The most useful result in the New York Fed's evaluation of LinkedIn's labor market tightness indicators is a negative one. Two of the indices share the same numerator, "attractive" job openings, and differ only in what sits below the line. Divide by the number of jobseekers and you get the best inflation predictor in the study, beating the vacancies-to-unemployment ratio on both price and wage inflation. Divide by the number of applications and you get one of the worst. The mechanism is the lesson. Applications are a behavioral margin that platforms, software and desperation all inflate. Jobseekers are a headcount, and a headcount is much closer to the scarcity that actually drives wage pressure.
Intuition gets this backwards. Application counts look like the richer, more granular signal: more observations, finer resolution, closer to the moment of matching. In practice that granularity mostly measures how cheap it has become to apply.
Three Indices, One Shared Numerator
Elizabeth B. Li, Giorgio Topa and Jacob P. Weber, writing on Liberty Street Economics on October 30, 2026, test three LinkedIn series against standard benchmarks over July 2019 to April 2026.
The Jobs per Application Index (JPAI) divides attractive openings in a month by applications submitted on LinkedIn. The Jobs per Jobseeker Index (JPJI) divides the same openings by the number of members who submitted at least one application that month. The LinkedIn Hiring Rate (LHR) counts members who added a new employer to their profile, scaled to the relevant member base and indexed to the average month of 2016.
The filter on the numerator matters more than it first appears. An opening counts as "attractive" only if the employer either paid to promote it or listed it manually, which means the poster spent money or time. It must also have drawn at least one application from a member in that local market. That screen is a direct answer to the phantom-vacancy problem that has dogged JOLTS openings: postings left up for branding, pipeline-building or compliance, with no live intent to hire. A costly signal from the employer, plus evidence of real demand from the worker side, is a reasonable definition of a job that exists.
The benchmarks are the vacancies-to-unemployment ratio (V/U), built from JOLTS openings and BLS unemployment, and, for the hiring series, the JOLTS hires rate.
Why the Denominator Decides the Ranking
The forecasting horse race is clean. In nowcasting, the authors run rolling regressions of inflation through the last month of official data on each tightness measure, then use the current month's reading to predict current inflation. They score each measure by root mean squared error relative to V/U. The JPJI posts the lowest nowcast error for both monthly CPI and quarterly Employment Cost Index (ECI) inflation. In one-period-ahead forecasting it again ranks first on both. The JPAI is consistently worse than V/U, with one exception: it edges V/U when forecasting ECI.
Since the numerator is identical, the gap between JPJI and JPAI comes entirely from the ratio of applications to jobseekers, that is, applications per person. That ratio is not a tightness variable. It moves with:
- Application technology. LinkedIn counts both Easy Apply submissions and click-throughs to external application pages. Any feature that lowers the cost of a click raises the denominator without changing how many people want work.
- Search intensity. When candidates feel insecure, each one sends more applications. That raises the JPAI's denominator even if the pool of seekers is unchanged, so the index tilts toward "loose" on sentiment rather than on supply.
- Automated and AI-assisted applying. Tools that generate and submit applications at volume push applications per seeker up mechanically. The study does not isolate this channel, and it should be treated as a plausible contributor rather than a measured one. It does point the same direction as the result.
Wage pressure comes from employers competing for a scarce body of workers, not for a scarce stack of résumés. A firm facing 400 applications from 40 people has the bargaining position of a firm facing 40 candidates. The JPJI sees 40. The JPAI sees 400 and concludes the market is ten times looser than it is.
This is also why the JPJI's correlation with V/U, 0.733 over the sample, is high but well short of identical. The two disagree in specific periods. The JPJI ran slightly above V/U in the early post-pandemic aftermath and slightly below it from late 2022 to mid-2024. That second window is when much of the debate over the "immaculate disinflation" centered on whether JOLTS openings were overstating labor demand. A reading that the JPJI's filtered openings were stripping out exactly the soft postings that inflated V/U fits the pattern. The authors do not make that claim, so it stays an interpretation, but it is the obvious place to look for why the LinkedIn index forecast better.
Hiring Flows Versus Hiring Pressure
The hiring-rate comparison produces a second puzzle. The LHR and the JOLTS hires rate are nearly the same series. Normalized and measured from the start of 2021, their correlation is 0.95. Yet they rank very differently. The LHR beats V/U in forecasting both CPI and ECI and beats it in nowcasting CPI, though it trails V/U for the ECI nowcast. The JOLTS hires rate is worse than V/U across the board.
Two series correlated at 0.95 can still diverge in predictive power, because forecasting accuracy lives in the residual 5% and in timing. A small, persistent lead in turning points, or less noise month to month, is enough to separate them in RMSE terms. The excerpt does not establish which of these is at work, and it does not say whether the exercise used first-release or revised data. If it used revised data, some of the JOLTS hires rate's real-world handicap, its revisions and survey response problems, would not even show up in the scores.
The broader point holds regardless. Hiring is a flow and tightness is a ratio of stocks. A high hires rate can reflect a fluid market with easy matching rather than a scarce one with bidding wars. That ambiguity explains why both hiring measures sit below the JPJI, and why the LHR struggles specifically on the wage side in nowcasting.
The Timing Advantage Is the Policy Story
Accuracy is only half of the result. LinkedIn can publish a month's data in the first week of the following month: August figures in early September, in the authors' example. JOLTS arrives with roughly a five-week lag. For a central bank that leans on V/U as a summary of labor tightness, a series that is at least as accurate and lands about a month sooner is a material upgrade in the information set.
The channel to markets runs through the policy path. Tightness measures feed wage and inflation forecasts, and those forecasts feed rate expectations at the front end of the curve. An indicator that front-runs JOLTS by a month gives forecasters a cleaner early read on whether labor scarcity is reasserting itself or fading. How heavily the Fed will lean on a proprietary series is unknown. The fact that its own research staff published a favorable evaluation is the first step toward these numbers showing up in staff nowcasts and, from there, in the questions asked at press conferences.
Where the Evidence Is Thinner Than the Ranking
The strongest case against over-reading the result is the sample. July 2019 to April 2026 holds one pandemic collapse, one inflation surge and one disinflation, essentially a single giant cycle. Any variable that rose and fell with the 2021-2022 episode will score well on RMSE, and ranking measures by average error over one dominant event is fragile. The ECI is quarterly, so wage-side comparisons rest on fewer than thirty observations. The excerpt also reports relative errors against a V/U baseline of 100 without the magnitudes, so the size of the JPJI's edge cannot be judged from the text available here.
Representativeness is the second weakness, and the authors flag it. LinkedIn's user base skews toward professional and white-collar work. The service, construction and hospitality labor markets that drove much of the 2021-2022 wage surge are thinly represented. The JPJI may have forecast well partly because broad tightness happened to move together across sectors in this period. In a downturn concentrated in one segment, a professional-services hiring freeze for example, a LinkedIn-weighted gauge could overstate or understate aggregate slack.
Third, the definitions belong to the platform. The "attractive opening" screen depends on employers paying for promotion, which makes the numerator sensitive to recruiting budgets and LinkedIn's pricing. Product changes to Easy Apply or membership composition can shift the series in ways unrelated to the economy. Official statistics carry documented methodology and revision histories. A vendor index can change underneath its users.
None of this overturns the denominator result, which is the most robust finding precisely because it compares two series sharing the same numerator, platform and sample. Whatever biases the attractive-openings count carries, they hit the JPJI and the JPAI equally. The gap between them is about how tightness should be measured, and it favors people over applications.
What Would Confirm or Break the Read
The test is out-of-sample performance in a regime that does not look like 2021. If the JPJI keeps beating V/U on CPI and ECI through a period of modest, sector-specific softening, the case for treating it as a primary tightness gauge becomes strong, and its one-month lead becomes a genuine early signal for wage forecasts and front-end rate expectations.
The most informative moment will be a divergence. When the JPJI and V/U next separate by a meaningful margin, the subsequent ECI prints will show which one was measuring scarcity and which one was measuring noise. And if applications per jobseeker keep climbing as automated applying spreads, the gap between the JPJI and the JPAI should widen further. That would be the clearest evidence yet that the application count tracks the cost of a click rather than the price of labor.
