Plan training on a forecast, not on last year's vacancies.

Job forecasts and course plans are made separately

The ILO reports that youth unemployment rose to 12.4% in 2025, about 67 million young people, and that more than 257 million young people were not in employment, education or training. In the WEF survey, 63% of employers named skill gaps as a major barrier to changing their business, more than any other barrier.

We think the forecast and the course plan should be built in the same units, skill by skill and region by region. Job forecasts count occupations, while courses teach skills.

What we would model, and with which data

Official projections come first, and we would add regional and skill detail.

• Labor force surveys and projections from national statistics offices. The US Bureau of Labor Statistics projects employment ten years ahead, and its 2025 to 2035 release projects 5.9 million more jobs.
• Regional forecasts, such as Cedefop's projections of employment by sector and occupation for the EU, Norway, Iceland, North Macedonia, Switzerland and Türkiye.
• Skills taxonomies. ESCO links 3,039 occupations to 13,939 skills in 28 languages, and O*NET profiles more than 900 US occupations.
• Job advertisements, read for the skills employers ask for now.
• Course catalogs and graduate numbers from colleges and training providers.
• Employer workforce plans, shared through industry bodies.

How the model is built and checked

Nobody should plan with a forecast that hides its range or its past errors. Statistical forecasting projects employment by occupation and region from past trends, population and economic scenarios, and gives each forecast as a range. A skills extraction step reads job advertisements and course descriptions and maps both to the same taxonomy, so demand and supply are compared skill by skill. Text embeddings, which turn sentences into numbers so that similar meanings sit close together, match each syllabus to the skills it teaches. Back-testing runs the model from an earlier year and compares its forecast with the jobs that appeared, and with official projections. Employers and educators review the skill mappings before anyone uses them.

What decision-makers get

• Labor ministries get occupation and skill forecasts by region, with ranges.
• Colleges get the skills employers ask for that current courses do not teach, and the courses closest to filling each gap.
• Employers get the internal moves that need the least retraining. Half of the employers in the WEF survey plan to move staff from declining to growing roles.
• Career services get plain guidance on growing occupations and the training that leads to them.

A first project picks one region and one sector with a shortage that employers and a college both feel. We would build and back-test the forecast, review it with both groups and hand over a scorecard of past accuracy, the largest skill gaps and the courses closest to closing them. The partners then decide whether to continue.

How it runs safely

Guidance affects a person's career, so it explains options and never decides who gets a course place or a job.

• Forecasts use aggregate data. Individual records, such as graduate outcomes, are used only under a data-sharing agreement, inside the data holder's systems, with personal details masked.
• Fairness checks test whether guidance differs by gender, age, disability or region for people with the same skills.
• Each forecast is published with its assumptions and version, so others can check it.
• In the EU, the AI Act classes AI that decides admission to education or training, or screens job applicants, as high-risk. This page is not legal advice.

Open-source building blocks

Open code lets a ministry keep and rerun its own forecast.

• Nesta's Skills Extractor, which pulls skill phrases from job advertisements and maps them to ESCO or Lightcast Open Skills.
• StatsForecast, a library of statistical forecasting models.
• Sentence Transformers, a library for text embeddings.
• Fairlearn, for checking whether guidance treats groups differently.