC projections; CBO outlook). These are no-shock anchors, not unconditional forecasts of the business-cycle phase in 2035.

The post-pandemic gap has a plausible durable cause. New York Fed research estimates that remote work explains 64% of the increase in unemployment for a closely related young-college-graduate group between 2017–19 and 2022–24, through weaker training and mentorship on distributed teams (New York Fed remote-work study). The timing predates broad use of generative AI.

AI is an upper-tail risk, not a clean central estimate. New York Fed job-posting data through early 2026 showed no junior-versus-senior divergence inside highly exposed occupations (New York Fed AI postings study). Anthropic found no systematic unemployment increase in highly exposed work, but estimated a 14% relative decline in young workers' job-finding rate in the most exposed occupations (Anthropic study). Stanford's balanced ADP panel also found weaker employment growth for workers aged 22–25 in highly exposed occupations, but the panel covers 25,000 continuing firms rather than the whole economy (Stanford Digital Economy Lab).

Two forces pull the forecast down. BLS projects employment in occupations typically requiring a bachelor's degree to grow 5.6% from 2024 to 2034, versus 3.1% for all occupations, though that covers all jobs rather than only entry-level roles (BLS projections). WICHE projects high-school graduates to peak at 3.9 million in 2025 and fall 13% by 2041, which should gradually restrain the supply of new degree holders (WICHE projections).

I combined the evidence as a four-component lognormal mixture. The weights reflect model disagreement, not literal observable regimes.

The resulting expected value is 5.27% and the median is 5.00%. The central 50% interval is 4.07%–6.14%, the central 80% interval is 3.39%–7.42%, and the central 90% interval is 3.04%–8.38%. The distribution assigns 27% to an outcome above 6%, 14% above 7%, 7% above 8%, and 2% above 10%.

What's non-obvious

The current weakness is not simply the usual effect of a soft economy. The recent-graduate rate used to sit well below the all-workers comparator, but it now sits above it. A forecast that mechanically returns to the 1990–2025 mean misses that sign reversal; a forecast that carries 5.7% forward for nine years misses the series' fast historical mean reversion.

The evidence also does not support adding a large AI penalty directly to the median. The deterioration began before ChatGPT, and current posting and CPS evidence remains mixed. I put AI mainly into the right tail. The central forecast stays near 5% because remote-work frictions look persistent, while employer adaptation, faster growth in bachelor's-level occupations, and a smaller youth pipeline offset them.

Limitations

The target has only 36 complete annual observations and four clear recession episodes, including the unique 2020 shock. That makes the upper tail much less certain than the center. The pseudo-backtests confirm that nine-year intervals built from one model under-cover when the business-cycle regime changes.

The New York Fed does not publish an exact sampling error for this narrow CPS subgroup. Its three-month smoothing and the four-quarter resolution average reduce noise, but do not remove it. The metadata also says October 2025 was estimated because the CPS observation was missing.

The largest unresolved question is structural. Remote work, AI, college enrollment, immigration, and employer training could interact in ways absent from the 1990–2026 sample. The numeric distribution is conditional on the named series, or an eligible direct successor, producing a valid resolution rather than the question being annulled.