Skip to content
← Back to CareerStar

How scores are calculated

CareerStar sizes up a career the way an investor sizes up a stock — what’s the payoff, what’s the risk — and blends the result with how well the path fits you. Everything below is plain arithmetic: the same inputs always produce the same score, and no AI ever computes a number. This page walks through all of it.

What is this page & how do I use it?

This page is the math behind every CareerStar score — the full model, its weights, the data sources, and the tests it was put through, written so you can check the work rather than take my word for it.

How to use it: skim the section headings and stop where you’re skeptical. If you only read one section, read the back-test — it shows how the model fared against a real decade, including what it got wrong.

The model

Each occupation gets a 0–100 score built from five measured ingredients. Every ingredient is converted to a number between 0 and 1 before any arithmetic happens — 0 meaning the bottom of the range, 1 the top — so that things measured in different units (percent growth, dollars of pay) can be combined at all. The five, and where each number comes from:

  • growth — the U.S. Bureau of Labor Statistics’ projected employment growth for this occupation over 2024–2034, expressed as a percentile rank against all 730 occupations in the set. A growth of 0.92 means “projected to grow faster than 92% of all careers” — not 92% growth.
  • pay — the occupation’s median annual pay (also BLS), as the same kind of percentile rank. 0.50 means the pay is exactly in the middle of all careers.
  • exposure — the estimated share of the job’s tasks that AI could meaningfully perform or speed up, taken from a published research dataset (Eloundou et al. 2023 — see data sources). It already arrives as a 0–1 number: 0.8 means roughly 80% of the job’s tasks are AI-touchable. (Touchable, not doomed — a task AI can help with is not a job that disappears.)
  • volatility — a stand-in for how unstable the field is. No public dataset measures this directly, so it’s built from the growth number: volatility = 1 − growth. A field at the 10th percentile for growth (i.e. shrinking relative to others) gets volatility 0.90 — a shrinking field is treated as a riskier place to stand.
  • Fit — how well the occupation’s real skill profile matches the interests you typed in, from 0 (no overlap) to 1 (strong match), measured in O*NET capability-space (explained in the next section).

Those five ingredients feed four lines of arithmetic:

Return = wGrowth·growth + wPay·pay
Risk   = wExposure·exposure + wVolatility·volatility
RAV    = Return · (1 − γ·Risk)          (risk-adjusted return)
Score  = 100 · [ α·RAV + (1 − α)·Fit ]

Everything else in those lines — the w’s, the Greek letters — is a weight: a fixed number that says how much an ingredient counts. They look mysterious, but each is just a name for one number, and here they all are, straight from the code:

  • wGrowth = 0.5, wPay = 0.5 — Return is a plain 50/50 average of the growth rank and the pay rank. Neither is favored.
  • wExposure = 0.7, wVolatility = 0.3 — Risk is 70% AI exposure, 30% volatility. Why exposure dominates is argued below.
  • γ (“gamma”) = 0.6 — how hard risk is allowed to punish return. In line three, a career’s Return gets multiplied by (1 − 0.6·Risk): at zero risk nothing is lost, and even at the maximum possible risk (Risk = 1) the career keeps 40% of its Return. RAV — risk-adjusted return — is just the name for Return after this discount.
  • α (“alpha”) = 0.7 — the final blend: the score is 70% market reality (RAV) and 30% personal fit. Since RAV and Fit are both 0–1 numbers, multiplying the blend by 100 is what makes the familiar 0–100 score.

A worked example, start to finish. Take a career that ranks at the 90th percentile for growth and the 70th for pay, has AI exposure 0.50, and matches your interests at Fit = 0.60:

Return     = 0.5·0.90 + 0.5·0.70        = 0.80
volatility = 1 − 0.90                   = 0.10
Risk       = 0.7·0.50 + 0.3·0.10        = 0.38
RAV        = 0.80 · (1 − 0.6·0.38)      = 0.62
Score      = 100 · (0.7·0.62 + 0.3·0.60) ≈ 61

Four short lines, but each one is a decision, and each had an obvious alternative I turned down:

  • Percentile ranks, not raw units. Growth is a percent and pay is dollars, so you can’t average them directly. Instead, each career is ranked against all 730 others — “92nd percentile for pay” — which also means a score never says a career is good in the abstract. It says how the career stacks up against everything else you could pick, which is the decision you’re actually making. The obvious alternative — stretching raw values onto a ruler running from the smallest to the largest — would let one extreme career (surgeons’ pay) squash every other career into the bottom of the scale.
  • Risk discounts return; it isn’t subtracted. RAV multiplies: Return · (1 − γ·Risk). Subtracting risk would let a safe dead-end career outrank a risky booming one purely on safety. Multiplying makes risk matter in proportion to what’s at stake — the same logic as a risk-adjusted return in investing. And γ = 0.6 caps the damage: even maximum risk erases at most 60% of a career’s return, never all of it.
  • Exposure dominates risk (0.7 vs 0.3) because AI disruption is the question this site exists to ask — and because the volatility term is the weaker measurement, a constructed proxy (see limitations). The weaker measurement gets the minority weight.
  • Market outweighs fit (α = 0.7) because this is a viability rating that gets personalized, not a personality quiz with salary data sprinkled on. Reality carries the score; your interests tilt it.

Now the honest part: nothing proves these particular weights are correct. There is no answer key of “right career choices” to tune them against, so 0.7 and 0.6 are judgment calls — written down in the code for anyone to read, not hidden. That’s why so much of this page is stress-testing: every comparison is re-run 729 times with the weights nudged in every direction, five rival formulas score it under different philosophies entirely, and the model was graded against a real decade of history. I can’t prove my weights are right, so I show you what happens when they’re wrong.

How Fit is measured

The U.S. Department of Labor’s O*NET database rates every occupation on the same checklist of 68 abilities: 35 skills (Critical Thinking, Programming, Mathematics…) and 33 knowledge areas (Economics & Accounting, Engineering, Medicine, Law…), each scored by how important it is to that job. So every career comes with a 68-number profile of what the work actually demands.

When you type in your interests, CareerStar translates them onto that same 68-item checklist using a fixed, published word list (“coding” maps to Programming, “helping people” maps to Social Perceptiveness, and so on). Fit is then simply how much your profile and the career’s profile overlap.

One subtlety matters. Nearly every professional job rates high on the common abilities — reading, speaking, critical thinking — so raw overlap would make every office job look like a match for everyone. To fix that, CareerStar counts a match on a rare ability (one few other jobs demand, like Medicine or Programming) far more heavily than a match on one every job shares. That’s what lets fit tell finance from engineering, not just desk-work from the trades. It’s still a modeling choice, and I say so — two similar fields can both score high.

The AI-moat rating

A moat is the water around a castle — investors (Morningstar, most famously) use the word for how well a company can defend its position. CareerStar borrows it for careers: how defensible is this work against AI? The rating comes from a fixed formula, not a judgment call, built from two questions:

  • How much of the work can AI touch? (the same 0–1 exposure number from the score — low exposure is shelter)
  • Are the career’s defining abilities rare? A job built on abilities few other jobs demand (surgery, say) is harder to substitute than one built on abilities every job shares.
defensibility = 0.6·(shelter from AI) + 0.4·(rarity of its top abilities)

Wide moat   = top ~20% of careers on defensibility
Narrow moat = middle ~50%
No moat     = bottom ~30%

The cutoff lines between wide, narrow, and none were set once, on the full 730-career distribution, and then frozen — they don’t move to make any career look better. The model makes uncomfortable calls: it gives software developers no moat (AI can already touch most of their tasks) even though their overall score stays high — and I ship those calls rather than tune them away. For 25 newer occupation codes O*NET hasn’t rated yet (software developers among them), the ability profile is estimated from closely related occupations and flagged as an estimate in the data.

Score uncertainty is also stated in words on every card — Low (±5 or less), Medium (±6–8), High (±9–10), Very high (±11+) — wider with high AI exposure or weak fit.

What the rated universe looks like

Here is how all 730 occupations actually spread out under the default settings — how many land in each score range, and how many get each moat rating. (Star ratings are graded on a curve on top of these scores, the way a professor curves grades: the top slice of careers gets five stars regardless of the raw number.)

Score distribution (0–100, bins of 10)
AI-moat split
wide 152 (21%)narrow 358 (49%)none 220 (30%)

Robustness (does the answer survive?)

The weights above are a deliberate choice, so the fair test is whether the answer survives when you disagree with them. For every comparison, CareerStar re-runs the whole calculation 729 times: each of the six weights is tried at three settings — as chosen, 20% higher, and 20% lower — and 3 × 3 × 3 × 3 × 3 × 3 makes 729 combinations. Then it reports how often each career keeps its place in the ranking. A result that holds across all 729 runs is solid; one that shuffles is flagged as a close call rather than sold as a verdict. You can watch this live by moving the sliders on the results screen.

Five rival models (model risk, made visible)

Any single formula is one opinion about how much AI risk should count. So every comparison is also scored under four rival models — different philosophies, not weight tweaks — and the app shows when they agree (conviction) and when they crown different winners (genuine uncertainty):

ModelPhilosophyFormula
StandardThe headline model: reward (growth + pay), marked down by AI risk, blended with your fit.100·[α·(Return·(1−γ·Risk)) + (1−α)·Fit]
Growth optimistGrowth + pay only. Deliberately ignores AI risk, betting the official forecasts already account for it.0.5·growth + 0.5·pay
DefensiveSafety first: if AI hits hard, surviving beats thriving. Shelter from AI counts most; reward is secondary.0.5·moat + 0.3·resilience + 0.2·pay
EfficiencyReward per unit of risk — a career must earn its risk to score high. (Investors call this a Sharpe ratio.)100·Return / (Return + Risk)
Plain averageThe baseline: a simple average of everything. If a clever model can't beat this, that's worth knowing.(growth + pay + resilience + fit) / 4

One of the five is deliberately dumb: it just averages everything, equally. It’s there as a baseline — if a fancier model can’t beat a plain average, all its extra machinery is doing nothing.

Data sources

  • Growth & pay — U.S. Bureau of Labor Statistics, Employment Projections 2024–2034 (public domain) — ~730 occupations across every field.
  • AI exposure — Eloundou et al. 2023, “GPTs are GPTs” (occupation-level exposure, β measure; MIT-licensed).
  • Skills & knowledge — O*NET 29.0 Database (U.S. DOL/ETA, CC BY 4.0), Skills and Knowledge importance ratings.
  • Occupations are keyed by O*NET-SOC code.

Does “AI exposure” just mean “decline”?

A fair objection: maybe the AI-exposure number is redundant — maybe the exposed jobs are simply the shrinking ones, and growth already tells the whole story. The data says no. Across all 730 occupations, knowing a job’s AI exposure tells you almost nothing about whether it’s projected to grow or shrink — the two are nearly unrelated. (Statistically: the correlation is 0.075on a scale where 0 means “no relationship at all” and 1 means “perfectly linked.”) Split all jobs into four buckets from least- to most-exposed, and average projected growth is about the same in every bucket — 0%, 1.6%, 3.6%, 1.3% from least to most exposed — no downward slide.

So AI exposure captures something the growth forecast doesn’t. That’s why the score adjusts for it instead of just ranking careers by growth — and why “exposed to AI” really doesn’t mean “doomed to shrink” in this data. (This isn’t left as an assertion, either — the model was run against a real decade of history. See the next section.)

The back-test: 2014 → 2024, a real decade

A back-test is how you check a model without waiting ten years: rewind the clock, let the model make its calls using only what was knowable then, and grade it against what actually happened. So I dug the government’s official 2014 career forecasts out of an internet archive, had the model score the 2014 labor market using only that 2014 information, and then compared its scores against what actually happened by 2024 — 647 occupations tracked across the decade.

0.39
link between 2014 score and what happened (0 = none, 1 = perfect)
48%
of the careers that shrank were flagged (guessing gets 33%)
37 vs 46
typical 2014 score of careers that shrank vs ones that grew
0.099
AI exposure’s link to that decade (expected ≈ 0 — it was pre-ChatGPT)

Read it honestly. Careers the model scored higher in 2014 really did fare better over the decade — the link between score and outcome was 0.39 on that 0-to-1 scale, solidly above zero but far from prophecy. And it flagged 48% of the careers that went on to shrink, roughly one and a half times better than guessing. One wrinkle I’m not going to bury: the government’s raw growth forecast alone did slightly better (0.411) than the full model. Why? Because the model’s AI-risk adjustment did nothing for that particular decade — 2014–24 ended before ChatGPT changed anything, so penalizing AI-exposed careers couldn’t help yet. That’s the model working as designed: AI exposure is a bet about the future, and a test run on the past can’t score it. The back-test proves out the part of the model history can test, and is upfront about the part it can’t.

The misses, named: the model liked oil-and-gas roles in 2014 (rotary drill operators, oil and gas, score 64) that fell 52% when oil prices crashed — the kind of economy-wide shock no career-by-career model can see coming. And it under-rated couriers (+166% on the e-commerce boom). A few of the wildest swings are bookkeeping rather than reality — the government occasionally renames or regroups occupations between decades, which can make a career look like it exploded or vanished. Full numbers, misses, and false alarms: data/backtest.json.

Check my math — download the data

Every rating on this site is in one open CSV: careerstar-ratings.csv — all 730 occupations with the default-weight score, stars, percentile, AI-moat, all five model scores, consensus, growth, pay, exposure, and a flag for the 25 careers whose capability vectors are estimated. If a number looks wrong, the formulas are on this page and the code is on GitHub — auditing it is encouraged. Derived ratings are CC BY 4.0 (cite CareerStar); upstream data keeps its own licenses (BLS public domain, O*NET CC BY 4.0, Eloundou MIT).

Limitations (read these)

  • Exposure is not job loss. AI exposure measures the share of tasks a model could touch — not whether the job disappears.
  • It’s an estimate, not a prediction. No one can see the future of the labor market; this is a grounded, transparent snapshot.
  • Fit is a model, not a verdict.It uses the government’s real skill data, but the word list that translates your interests onto that skill checklist was written by hand, so fit is a defensible estimate — and two similar fields (say finance and engineering) can score alike.
  • The volatility number is a stand-in. Nobody publishes a direct measure of how unstable each career is, so this model uses the best available substitute: treating shrinking fields as less stable ones.

Built with an AI-native workflow. The language model only writes the plain-English explanations — it never computes a score.