Method & data

What this tool estimates

For a target occupation and a U.S. region, the tool estimates how many workers already hold skill-adjacent jobs, how many of them change occupations in a typical year, what wage incentive a move would offer, and which specific skills a training program must close. It is an open re-implementation of the Workforce Supply Chains methodology developed at Carnegie Mellon University's Heinz College.

Lineage & citations

• Carnegie Mellon University, “Identifying the workers we need — and where to find them” (Work That Matters series), describing the original tool at workforcesupplychains.org.
• Gonchar, E., Combemale, C., & Krishnan, R. (2023). Workforce Insight Tool Methods Report. SSRN 4671063 — the analytical foundations: regional labor stocks by occupation connected to O*NET-derived skill-requirement similarity.
• Gonchar, E., Combemale, C., & Krishnan, R. (2025). Build It and They Will Come? U.S. Regional Labor Composition & Readiness to Meet Skill Demand Shocks from CHIPS and Science. SSRN 5126344 — applies the method to semiconductor production jobs, incorporating occupational transition rates and regional wage distributions.

How skill similarity is computed

• Each of 734 detailed SOC occupations is a 161-dimension vector of O*NET 29.3 level ratings across four domains: skills (35), knowledge (33), abilities (52) and work activities (41), averaged from 8-digit O*NET-SOC codes to 6-digit SOC.
• Similarity between two occupations is the cosine of their mean-centered vectors. Centering removes the baseline requirements every job shares (reading, communication), so similarity reflects what is distinctive about each occupation.
• The default threshold of 0.65 was calibrated against the figures CMU published for electricians in Pittsburgh: ~46,000 workers in adjacent occupations (this rebuild: 45,160), fewer than 3,000 annual transitions (≈2,100–3,000), and ~130,000 reachable with expanded training (133,610 at threshold 0.40).
• Training gaps are the positive differences between the target's and a feeder's O*NET levels, weighted by regional feeder employment — the “most salient missing skills.”

Data sources

• Regional employment & wages: BLS Occupational Employment and Wage Statistics, May 2024 — all 393 metropolitan and 137 nonmetropolitan areas, detailed occupations.
• Annual transition propensity: BLS Employment Projections, Table 1.10 — the occupational-transfer rate (2024–34 annual average), i.e. the share of an occupation's workforce that moves to a different occupation each year.
• Skill profiles: O*NET 29.3 database (U.S. Department of Labor / O*NET Resource Center).
• Crusoe sites: compiled June 2026 from the Crusoe newsroom and Data Center Dynamics coverage; legacy flare-gas modular sites are excluded (sold to NYDIG, March 2025).
• Cotulla footprint: cotullaeducation.com, AIM campuses, Centura College, Tidewater Tech, and the American Lineman College acquisition release (Oct 2025).

First-time workforce entrants

The skill-adjacency model counts only career-switchers. Most intake for an entry-level trades program is instead people entering the workforce, so the tool adds a second supply stream, estimated by scaling national ratios to each region's total OEWS employment:

• Commute-radius catchments (Crusoe site regions): a training program recruits graduates from every high school within driving distance, not just the OEWS metro, so each Crusoe site uses a ~60-mile catchment: county population sums (Census 2024-vintage estimates, rounded) × 1.23% turning 18 yearly × 87% HS graduation (NCES) × 38.6% not enrolling in college (EducationData.org — 61.4% enrollment, a 30-year low). Catchment populations are a research snapshot — verify before citing.
• Other regions (fallback): annual non-college-bound entrants ≈ total regional employment × 2.8% (the ~4.2M-person 18-year-old cohort against ~151.7M OEWS jobs) × 38.6%.
• Unemployed reserve: catchment (or regional) employment × u/(1−u) at the state unemployment rate (TX 4.4% / OK 4.2% / CO 3.9% / LA 4.5%, mid-2026; 4.2% national default).
• These are estimates, not measurements — they assume local age structure mirrors the nation. Appetite evidence (98% of Gen Z would pursue trades if training were funded — Cengage; 52% of pre-apprentices waitlisted — DEWALT) is shown as context and never multiplied into the math.

Recruitment funnel assumptions

The funnel card works backwards from an annual hiring target using benchmarked conversion rates (each editable in the card). Media cost = impressions × CPM / 1,000 — media cost only; recruiter time and training delivery are excluded.

• Impressions → clicks: 2.5% — midpoint of the general 1–5% digital job-ad CTR range (assumption; replace with smoke-test actuals when available).
• Clicks → applications: 4.22% — Appcast 2025 Recruitment Marketing Benchmark, Construction & Skilled Trades segment, 2024 (trades-specific).
• Applications → interviews: 8.4%; interviews → offers: 36%; offers → accepted: 90% — Jobvite 2017 Recruiting Funnel Benchmark. These are cross-industry ATS-era figures for experienced hiring, not trades- or program-specific — deliberately conservative. The “program intake rates” preset (30% / 60%) models inclusive pre-apprenticeship screening for aptitude rather than experience; applicant abundance supports it (IBEW Local 134: 1,000–4,000 applicants for ~200 slots; Meta LevelUp: 35,000 applications for 1,000 seats in one week).
• Hires → deployment-ready completers: 45% — Mathematica’s 10-state DOL Registered Apprenticeship study. That figure describes multi-year registered apprenticeships; shorter pre-apprenticeship programs typically complete at higher rates, so treat it as a conservative floor.
• Reconciliation: at benchmark rates, one hire requires ≈34,828 impressions. The source workbook’s Scenario 1 (2,150 hires/yr across TX/OK/CO/LA) therefore needs ≈74.9M impressions — $915,900 at a $12.23 CPM, i.e. the workbook’s $426 marketing cost per hire ($947 per deployment-ready completer). Scenario 2 (4,500 hires/yr) reproduces $651/hire at an $18.69 CPM (higher CPM reflecting audience saturation at roughly double the volume).
• Crusoe region presets split each state’s scenario target across Crusoe regions by hand-set shares (shown in the card) — scenario allocations, not forecasts. Impressions are ad views, not unique people; applications need not come from the skill-adjacent pool, so the supply comparison is context, not arithmetic on the same population.

Assumptions & caveats

• “Data center technician” is not an official SOC occupation; it is mapped to SOC 15-1232 (Computer User Support Specialists), the closest match used by training programs.
• O*NET publishes no ratings for SOC 51-2028 (industrial wirers / electrical assemblers); its skill profile is proxied from the mean of five related assembler and repairer occupations. Its employment and wages are real OEWS data.
• Est. annual movers applies each feeder's economy-wide transfer rate to its regional employment; it estimates churn available to recruit, not moves into the target specifically.
• Skill transferability is not guaranteed labor supply — wages, licensing, unionization and local conditions gate actual moves (the original tool's FAQ makes the same caveat).
• Training-program listings are an agent-assisted research snapshot (June 2026) covering Cotulla campus metros and Crusoe site regions — verify before citing. Where a regional median wage is suppressed, the tool falls back to an employment-weighted national median, or to your manually entered offered wage.