Methodology & Data Changelog

AI Exposure Blending

How FutureGrid measures the degree to which AI tools are already used in, or could affect, an occupation.

Consensus & Gap

The 'consensus' lens is the unweighted average of Usage, Capability, and Ability — the three modern, AI-specific lenses — for occupations where all three are available. When one or more lenses are missing, consensus is the average of the available modern lenses. The automation baseline is intentionally excluded from the consensus average because it measures a broader, older concept of task computerization.

The 'gap' (Capability minus Usage) indicates how much further AI could displace occupation tasks beyond current observed usage. A high positive gap suggests the occupation is structurally exposed but workers have not yet adopted AI heavily.

WARN Pressure Ranking

Which states are ranked, how the pressure score is built, and why many jurisdictions show null instead of a score.

Ranking Eligibility

A state is eligible for ranking only if all three conditions are met: (1) the state has machine-readable WARN Act notice data available (not manual-only or unavailable), (2) BLS LAUS labor-force data is valid for the latest reported month, and (3) at least one WARN notice falls within the rolling 12-month lookback window ending at the most recent data build date. States that fail any condition receive a null pressure score — NOT zero.

Manual-only states (those that publish WARN data as PDF or HTML tables only) and states with no machine-readable coverage are not ranked. Their null value signals a data gap, not the absence of WARN activity.

Pressure Score Formula

For eligible states, the pressure score blends WARN employee notices (normalized per 10 000 labor-force) with the year-over-year change in the state unemployment rate. Both components are percentile-ranked across all eligible states in the current build. Higher scores indicate both more WARN-notice volume relative to labor force and worsening unemployment rates.

12-Month WARN Window

Only notices with an effective date falling within the 12 months prior to the data build date are counted. Notices outside this window are excluded to keep the ranking current and comparable across builds.

Market-Signal Scoring

How the 0–100 Market AI Sensitivity Score is constructed and what it does — and does not — mean.

Score Construction

The marketAiSensitivityScore is a clamped 0–100 descriptive blend of two normalized components: 65% weight on the sector ETF's excess total return versus the S&P 500 (SPY) benchmark, and 35% weight on the employment-weighted AI exposure of occupations mapped to that sector. Both components are min-max normalized across all available sector ETFs in the current build window.

Data Source

The benchmark comparison window starts from a fixed date (typically 12 months prior to data build) and uses daily price observations sourced from the Yahoo Finance chart endpoint. Because the Yahoo Finance endpoint is unofficial and redistribution is prohibited under Yahoo's Terms of Service, the raw market-signal dataset is not available for download.

Forecast, Regression & Disruption Index

How employment forecasts, the Pearson correlation, and the Disruption Index are computed.

Linear Regression & Pearson Correlation

For each scatter analysis (AI exposure vs. employment growth; AI exposure vs. wage growth), FutureGrid applies ordinary least-squares linear regression using all occupations with finite values for both variables. The Pearson r coefficient is computed from the same pairs. These are purely descriptive statistics on historical BLS OEWS data (2016–2025) — they describe associations, not causal relationships.

2030 Employment Forecast

The baseline 2030 forecast extrapolates each occupation's historical employment trend (a CAGR computed from its BLS OEWS history) linearly from 2026 to 2030. The AI-adjusted forecast applies a sensitivity multiplier (default 0.5) that nudges the projected growth rate downward proportional to the occupation's AI exposure score. Both are projections based on trend extrapolation — not economic models. Actual 2030 employment will differ.

Disruption Index

The AI Disruption Index score (0–100) for each occupation is a weighted combination of four percentile-ranked components: AI exposure (40%), employment decline rate (25%), wage stagnation (20%), and absence of a BLS 'Bright Outlook' designation (15%). Each component is min-max normalized across all occupations with complete data. The index describes relative structural pressure — it is not a prediction of layoffs.

H-1B Work-Visa Trends

How the certified H-1B LCA metric is derived from DOL OFLC disclosure data, and why it is a demand signal for filings rather than a count of visa approvals.

What the Metric Counts

The headline metric is the number of certified H-1B Labor Condition Applications (LCAs) per fiscal year. An LCA is the wage-and-working-conditions attestation an employer files with the Department of Labor's Office of Foreign Labor Certification (OFLC) before petitioning for an H-1B worker. A certified LCA is an approved filing — it is NOT a visa approval, a petition, or a hire, and a single LCA can list multiple worker positions.

Per-Quarter Distinct-Case Summation

OFLC publishes LCA disclosure files quarterly for FY2020 onward and as a single annual workbook for FY2016–FY2019. The quarterly files are per-quarter snapshots, not cumulative, so a full fiscal-year figure is computed as the union of DISTINCT case numbers across all four quarterly files — de-duplicating cases that appear in more than one quarter. FY2016–FY2019 use the annual disclosure workbook directly. SOC occupation codes are normalized to the 2018 SOC vintage via a crosswalk so all ten years are comparable.

Wage Annualization

Offered wages are reported on hourly, weekly, bi-weekly, monthly, or yearly bases. Every wage is annualized to a common yearly figure before the median and the 25th/75th percentiles are computed, so the wage trend reflects like-for-like full-year compensation across the decade.

AI-Exposure Tier Join

Each occupation's SOC code is joined to FutureGrid's occupation-snapshot automation-risk tiers (Low, Medium, High, Very High) to show certified-LCA volume by AI-exposure tier. About 72% of H-1B SOC codes match; unmatched codes are bucketed as 'Unclassified'. This is a descriptive overlay of two independent datasets, not a causal claim about AI and visa demand.

Data Changelog

Dataset vintages sourced from the provenance registry (data/provenance.json). Sorted by most-recent build date.

Download Data

Compliance-cleared datasets are available for download. Files flagged with redistribution restrictions are listed as unavailable.

Available Downloads

Flagged / Restricted

See COMPLIANCE.md: data/COMPLIANCE.md