How AvgPay reaches its conclusions.

We turn career information into our own models, scores and comparisons. This page explains the calculations, the guardrails and the underlying research.

Salary forecasts: start with the model that survives testing

We forecast the next published national annual median wage for an occupation's existing statistical category. This is not an individual's earnings after a promotion, total compensation, an offer estimate or a prediction of AI job losses.

Comparable national occupational releases begin in 2022, after the 2021 MB3 estimation change and the 2022 wage-processing change. Matching uses the same detailed SOC. Historical estimates share overlapping survey panels and therefore do not form an independent worker wage panel. We currently support 808 occupations with a complete matched history.

Time-based evaluation: wage and completed CPI observations through 2023 were used to predict the 2024 release for model selection and range calibration. Observations through 2024 then predicted the held-out 2025 release. Each origin is assumed to be May 1 of the following calendar year, when the inputs would be available. The one-year horizon is between wage reference years; because publication lags, the operational lead can be short. We do not use future inflation, future projections, current job postings or later-released employment data at earlier origins.

The selected model is unchanged real. The inflation-persistence baseline calculates median(year + 1) = latest observed median × latest completed annual CPI / prior annual CPI. We use an occupation-specific trend only when validation supports it. Future purchasing power uses the displayed inflation assumption; when inflation persistence is selected, the model implies unchanged real wages. This is an assumption, not a macroeconomic inflation forecast.

Equal-weight occupation errors; current matched population, not employment weighted
Model2024 development MAPE2025 holdout MAPE2025 MAE80% interval coverage
unchanged nominal4.19%4.03%$ 2,87381.4%
unchanged real (selected)2.75%2.68%$ 1,91381.2%
recent trend4.26%4.34%$ 2,97582.3%
half trend half inflation3.25%3.27%$ 2,27281.3%

We compare unchanged nominal wages, unchanged real wages using inflation persistence, continuation of the most recent occupational change, and an equal log-rate blend of the latter two. A challenger needs at least a 3% relative improvement over the better simple baseline in development testing. Its choice is frozen before the held-out outcome. A challenger that then fails to outperform the baseline falls back to that baseline; any such fallback is reported in the downloadable validation artifact.

Plausible range: the intended 80% interval uses the finite-sample upper 80th percentile of absolute log prediction errors from the 2024 calibration release. Lower and upper wages multiply the point estimate by exp(−q) and exp(q). We measure coverage on 2025 without using 2025 residuals to set the width. Mean absolute percentage error (MAPE) measures absolute error divided by the observed outcome; mean absolute error (MAE) is in dollars.

There is only one held-out calendar release, and all careers share economic conditions and survey design. The number of occupations is not the number of independent years. Model selection and interval calibration share the 2024 development fold, so these are experimental empirical ranges, not guaranteed confidence intervals or occupation-specific accuracy claims. We also condition on occupations observable in the latest catalog; discontinued and suppressed series are excluded.

Horizon and location gates: this model version separately validates only one year ahead. Three- and five-year estimates remain scenarios until they receive their own range calibration and temporal holdout evaluation. Local histories have not been harmonized across the 2024 geographic revision, so local estimates are also scenarios. Editing wage growth converts any result to a scenario. A missing matched history produces an explicitly labeled scenario.

Point estimates display to the nearest $100. One-year estimates may be nowcasts of an unpublished past reference period. Forecast input hashes are checked against canonical wage, CPI and history files on every build. Download forecasts and validation · Download matched release history · OEWS comparability limitations

Regional buying power and working conditions

Price-adjusted wage equivalent: metro median wage ÷ (all-items regional price parity / 100). The premium is that equivalent divided by the national occupational median, minus one. Regional prices cover 438 state and metro geographies in 2024. We join exact state FIPS or metro CBSA identifiers, never fuzzy city names; wage and price years remain separate.

Price and wage vintages can differ by at most one year. When regional prices lag the wage release, the result is labeled as a lagged price proxy. We require at least 300 estimated workers in each occupation/metro. The price index describes relative regional consumer prices, not a household budget, taxes, disposable income, housing affordability or an individual's spending. Missing regions and suppressed values are omitted. Both series use the OMB 2023 area delineations; future geography revisions require review before a join.

Working conditions: exact O*NET-SOC dossier matches, with no averaging across specialties. We display unsuppressed CX mean context ratings on the original 1–5 scale, including question-specific endpoint labels, sample size where available and survey update month. A mean is not a percentage. Different questions have different meanings; we do not sum them into an arbitrary fit score. People should use them to ask better questions about individual employers.

The quarterly source version is checked against the existing occupation profile release. New source files preserve checksums and licenses. The regional pipeline checks for annual releases monthly; forecast and work-context outputs rebuild after relevant input changes. Failed imports preserve the prior valid publication. See BEA regional price definitions and O*NET database license.

The three things to know

1

It's actual survey data.

Salary figures come from the Bureau of Labor Statistics' Occupational Employment and Wage Statistics (OEWS) survey, May 2025. They are estimates for occupations, not invented offers.

2

Median means the middle.

For an occupation, half the workers have wages above its median and half below. The 10th–90th percentile range covers the middle 80% of reported wages.

3

It isn't your guaranteed pay.

The survey does not tell us what a specific employer will offer. Benefits, stock awards, taxes, experience and local living costs can change the picture.

Why a career has more than one name

AvgPay uses concise, recognizable job titles to make salary research easier to find and understand. The published wage statistics are still keyed to their original Standard Occupational Classification (SOC) occupation groups. A familiar name such as Financial Analyst can describe a broader published wage group, Financial and Investment Analysts. When the names differ, the official coverage is identified on the career page.

Alternate titles, common abbreviations and reported role names help people find the appropriate guide. We do not create a separate page for every title, and an alias does not prove that two narrowly defined positions earn the same amount. Each fully researched occupation has one canonical page and one permanent redirect from its previous occupation-category URL.

The broader national library includes hundreds of additional verified wage occupations. Until those careers have sufficiently differentiated research, they are searchable in the explorer but do not generate thin, indexed detail pages. These naming choices improve usability; they are not claims of independently measured keyword volume or guaranteed search visibility.

How our original career scores work

AvgPay Pay Position: original occupation benchmarks

Pay Position compares verified BLS national occupational wage records rather than rewriting published occupation descriptions. Every occupational group is identified by its SOC code and most recent validated OEWS vintage.

Field pay rank equals one plus the number of occupations within the same two-digit SOC career field with a higher reported annual median. The numerator and denominator count occupations, not workers, and tied medians share a rank. The field benchmark is the sum of each occupation's median multiplied by its published employment, divided by total published employment in the field. A career's percentage difference is (its own median ÷ the field benchmark − 1) × 100. This is an employment-weighted index of occupation medians, not the median wage paid to individual workers in that field.

Relative wage range compares the published 90th-percentile wage with the 10th percentile: (p90 ÷ p10 − 1) × 100. We also rank that percentage against other occupations in the same field when at least five have both percentiles. Suppressed source values are never estimated. A wider published range does not mean stronger career progression or guaranteed pay increases.

Related higher-paying occupations are selected from the same broad field by their published median. These are comparisons, not proven promotions. Skills, education, licensing, location, experience and fit must be considered independently. All inputs recalculate from verified source data when a new release passes publication checks.

AvgPay Opportunity Index (0–100)

We rank each currently covered occupation against the others on four metrics: national median annual pay (35% weight), projected 2025–2035 employment growth (30%), average annual projected openings per 1,000 base-year workers (25%), and the number of metro markets with eligible wage/employment observations (10%). Each component is a percentile within the curated occupation sample, using midranks for ties. The weighted sum is rounded to the nearest integer. It is not a personal hiring score, forecast of return on education, measure of living standards, or national ranking of all occupations. The results change as coverage expands.

AI task exposure score (0–100): source and limitations

The AI risk score comes from the published occupation file of GPTs are GPTs (Eloundou, Manning, Mishkin and Rock; OpenAI and the University of Pennsylvania), released under the MIT license at github.com/openai/GPTs-are-GPTs. The study broke each job into its O*NET tasks and asked whether a large language model could do each task at least twice as fast with the same quality. Trained people and GPT-4 each rated every task.

We show the study's main measure: the share of a job's tasks a chatbot alone could do that much faster, plus half of the tasks that would also need AI software built on top. We average the human and GPT-4 ratings, average O*NET detail codes into the SOC code our pay data uses, and multiply by 100. The two other numbers on each job are the same calculation with none (chatbot alone) or all (with AI software) of those extra tasks. Levels: Low under 15, Moderate 15 to 34, High 35 to 59, Very high 60 and up. Field and national averages are weighted by the number of people in each job. Where the study does not list a SOC code (mostly "all other" groups), we use the average of related published jobs and say so on the page.

This is not the odds of losing a job. These are theoretical task-speedup ratings from research using 2023-era models. They do not measure actual adoption, work already automated, future job cuts or the capabilities of systems in 2026. We do not use them to predict individual employment outcomes.

Score bands are chosen by AvgPay, not calibrated probabilities. Low means 0–14, Moderate 15–34, High 35–59 and Very high 60–100. Of the 825 occupations, 58 lack direct ratings in the source and have related-occupation estimates, identified on individual results. Comparisons between jobs do not prove that changing occupations would be feasible or safer; education, licenses and experience must be considered separately.

Inflation and real pay comparisons

We use U.S. city average annual-average CPI-U from 2021 through 2025. For salary value S in starting year A, the equivalent spending power in ending year B is S × (CPI_B/CPI_A). Cumulative inflation is (CPI_B/CPI_A − 1) × 100. This is national all-items consumer inflation, not regional prices, tax-inclusive take-home pay, or an occupation-specific wage series. Illustrative salary inputs do not imply that the current median was earned in the earlier year.

When validated historical wage releases exist, the inflation-adjusted comparison of published national median estimates is ((median_B/median_A) ÷ (CPI_B/CPI_A) − 1) × 100. We do not extrapolate absent historical wage data. Because successive OEWS releases incorporate overlapping three-year samples, even inflation-adjusted changes in those published estimates are not direct measures of raises for individual workers.

Skills and knowledge connections

Each skills page is computed by matching a named capability across the curated occupation profiles. Importance scores are occupational ratings on a published 1–5 scale and comparisons are only among roles for which that rating is available. Related skills are ranked by the count of covered occupations that also report them; co-occurrence is not causality or training advice.

The AvgPay Career Intelligence Graph

Canonical identity: Each supported U.S. occupation uses its detailed SOC code and one canonical salary URL. Government names, editorial labels and O*NET reported/alternative titles are search mappings, not automatically different paid jobs, specializations or levels. Normalized titles that appear under more than one occupation remain ambiguous until additional evidence distinguishes them.

Documented related occupations: An O*NET published relationship means two occupations are described as related in its database. It does not imply hiring frequency, an observed move, a shared license or a promotion path. Connections use the actual source occupation identifiers and are labeled verified relation.

Modeled similarity: For independently profiled occupations in the same SOC major field, AvgPay compares their source-rated skills and knowledge. We take the union of their 15 strongest skill ratings and 12 strongest knowledge ratings that are at least 2.5 of 5, deduplicating common source concept IDs. The comparison is a weighted Jaccard index: sum of the lower value on each shared concept divided by the sum of the higher value over all concepts in either profile, multiplied by 100. We require at least five shared concepts, at least three shared skills and a minimum 42-point score, with no more than a one-level O*NET job-zone difference when both are reported. These display thresholds are editorial research filters, not validated employment-transition thresholds or probabilities. The top six per occupation are presented as derived research comparisons.

Preparation and skill differences: O*NET job-zone differences describe broad occupational preparation, not a personal skills gap or regulatory eligibility. The skills shown as a different emphasis are higher-rated in the target occupation than in the source, or are not among the source's prominent ratings. They are not an assessment of what any individual visitor knows.

Source rights: The O*NET occupational database is used under CC BY 4.0 with attribution to the U.S. Department of Labor/Employment and Training Administration. AvgPay modifies the source for matching, indexing and derived comparisons; USDOL/ETA has not endorsed or validated these modifications. Published wages remain the distinct BLS OEWS estimate for the full statistical occupation, not a modeled salary for an alias.

How the AvgPay Career Path Explorer works

Each career path begins with documented entry education, related experience and training, then lists concrete occupational tasks and skills. Related careers are drawn from occupational relationships; where we lack direct links, some options are labeled as comparisons based on occupational knowledge overlap. We compare the national median pay of different occupations, not typical promotions or personal salary increases.

Potential preparation gaps compare published job requirements. They cannot prove what a visitor personally knows or whether they are qualified, nor do they substitute for professional licensing, credentials, hiring experience or local opportunities. We do not invent how many years a transition takes.

Freshness and responsible publication

AvgPay is designed to check wage, occupation, skills, outlook and CPI releases daily. New data becomes part of the website only after source identity, dates, coverage, career relationships, inflation math, navigation and production-build checks pass. An annual release remains annual: refreshing the page cannot turn it into live salary information.

Automated checks currently depend on GitHub Actions runner availability. Its runner-provisioning issue means the scheduled checks are configured but not yet confirmed operational. Until a verified refresh run succeeds, the website serves the last validated snapshots. A source's publication date—not today's visit date—describes the data's age.

Our skill detail pages remain internally useful but are excluded from search indexing until each provides independently reviewable, differentiated value. No content is created solely to target keyword combinations. Explore the skills browser or see our published career research in the Decision Lab.

Published data and attribution

BLS occupational wages7,623 verified wage observations · 18 curated occupations · May 2025

Official BLS data ↗

O*NET occupational information755 published occupation profiles · version 31.0. We currently use occupational fields for related-job discovery; numerical skill scores are withheld.

Official O*NET data ↗

O*NET occupation profiles744 curated occupation dossiers · 14 versioned official tables · release 31.0. Includes descriptions, job titles, tasks, related occupations and preparation statistics.

Official O*NET database ↗

BLS Employment Projections825 occupation projections · 2025–2035. Employment projections are separate from occupational wages and not live job postings.

Official BLS outlook table ↗

CPI-U annual-average inflation2021–2025 national all-items consumer price index; annual averages, unadjusted. Used only for purchasing-power comparisons.

Official BLS CPI annual table ↗

More about how it works

How do you compare pay between cities?

We compare the published median wage for the same occupation in a metro area with the U.S. median. The percentage is based on observed wages only and is not adjusted for local living costs, taxes, equity or benefits.

What does “pay varies by city” measure?

We include metro areas with at least 300 workers reported in the occupation, then compare the 90th and 10th percentiles of metro median wages. Each eligible metro has equal weight. The result describes differences between metro medians, not a range for individual workers. This model is currently version 1.1.

What is “local career advantage”?

For an occupation in a metro, calculate the percentage difference between its local median and its own U.S. median. Then subtract the median local wage premium across at least eight other AvgPay occupations in that metro. All the included occupation-market observations require at least 300 estimated workers. The difference in percentage points reveals unusual relative pay within our limited curated comparison set. It is not a purchasing-power adjustment, an employment forecast or a statistic for the entire labor force.

What is the high-pay, substantial-employment frontier?

A metro is on the frontier when no other eligible metro has both a higher annual median wage and a higher estimated employment count for the same occupation. This is a deterministic Pareto comparison, not a ranking of job openings, living costs or places where you are most likely to be hired.

How do you compare national pay ranges?

We divide a job's published national 90th-percentile annual wage by its 10th-percentile annual wage and subtract one. The result measures the spread across workers; it is not expected career earnings growth. We do not fill in suppressed percentile values.

How are historical wage trends calculated?

Only after official historic BLS OEWS archives pass source, completeness and occupation-identity checks do we publish nominal changes in published annual median estimates. BLS uses overlapping three-year survey samples and warns that OEWS is not designed to measure short-term wage changes directly. If these records are unavailable, the timeline stays unavailable rather than invented. When validated history exists, AvgPay also reports the CPI-adjusted change in the published estimates, never an estimate of a worker's personal raise.

Where do job descriptions, alternate titles and task lists come from?

We use official O*NET Occupation Data, Job Titles, Sample of Reported Titles and Task Statements. Those sources describe an occupational classification; a specific employer may use a different title or assign narrower duties. Tasks are not a measured percentage of anyone's workday. Different O*NET-SOC suboccupations are not automatically combined.

What is an occupation's “knowledge fingerprint”?

AvgPay subtracts the median importance rating for an O*NET knowledge area across at least eight other covered occupations with that published measure from the occupation's own rating. This highlights subject areas that are unusually important for a role in our limited comparison set. It is not a nationally standardized percentile, a judgment of a person's skill, or a prediction of wages.

Are the listed career directions actual promotion paths?

No. They come from the official O*NET Related Occupations network. They are useful adjacent directions to investigate, but the links are not time-ordered career histories, verified promotion sequences, licensing determinations or probabilities of being hired. A comparison of knowledge areas only describes different occupational requirements; it does not assess a visitor's abilities.

How are education and preparation requirements represented?

We show two different sources: BLS typical entry-level education, related experience and on-the-job training classifications; and O*NET job zones plus respondent-assessed education distributions. O*NET education percentages do not represent the percentage of workers holding a degree or legal licensing requirements. Individual employers and regulated professions can impose additional requirements.

Are employment projections the same as job openings right now?

No. BLS's 2025–2035 projections estimate changes in total occupational employment and average annual openings that include replacement needs. A field can have declining projected employment but positive annual openings. AvgPay calculates annual openings per 1,000 baseline workers as a descriptive, normalized comparison, not an actual vacancy rate or the chance of being hired. These projections are separate from May 2025 OEWS wages.

How do you find related careers?

We currently show other occupations within conservative broad occupational fields. An earlier general-skill similarity score was insufficiently discriminative across unrelated jobs, so we stopped publishing numerical career-match scores. A future model must validate detailed tasks, knowledge, credentials and training requirements before suggesting a transition.

How do automatic data updates work?

Our scheduled daily process is designed to check for new BLS and O*NET releases, normalizes source data and requires duplicate, coverage, checksum and wage plausibility checks before publication. A separate historical-data stage only publishes fully verified archives. A hosted GitHub runner is required to execute the schedule; the current runner-provisioning issue means the automatic refresh has not yet been demonstrated end to end. These updates follow source releases, not live hiring.

How does the job-change calculator work?

It compares the pay, costs, growth assumptions and time horizon you enter. The displayed scenario is arithmetic, not a prediction, and does not include taxes, equity, unemployment risk or inflation unless you account for those effects yourself.

Ready to look up a salary?

Browse covered jobs →