How every number is computed

This edition analyses publicly accessible job postings on 104 Job Bank (104 人力銀行). SkillTrendLab is an independent research and analysis service with no affiliation, partnership, sponsorship or licensing relationship with 104.

The site does not provide or republish any original posting; it publishes aggregated statistics and trends only. There is one source at present, so the numbers describe public postings on that platform and are not the whole Taiwanese market. When another source is added, it is announced here.

An edition is one snapshot of the postings open on a given date. The numbers describe how many postings were open at that moment and what they looked like, not the flow of new postings and not the number of people hired.

A posting counts once per edition; if it is still open in the next edition, it counts again there. A skill is said to rise or fall by comparing this edition with the nearest edition at least 30 (or 90) days earlier, in percentage points.

The sample is this edition's number of core AI postings, stated at the top of this page; every share divides by it.

A job platform’s keyword search is fuzzy: searching "AI" returns estate agents and interior designers. Search results are only candidates. The decision is made by this site’s own rule, kept under version control with tests.

• A technical title carrying an AI term (artificial intelligence, machine learning, deep learning, LLM, generative, NLP, computer vision, data science, algorithm, MLOps and so on) is a core posting.
• A title with the bare word "AI" counts only when it is also a technical title (engineer, developer, researcher, analyst, data, systems, product manager and the like). Non-technical jobs wearing AI as a badge, such as "AI customer service" or "AI marketing", do not count.
• Teaching roles (cram schools, universities, lecturers, tutors) and staffing-agency postings are excluded: the former teach AI rather than do it, the latter post the same job several times over.
• A technical title with no AI term that mentions AI only in the description is flagged as "related" for internal quality checks and stays out of the published statistics.

The skill dictionary currently holds 190 skills, each with Chinese and English names, a category and aliases, under version control. Correcting the dictionary is correcting the method: the method version changes and every edition is recomputed.

• A posting "has" a skill on one of two kinds of evidence: the employer’s own skill tag maps to the dictionary, or the title or description matches its text pattern. Text matching requires whole words ("C" does not match "Vitamin C", "Go" does not match "Google") and excludes known false contexts.
• A skill counts once per posting however often it is mentioned. Penetration = postings with the skill ÷ core postings in the edition.
• Change is in percentage points: this edition’s penetration minus the penetration in the nearest edition 30 (or 90) days earlier. Skills with fewer than 30 postings are left off the movers list.
• The homepage ranking lists technology skills only (languages, frameworks, databases, cloud, data engineering, machine learning, LLM, DevOps, tools); competencies such as project management stay in the full index.

A role is decided from the title by an ordered list of rules: the first rule that matches wins. There is an AI track and a data track; a title that matches no rule goes to "other AI postings".

• Role share = postings in the role ÷ core postings in the edition. The "most asked" skills of a role are the skills with the highest penetration among that role’s postings.
• Salary statistics use only postings that state a monthly or annual figure (annual converted to monthly); negotiable, hourly and daily pay are excluded. Every salary figure is shown with the share of postings that disclosed one, because postings that say "negotiable" tend to pay more and the sample is biased.
• Below 30 postings disclosing pay, a role publishes no salary statistics for that edition: quartiles off a dozen postings are noise, and a number on a page is read as a fact.
• Annual pay is divided by twelve. Monthly figures under 10,000 or over 2,000,000 TWD are treated as typos and excluded.

Industries: the platform’s hundred-odd industry labels are folded into 18 groups; the mapping is under version control.

Places: by the city or county of the work location, 22 cities and counties in 3 regions. Postings abroad or without a city are left out of the place statistics.

A skill or role gets a page of its own once its sample is large enough and spread across enough companies. Below that it still appears in the rankings, without a page.

History gates something else: whether the page enters the sitemap and is offered to search engines. A page with one edition behind it can be read, and says on its face that it cannot yet speak about change.

Every edition passes a consistency check before it is published: fewer than 500 core postings, a population more than 25% away from the previous edition, or a top-20 skill moving more than 15 percentage points in one edition holds the edition back until a person has read the diff report.

Every number carries the method version and dictionary fingerprint that produced it (top of this page). When a rule or the dictionary changes, the version changes and every past edition is recomputed under the new rules, so editions are always measured with the same ruler.

Changes to the method are recorded on this page.

This is not a job board. It does not match candidates to jobs and never sends a reader to a posting. None of the following appears on the site:

• The content, title or link of any individual posting.
• A searchable list of postings.
• Lists of companies, or the postings of any one company.
• Original job descriptions, in full or in part.
• The source platform’s logo, interface or screenshots.

• One source: the numbers describe public postings on one platform, not the whole Taiwanese market.
• Stock, not flow: open postings are not hiring demand or hires, and a posting open for months is counted in every edition.
• Text matching has limits: a tool missing from the dictionary is invisible until it is added, and synonyms or abbreviations can be missed.
• Salary bias: only postings that disclose pay enter the statistics; negotiable postings are absent.
• A strict title rule: postings that mention AI only in the description are outside the population, so the numbers lean towards jobs that name AI explicitly.