The Skill-Demand Trajectory Score: Rising, Plateauing, or Fading

A skill's mentions in job postings are climbing. Before you build a hiring plan or a training budget on it, you need to know whether that rise is a real, sustained trend or short-lived hype that fades in two quarters. This guide is for workforce planners, talent-intelligence analysts, and strategy teams sizing a skill market. It hands you a repeatable rubric to grade one named skill's demand trajectory from public data and defend a rising, plateauing, or fading verdict in front of a skeptical finance partner.

Most public content on skill demand is either a vendor listicle ranking "top skills" or an academic paper on extraction methods. Neither hands you a scored call on the one skill in front of you. This guide converts the scattered methodology into a single verdict: track share not levels, test multi-quarter persistence, adjust for season, measure cross-role spread, and correct for supply.

What the trajectory score decides

The score answers one question: is a skill's rising presence in job postings a durable trend worth planning around, or hype that will fade? The verdict has three grades and one attached expiry date.

The three grades are not a mood ring. Each maps to a specific evidence pattern you can defend line by line.

The word to underline is share. Indeed's public index defines its data as the percentage change in seasonally-adjusted job postings since February 1, 2020, using a seven-day trailing average, where a reading of 101 signifies postings are 1% higher than the baseline. That framing, share indexed to a fixed baseline, is the spine of every honest skill-demand read. Raw counts lie because total hiring volume moves under your feet.

The five dimensions that decide it

Score a skill on five dimensions: normalization, persistence, seasonality, spread, and supply. Each one proves something specific, and each one has a signature failure that looks like a signal until you check it.

• Normalization. Proves the trend is not just total-volume growth. It lies when raw counts rise but share stays flat.
• Persistence. Proves the trend has legs across time. It lies when a single quarter spikes and reverts.
• Seasonality. Proves the rise is not the annual Q1 hiring surge. It lies when a January bump reads as emergence.
• Spread. Proves the skill is a real capability cluster, not a niche fad. It lies when high demand share sits in one occupation and near-zero everywhere else.
• Supply. Proves demand outpaces the people who can fill it. It lies when postings rise but profiles rise faster, which is saturation dressed as shortage.

Read the stack bottom-up. If normalization fails, nothing above it matters, because you are looking at hiring volume, not a skill trend. If spread fails, a real share rise is still just one team's experiment. Supply sits on top because it converts a demand verdict into a sourcing verdict: a rising skill with abundant supply is easy to hire and cheap to ignore.

How to score each dimension

For each dimension, give a plain pass or fail, then a one-line reason. A skill needs a pass on normalization and seasonality to be scored at all; those are gates, not points. Persistence, spread, and supply then decide between rising, plateauing, and fading. Three passes on those three is rising. One or two is plateauing. Zero, or a fail on either gate, is fading.

The normalization anchors you can actually cite

Three public sources set the standard for how to normalize and seasonally adjust posting data, and each carries a different revision risk you must annotate. Pick one convention and hold it constant across your window.

Two things in that table matter more than they look. First, the seasonal method depends on data frequency. JOLTS uses X-13ARIMA-SEATS with a concurrent methodology, where new seasonal factors are calculated each month using all relevant data up to and including the current period. But Indeed seasonally adjusts each daily series using methodology developed by the Deutsche Bundesbank for daily time series data, with projected seasonal factors based on the preceding three years, adopted in November 2024. Applying monthly X-13 logic to a daily posting series misreads intra-week and holiday effects as trend.

Second, the revision column is where phantom growth hides. Lightcast reclassifies postings data every four weeks, reprocessing all historic and current raw data with the most up-to-date classifiers, so the skill 'Generative AI Agents' introduced in January 2025 was identified in all historic data. That is not a bug, it is the design, but it means a skill's entire back-series can move without a single employer changing behavior.

The scoring procedure

Run these eight steps in order. Steps one through four are gates that clean the data; steps five through seven produce the evidence; step eight is the verdict. Budget most of your time on cleaning, because that is where the false positives live.

There is a real disagreement about order. LinkedIn anchors persistence on a multi-year CAGR first, computing the share of hiring and compound annual growth rate for each occupation between 2015 and 2019 to identify Emerging Jobs. Practitioner guidance prioritizes rolling continuous monitoring over any single snapshot. I resolve it this way: use the multi-year CAGR to establish that a trend exists, but never let a static multi-year figure stand as your current read. Both belong in the score, at different jobs.

Reading spread and adjacency

Spread and adjacency separate a real capability cluster from a single-team fad, and they are the two signals vendors rarely hand you pre-computed. Compute them yourself from the postings you already scoped.

Spread uses demand share per cluster. Let n_g(s) denote the number of job postings in group g that require skill s and N_g the total postings in g; the demand share f_g(s) = n_g(s)/N_g gives the proportion of postings requesting skill s. Compute that for each occupation cluster. A skill that appears across many clusters at a modest share is broader and more durable than a skill spiking in exactly one.

Adjacency uses co-occurrence. An IZA study applied principal component analysis and found that programming knowledge and reliableness stick out as these skills are often mentioned in isolation. Isolation signals experimentation; embedding in an established cluster signals capability building. Lightcast extracts on average 13 skills per posting, which gives you the denominator for composition: if your skill rarely shares a posting with any established peer, treat the rise with suspicion.

Isolation is the tell for hype: a skill that never shares a posting is a skill no one has learned to use yet.

Refolk removes the friction that adjacency analysis usually carries: instead of scraping profiles and building a co-occurrence matrix by hand, you ask for the people who list both skills and read the overlap directly. That co-listing count is the supply-side denominator the rest of your score needs.

The supply-side counterweight

When demand and supply both rise, use a shortage signal, not demand alone, because rising demand plus faster-rising supply is saturation. The OECD Skills for Jobs shortage and surplus indicators are constructed from five sub-indices - wage growth, employment growth, hours worked growth, unemployment rate, and under-qualification growth - comparing each occupation's time series to the economy-wide trend.

You do not need all five to sanity-check a verdict. Profile-count growth against posting-share growth is enough to catch the saturation trap. Lightcast illustrates the gap directly: almost 20% of employers are looking for physician assistants with experience in radiology, but less than 5% of profiles list it as an acquired skill. That is a genuine shortage. The opposite pattern, profiles climbing faster than postings, is a fading verdict waiting to happen.

Refolk's index supplies the profile-count denominator, and the numbers reshape a trajectory call. Even a hype-candidate skill can be genuinely scarce.

Read that table carefully before you write off a hyped skill. In Refolk's index, Prompt Engineering supply sits at just 0.43x Kubernetes, under half the installed base of an established skill. Scarcity survives hype: a rising-demand verdict can still be defensible on supply grounds even when the skill feels like a bubble, because the people who can actually do it are thin on the ground.

Geography compounds this. The same skill carries wildly different supply by market.

Generative AI profiles are 12.35x more common in the US than Germany. An identical demand trend implies a very different shelf life and sourcing difficulty in each market. A rising verdict in the US might be plateauing on supply; the same trend in Germany stays sharply undersupplied for longer.

How this goes wrong

Every false positive in skill-demand analysis is a signal that survives a lazy look and dies under a specific check. This is the most important section to internalize, because a confident wrong verdict costs more than no verdict. Here are the eight ways the score fails and the check that catches each.

Two of these deserve extra weight because they are invisible without domain knowledge.

Retroactive relabeling is the sneakiest. Because Lightcast reprocesses all historic data every four weeks and Indeed shifted from rule-based keyword matching to model-based extraction, a skill's back-series can rise without any change in employer behavior. Indeed itself warns that its skill taxonomy continuously evolves through new skills, refined categories, and a shift from rule-based keyword matching to model-based extraction, which may contribute to differences in results over time alongside genuine labor market trends. The mechanism is that retroactive relabeling moves the numerator and denominator at different dates. When you see a clean step change, check the date against a known method change before you call it growth.

Snapshot staleness is the most common and the most forgivable, which is why it needs a hard rule. The most common mistake is treating trend analysis as a quarterly snapshot exercise; markets move faster, and a three-month-old picture of skill demand can send your sourcing strategy in the wrong direction. The fix is continuous trend monitoring over static snapshots, tracking both the velocity and direction of change. Bake the revisit into the verdict itself.

Assigning shelf life and keeping the score current

Every verdict ships with an expiry date roughly one quarter out, because no public standard sets a longer safe horizon and the evidence says reads decay fast. A skill-demand score without a shelf life is a liability the moment the market moves.

There is no hard published figure for how long a read stays good, so state that plainly to your finance partner rather than inventing precision. What the evidence supports is a one-quarter cadence: with the Indeed index barely above pre-pandemic and negative year over year, a demand read decays quickly, and the practitioner consensus is that a three-month-old picture is already stale. Set the revisit date, note the velocity, and re-run the gates first.

Here is the rubric to hand a stakeholder verbatim.

Skill: __________  Aliases: __________  Universe (role/geo): __________
Denominator: postings mentioning skill / total postings in scope
Baseline index date: __________ (=100)

GATE 1  Normalization: PASS / FAIL  -  share moves, not just counts
GATE 2  Seasonality:  PASS / FAIL  -  survives X-13 or YoY at like quarters
(If either gate FAILS -> verdict = FADING, stop here.)

TEST 1  Persistence:  PASS / FAIL  -  growth streak: __ quarters | CAGR: __%
TEST 2  Spread:       PASS / FAIL  -  clusters with share > threshold: __
TEST 3  Adjacency:    PASS / FAIL  -  co-occurs with established skill: Y / N
TEST 4  Supply gap:   PASS / FAIL  -  demand share vs profile growth: ____

VERDICT: Rising (3 test passes) / Plateauing (1-2) / Fading (0)
Reindex breaks flagged: __________
SHELF LIFE / revisit date: __________ (~1 quarter)

Before you call the job done, run this final check. Each item is something a finance partner can and will ask about.

Keeping the score current is not re-doing the whole analysis every quarter. It is re-running the two gates and the supply check, watching for a break at any new reclassification date, and asking whether velocity has flipped. When the direction of change reverses for two consecutive reads after adjustment, the grade changes, and so does the plan built on it.