Methodology

See how the ProofIndex score is calculated, including every formula, weight, and data source.

How ProofIndex works

ProofIndex measures how exposed a local real estate market is to AI-driven income disruption. Each area gets two scores from 0 to 100: AMVS for mortgage exposure and AMVS-R for rental exposure.

Both scores use four pillars that cover different parts of the risk. We rank pillar scores within each country, so a score of 70 means "more exposed than 70% of areas in this country."

Composite Formula

AMVS = (0.35 × EVI) + (0.25 × ICR) + (0.25 × HLS) + (0.15 × (100 − RF))

AMVS-R = (0.35 × EVI) + (0.25 × ICR) + (0.25 × RME) + (0.15 × (100 − RF))

Pillar 1: Employment Vulnerability Index (EVI)

Weight: 35%

Measures the share of local jobs that are exposed to AI automation, weighted by each occupation's AI Exposure Coefficient (AEC).

AI-Exposed Workforce Share · 40%
The weighted sum of (employment × AEC score) / total employment. Occupations with higher AEC scores contribute more than their headcount alone.
White-Collar Concentration · 25%
Percentage of the local workforce in white-collar occupations. Higher concentration = more jobs susceptible to AI cognitive automation.
Sector Concentration (HHI) · 15%
Herfindahl-Hirschman Index of AI-exposed sectors. A local economy that depends on fewer AI-vulnerable industries is more fragile.
Top Employer Risk · 10%
Whether the area's largest employers operate in AI-exposed sectors. A single large employer in tech or finance creates outsized risk.
Remote Work Prevalence · 10%
Percentage of workers primarily remote/hybrid. Higher remote work = more jobs that could be performed by AI from anywhere.

Pillar 2: Income Concentration Risk (ICR)

Weight: 25%

Measures how much of the area's total income depends on AI-vulnerable sectors and occupations.

AI-Sector Payroll Share · 35%
Percentage of total local payroll from AI-exposed sectors. Payroll captures the value of local income, so a small group of highly paid tech workers can account for a large share of local spending power.
Top-Decile Income Share · 25%
Share of total income held by the top 10% of earners. In tech-heavy areas, losing some of this AI-exposed income can also reduce spending across the local economy.
Median Income Ratio · 20%
Local median household income divided by national median. Higher ratios indicate an economy built on high-paying (often AI-vulnerable) jobs.
Dual-Income AI Households · 20%
Estimated percentage of households where both earners work in AI-exposed occupations. Both incomes could be disrupted at the same time.

Pillar 3: Housing Leverage Score (HLS)

Weight: 25% · AMVS (Mortgage)

Measures how exposed local homeowners and the mortgage market are to income disruption.

Price-to-Income Ratio · 25%
Median home price divided by median household income. Higher ratios mean homeowners are more stretched and vulnerable to income loss.
Mortgage Debt-to-Income · 25%
Average mortgage payment as a percentage of gross household income. Higher DTI = less buffer to absorb income shocks.
3-Year Price Appreciation · 20%
Cumulative home price growth over the trailing 3 years. Recent rapid appreciation suggests prices may be disconnected from fundamentals.
High-LTV Share · 15%
Percentage of mortgages with a loan-to-value ratio above 80%. These homeowners have little equity to absorb a price decline.
Variable Rate Exposure · 15%
Percentage of mortgages on adjustable/variable rates. Especially relevant in the Czech Republic where most mortgages have 3-5 year fixation periods.

Pillar 3R: Rental Market Exposure (RME)

Weight: 25% · AMVS-R (Rental)

Measures how vulnerable the local rental market is to AI-driven tenant income loss.

Rent-to-Income Ratio · 20%
Median rent as a percentage of median renter household income. Higher ratios mean tenants have less buffer to absorb income disruption.
Tenant Occupation Profile · 25%
Percentage of local renters employed in AI-exposed occupations (Tier 1-2). This separates the rental score from the mortgage score because renters tend to be younger and more concentrated in tech work.
Yield Compression Risk · 15%
How close current rental yields are to historical lows. Compressed yields leave landlords with less room to absorb vacancy or rent reductions.
Vacancy Trends · 15%
Current vacancy rate weighted by its direction. A tighter market can be more exposed if demand falls suddenly.
Lease Repricing Velocity · 10%
How quickly rental rates adjust to market changes. Faster repricing means landlord income drops more quickly when demand falls.
Expat Dependency · 15%
Estimated share of rental demand from international workers, digital nomads, and corporate relocations. Especially relevant in Prague where expat demand drives a significant portion of the rental market.

Pillar 4: Resilience Factors (RF)

Weight: 15% (inverted)

Measures the area's ability to absorb economic shocks. Unlike other pillars, higher resilience reduces the overall score. The composite formula uses (100 - RF).

Economic Diversification · 30%
Inverse of the broad Herfindahl-Hirschman Index across all sectors. A diversified economy is less dependent on any single industry and better able to absorb sector-specific shocks.
Safety Net Strength · 25%
Unemployment benefit replacement rate multiplied by benefit duration. Stronger safety nets give displaced workers more time to retrain or find new employment.
Savings Proxy · 20%
Estimated household savings rate or liquid assets proxy. Areas with higher savings have more buffer to weather income disruption.
Blue-Collar Absorption · 15%
Ratio of blue-collar job openings to potential displaced workers. More blue-collar alternatives mean displaced white-collar workers have fallback options.
Retraining Infrastructure · 10%
Density of retraining programs, bootcamps, and vocational training per 100,000 population. More retraining capacity = faster workforce adaptation.

AI Exposure Coefficient (AEC)

Every occupation in the ProofIndex taxonomy receives an AI Exposure Coefficient (AEC) from 0.00 to 1.00, measuring how susceptible it is to AI automation.

The AEC is calculated from five weighted factors. The AI capability factor uses a dual-exposure model: it blends exposure to today's AI tools (35%) with exposure once complementary software is built on them (65%). Both come from the human task ratings in Eloundou et al. (2023), "GPTs are GPTs". They rate which tasks AI could speed up. They don't measure actual AI use, and they don't set a timeline. The other factors come from the O*NET 29.0 database.

Task Digitizability · 30%
Can the core tasks be performed entirely in a digital environment? 1.0 = fully digital, 0.0 = fully physical.
Routine Cognitive Work · 25%
How rule-based and predictable is the cognitive work? 1.0 = highly routine, 0.0 = requires constant novel judgment.
Output Measurability · 20%
Can output quality be objectively measured? 1.0 = easily measured (code that compiles), 0.0 = subjective (therapy outcomes).
AI Capability (Dual Exposure) · 15%
Blends two ratings from Eloundou et al. (2023): Current AI Exposure (the share of a role's tasks that today's AI tools could speed up directly) at 35% weight, and AI Potential (the same rating once complementary software is built on those tools) at 65% weight. Neither measures actual AI use or sets a timeline.
Relationship Dependency · 10% (inverse)
Does the role require deep human trust or physical presence? Higher values are protective. For example, a therapist (0.9) is harder to automate than a data entry clerk (0.1).
Critical · 0.85–1.00
Nearly all core tasks are exposed to AI tools
High · 0.65–0.84
Many core tasks are exposed to AI tools
Moderate · 0.40–0.64
Some core tasks are exposed, others are not
Low · 0.15–0.39
Few core tasks are exposed to AI tools
Minimal · 0.00–0.14
Core tasks resist AI automation

Normalization and risk tiers

All indicator values are ranked from 0 to 100 within each country. This means:

• Scores are relative. A 70 in the US is not the same absolute level as a 70 in Czechia
• Extreme values, such as San Francisco's, do not compress the rest of the distribution
• A score can be read as "more exposed than X% of areas in this country"
• Compare areas within the same country unless an absolute cross-country calibration is added

AI-Resilient · 0–20
Minimal exposure to AI-driven disruption
Low Risk · 21–40
Below-average exposure with resilience buffers
Moderate Risk · 41–60
Meaningful share of AI-vulnerable income
High Risk · 61–80
Significant AI-vulnerable job concentration
Critical · 81–100
Heavy dependence on AI-vulnerable occupations

What the score can and cannot tell you

ProofIndex is a relative AI-exposure screening index. It ranks areas within one country using workforce task exposure, income concentration, mortgage or rental leverage, and local resilience. The ranking shows relative exposure within that country.

The score does not forecast market outcomes. We tested it against the only later-period outcomes available to us: an archived 2026-Q1 score against the following quarter of FHFA house-price change across 424 US areas, and 13 reconstructed monthly Czech scores against district unemployment six months later across 72 districts. The score did not rank near-term deterioration better than a simple momentum or level baseline in either test. After accounting for those baselines, its added contribution was approximately zero. These short windows cannot settle a slow structural question, but they do not support a predictive claim. ProofIndex does not predict house prices, rents, vacancy, or mortgage defaults.

Keep two points in mind when reading a score. First, scores are ranked within a country. A Czech score of 70 and a US score of 70 sit in different distributions and are not directly comparable. Second, tiers can change. Between the archived 2026-Q1 and 2026-Q2 scores, 17% of areas changed mortgage risk tier. The average move was 3.4 points and the largest was 24.2 points. The rental score was much steadier, with 2.6% of areas changing tier. About a quarter of all areas sit within two points of a tier boundary, so an update can relabel an area even when its position barely moves. Read the score together with its drivers, not the tier alone.

Relative ranking within one country · Claim
The score says an area is more or less exposed than its country peers, and which pillars drive that. This is the strongest supported reading.
Not a market forecast · Limit
No claim is made about future house prices, rents, vacancy, delinquency, or defaults, at any horizon. Out-of-time testing found no edge over level and momentum baselines.
Tiers are labels, not probabilities · Limit
Tier boundaries cut a within-country percentile distribution, so every country averages around 50 by construction. A tier is a position, not a calibrated risk level, and 17% of areas changed tier in the one refresh we have archived.
What would change this · Trigger
Four or more archived score vintages, a genuine outcome series for the channel being claimed (delinquency or foreclosure for the mortgage score, rent or vacancy for the rental score), and a stable out-of-time edge over baselines in more than one window. Until all three hold, the claim stays where it is.

Validation, confidence, and limits

ProofIndex is built for market screening, monitoring, and underwriting discussions. We check whether rankings make sense, stay internally consistent, match stress-sensitive market traits, and clearly reflect the quality of their source data. Where suitable historical data exists, we also compare the scores with later housing-market outcomes. The score does not claim to cause or forecast changes in house prices, rents, vacancy, or defaults.

For each supported country, we compare the highest- and lowest-risk quartiles on affordability, mortgage burden, rental burden, and housing or rental exposure. We read resilience as a separate axis. High resilience does not cancel high exposure, but it may help an area absorb or delay a labor-market shock. Confidence reflects source quality, geographic coverage, recency, and whether a pillar uses official data, market data, estimates, or proxies. Our first later-period tests used US house prices and Czech district unemployment. They did not beat simple baselines, which is why we limit the claims above. Tests against rents, vacancy, and credit outcomes will require local outcome data that we do not yet have.

Face-validity ranking checks · Current
Known high-exposure technology, finance, and capital-city markets should rank above diversified or lower-income-concentration peers when the source data supports it.
High-risk vs low-risk benchmark checks · Current
Within each country, the top-risk quartile is compared with the low-risk quartile on stress-sensitive affordability, leverage, rent-burden, and rental exposure traits. RF is reported as offsetting capacity, not as a pass/fail risk trait.
Exposure versus resilience interpretation · Current
Markets are read on two axes: exposure pressure and resilience buffer. High-exposure, high-resilience markets can be prime but shock-sensitive; high-exposure, low-resilience markets are more fragile.
Country-relative normalization checks · Current
Scores are reviewed within each country so domestic outliers do not distort cross-country interpretation. A Czech score and US score are not treated as directly comparable absolute values.
Mortgage/rental consistency checks · Current
AMVS and AMVS-R should share employment, income, and resilience signals while diverging only where housing leverage or rental-market exposure justifies it.
Scenario monotonicity checks · Current
Slow, moderate, and fast adoption paths are checked so faster AI adoption does not accidentally reduce exposure unless a specific resilience mechanism explains the move.
Historical outcome evidence · Current
The US validation workflow compares current score cohorts with real FHFA house-price history. This is evidence for model review, not proof of causal forecasting.
Confidence tiers from source quality · Current
High confidence requires strong direct coverage. Medium and low confidence flag markets where derived indicators, regional interpolation, or proxy data carry more of the score.

Country source summaries

Czechia · Mixed 2024–Apr 2026
ČSÚ, MPSV, ČNB, Valuo and derived ProofIndex indicators · Regions and districts
Netherlands · Mixed 2024–2025
CBS and derived COROP-level indicators · COROP areas
Finland · Mixed 2023–2025
StatFin and derived regional housing/labor indicators · Sub-regional areas where available; municipality housing prices are aggregated where coverage is sufficient
United States · ACS 2020–2024, housing/labor through 2026
ACS/Census, Zillow/FHFA, BLS/FHFA and market indicators · States and supported MSAs

Data Sources

ProofIndex is built on official government statistics, real estate market data, and peer-reviewed research. All data sources are publicly available, but their vintages are not identical. Each score period uses the latest available official labor, income, population, and housing releases where available, and keeps existing market or derived proxies where official local coverage is incomplete.

For the Czech Republic, the pipeline fetches live unemployment data from the MPSV portal API and uses real per-district property prices from Valuo.cz. Wage and employment indicators are interpolated from 14 region-level data points using a dampened district factor based on real per-district unemployment and labor market tightness from MPSV. Property prices use direct per-district market data rather than interpolation. Rental vacancy pressure uses MPSV job-vacancy tightness as a derived local signal, not direct rental vacancy microdata.

For the United States, Employment Vulnerability (EVI) is modeled from ACS sector and education patterns rather than direct BLS OEWS occupation employment. EVI contributes 35% of the composite, so this component should be treated as directional. For the United States, Netherlands, and Finland, rental vacancy pressure uses country-specific labor and housing-market tightness signals already present in the pipeline: BLS LAUS plus FHFA/Zillow momentum for the US, CBS unemployment-benefit/job-vacancy/tenure signals for the Netherlands, and StatFin-backed housing price, income, tenure, and population-scale signals for Finland. These improve local AMVS-R differentiation but remain derived proxies, not direct residential rental vacancy series.

Known limitation: ProofIndex is a screening and monitoring signal, not a price forecast. It should be combined with local underwriting, rent rolls, financing structure, and supply pipeline analysis before making investment decisions.

Czech Republic — Employment & Economy

Unemployment by region · Apr 2026
MPSV — Ministry of Labour and Social Affairs · 14 regions
Monthly unemployment by district · Apr 2026
MPSV — Portal API (live) · 75 districts · data.mpsv.cz/portal
Job vacancies by district · Apr 2026
MPSV — Portal API (live) · 75 districts · data.mpsv.cz/portal
Wages by region · Q4 2025
ČSÚ — Czech Statistical Office · 14 regions · czso.cz
Employment by occupation (ISCO) · 2024
ČSÚ — Labour Force Survey (VŠPS) · 14 regions
Wage multipliers by occupation · 2024
ISPV — Earnings Information System · National · ispv.cz
GDP & regional accounts · 2024
ČSÚ — Regional Accounts · 14 regions
Population estimates · 2024
ČSÚ — Czech Statistical Office · 77 districts

Czech Republic — Housing & Mortgage Market

Apartment prices per district · Mar 2025 – Feb 2026
Valuo.cz (listing data) · 77 districts · www.valuo.cz/statistiky/zebricek-okresu
Rental estimates · 2025–2026
Derived from property prices via yield model · 77 districts
Mortgage rates & LTV distribution · 2024
ČNB — Czech National Bank (ARAD) · National/Regional · cnb.cz/arad/
Financial stability indicators · 2024
ČNB — Financial Stability Reports · National

Czech Republic — Resilience & Safety Net

Unemployment benefit parameters · Current
Czech labour law (Act No. 435/2004 Coll.) · National (statutory)
Retraining programs · 2024
MPSV — Labour Offices · Regional

Czech Republic — Geographic Boundaries

Administrative boundary polygons · 2024
CzechInvest / web-data · 14 regions + 77 districts · github.com/czechinvest/web-data

Occupation data used in scoring

O*NET Database (work activities and context) · 29.0
U.S. Department of Labor, Employment and Training Administration · SOC occupations (US) · www.onetcenter.org/database.html
LLM exposure ratings ("GPTs are GPTs") · 2023
Eloundou et al. (2023) — OpenAI / University of Pennsylvania · SOC occupations (US) · github.com/openai/GPTs-are-GPTs
SOC ↔ ISCO-08 occupation crosswalk · 2026
ProofIndex (manual curation) · Every occupation in the taxonomy

Background research (not a scoring input)

Labor market impacts of AI: observed vs theoretical exposure · Mar 2026
Anthropic (Massenkoff & McCrory, 2026) · SOC occupations (US) · www.anthropic.com/research/labor-market-impacts
Generative AI and Jobs: Refined Global Index of Occupational Exposure · May 2025
ILO — International Labour Organization · ISCO-08 (global) · www.ilo.org/publications/generative-ai-and-jobs-refined-global-index-occupational-exposure
New Jobs Creation in the AI Age · 2026
IMF Staff Discussion Note · Cross-national · www.imf.org/-/media/files/publications/sdn/2026/english/sdnea2026001.pdf
AI is simultaneously aiding and replacing workers · Feb 2026
Federal Reserve Bank of Dallas · US occupations · www.dallasfed.org/research/economics/2026/0224
Measuring workers' capacity to adapt to AI-driven job displacement · 2025
Brookings Institution · US occupations · www.brookings.edu/articles/measuring-us-workers-capacity-to-adapt-to-ai-driven-job-displacement/
AI exposure indices by occupation · 2023
Felten et al. (2023) — OECD · Cross-national

Source credits and licences

ProofIndex scores, rankings and derived indicators are our own calculations, not official statistics. The organizations below supplied source data. None of them has reviewed, approved or endorsed ProofIndex.

Methodology version 2026.2: Scores are updated quarterly as new employment, income, and housing data become available. The occupation taxonomy uses a dual-exposure model built from O*NET data and the task ratings in Eloundou et al. (2023).

AMVS, AMVS-R and AEC are ProofIndex's own methodology names, not industry standards.

See the evidence behind the model Browse the Signals feed