How We Calculate It
The Upper Limit Score measures proximity to a threshold where AI-driven productivity outpaces the purchasing power needed to sustain demand. A score of 0 means the economy is balanced. A score of 100 means we have reached The Upper Limit. This page documents every input, every weight, and every exclusion.
What It Measures
The score tracks the balance between two forces:
AI Productivity
How fast output per worker is growing, driven partly by AI and automation.
Purchasing Power
Whether household compensation and spending capacity keep pace with that growth.
# Score formula
Proximity = 100 - Σ(weight_i × normalized_i)
# Where each component is normalized:
normalized_i = clamp((value - min) / (max - min) × 100, 0, 100)
# Higher raw composite = more balanced. Invert for proximity.
Score Components
Five indicators, each with a documented weight, normalization range, and evidence-based rationale.
Productivity-Wage Gap
35%FRED Series
OPHNFB / COMPRNFB
Normalization Range
0.85 (severe divergence) to 1.0 (parity)
Current Value
~0.92
Direction
Higher ratio = closer to limit
Why this weight?
This is the core measurement of The Upper Limit thesis. When productivity (OPHNFB, output per hour) grows faster than real compensation (COMPRNFB, inflation-adjusted pay per hour), the gap represents economic output that is not translating into worker purchasing power. Bivens & Mishel (2015) document this divergence as the defining structural change in the post-1973 U.S. economy. We assign the highest weight because this single ratio captures the fundamental question the model asks.
Evidence
Bivens & Mishel (2015, EPI) show that net productivity grew 59.7% from 1979-2013 while median compensation grew 8.2%. The ratio of compensation to productivity has declined from near-parity in the late 1960s to approximately 0.92 today.
Real Compensation Growth
20%FRED Series
COMPRNFB (YoY change)
Normalization Range
-1% (contractionary) to +2% (strong)
Current Value
~+0.6%
Direction
Lower growth = closer to limit
Why this weight?
Year-over-year growth in real compensation measures whether workers are gaining purchasing power right now, independent of the cumulative gap. A level metric (the gap) tells you where you are; a flow metric (growth rate) tells you whether things are getting better or worse. We use YoY growth rather than the raw index to avoid base-year bias identified in our audit.
Evidence
Autor et al. (2020, QJE) demonstrate that the labor share decline accelerated in the 2000s, driven by market concentration. Real compensation growth has averaged roughly 0.5-1.5% since 2010, well below the 2%+ rates common before 1973.
Compensation Share of GDP
15%FRED Series
COE / GDP
Normalization Range
49% (concerning) to 53% (healthy)
Current Value
~50.5%
Direction
Lower share = closer to limit
Why this weight?
Compensation of Employees (COE) as a percent of GDP measures how much of total economic output goes to workers versus capital. This is a structural indicator that changes slowly. We weight it at 15% (reduced from an initial 20%) because it partially overlaps with the productivity-wage gap: a widening gap mechanically depresses compensation share over time. We keep it because it captures capital-labor dynamics that the gap ratio alone does not (e.g., rising profits, stock buybacks, and rent extraction).
Evidence
Karabarbounis & Neiman (2014, QJE) show a global decline in labor share driven by falling relative investment prices. Note: COE/GDP understates the canonical BLS labor share measure by 5-8 percentage points because GDP includes taxes on production and subsidies. We use COE/GDP because both series are quarterly and consistently reported.
Labor Force Participation
15%FRED Series
CIVPART
Normalization Range
60% (depressed) to 67% (full engagement)
Current Value
~62.5%
Direction
Lower participation = closer to limit
Why this weight?
LFPR measures whether working-age people are in the labor force at all. Low unemployment can mask structural displacement if discouraged workers drop out of the labor force entirely. This is the scenario The Upper Limit is most concerned about: AI-driven productivity gains that reduce labor demand so substantially that workers exit the market, not just between jobs but permanently.
Evidence
LFPR peaked at 67.3% in 2000 and has not recovered, declining to roughly 62.5%. Acemoglu & Restrepo (2019, JEP) model how automation can reduce the share of labor in the economy even while creating new task categories. The demographic component (aging population) partially explains the decline, but prime-age (25-54) LFPR also fell, suggesting structural causes.
Unemployment Rate (inverted)
15%FRED Series
UNRATE
Normalization Range
3.5% (full employment) to 7% (recessionary)
Current Value
~4.2%
Direction
Higher unemployment = closer to limit
Why this weight?
The U-3 unemployment rate is the most widely watched labor market indicator. We include it as a cyclical complement to the structural indicators above. Current low unemployment (~4.2%) actually pulls the score down (away from the limit), which is correct: low unemployment means labor markets are tight. However, unemployment is a lagging indicator and can change rapidly. We weight it equally with LFPR to balance cyclical and structural labor market signals.
Evidence
The U-3 rate captures active job seekers but misses discouraged workers (counted in LFPR) and underemployment. We normalize against a 3.5-7% range, where 3.5% represents historically full employment and 7% represents recessionary conditions (the approximate average during the 2008-2010 period).
Excluded Components
These metrics are tracked on the Dashboard but deliberately excluded from the score. Each exclusion has a specific reason.
Consumer Spending (% GDP)
PCE / GDPConsumer spending (PCE/GDP) at ~68% is tracked on the dashboard but excluded from the score. It has remained remarkably stable at 65-70% of GDP since 1960, driven partly by government transfer payments and consumer debt. Including it would add stability bias, making the score appear healthier than the underlying income dynamics suggest. The spending that matters is whether it is sustainable, which compensation growth and the wage gap already capture.
Corporate Profit Share
CP / GDPAfter-tax corporate profits as a share of GDP (CP/GDP, ~12%) are tracked on the dashboard. We exclude it from the score because it is roughly the inverse of compensation share: when labor share falls, profit share rises. Including both would double-count the same dynamic. We chose compensation share because it directly measures worker outcomes, which is what The Upper Limit tracks.
Inflation (CPI)
CPIAUCSLCPI year-over-year change is tracked on the dashboard. We exclude it from the score because real compensation (COMPRNFB) is already inflation-adjusted. Including CPI separately would penalize inflation twice. However, inflation does matter for purchasing power perception. If the model evolves to use nominal wages instead of real compensation, CPI would need to be reintroduced.
GDP Growth
GDPGDP growth is tracked on the dashboard but excluded from the score. The Upper Limit thesis is specifically about distribution, not growth. GDP can grow while wages stagnate. Including GDP growth would dilute the distributional signal. As Summers (2014) argues in the secular stagnation framework, aggregate growth metrics can mask fundamental demand-side weaknesses.
AI Adoption Index
N/A (no reliable series)No reliable, high-frequency, standardized measure of AI adoption exists in public data. Proxy measures (Stanford HAI surveys, compute spending, developer adoption) are annual, self-reported, and inconsistently defined. We removed a proxy metric from the dashboard because displaying an unreliable number alongside real FRED data would undermine credibility. Instead, AI milestones are plotted as event markers on all charts, showing when major AI systems launched. When BLS or Census develops standardized AI adoption metrics, we plan to incorporate them.
Wealth Concentration (Top 1% Share)
WFRBST01134The Federal Reserve publishes quarterly top 1% net worth share (WFRBST01134, currently ~31.6%). We plan to add this in a future revision. Rising wealth concentration can mask aggregate health: median purchasing power may be declining even when averages look stable. The challenge is calibration: the series begins in 1989, giving limited historical range for normalization.
Disposable Income Per Capita
A229RX0Real disposable income (A229RX0) is tracked on the dashboard. We exclude it from the score because it includes government transfer payments (Social Security, unemployment insurance, stimulus), which can sustain purchasing power artificially during periods of labor market weakness. The Upper Limit is concerned with whether the economy can sustain demand from earned income, not from transfers. Compensation growth captures the earned income signal.
Known Limitations
The model uses average compensation, not median. Top-earner gains can pull the average up while median workers see stagnation. If median compensation data becomes available at quarterly frequency, it should replace the average.
COE/GDP understates the canonical BLS labor share by 5-8 percentage points because GDP includes taxes on production and subsidies. Users familiar with the standard ~57% labor share figure should note this difference.
The model treats each indicator independently. Cross-indicator correlations (e.g., unemployment often rises as LFPR falls during recessions) are not modeled. This can cause the score to overreact during synchronized downturns.
Normalization ranges are calibrated to post-1960 U.S. data. Unprecedented structural changes (AI-driven mass displacement, UBI implementation) could push values outside these ranges.
Historical scores are retroactively computed. If the formula or weights change, all historical scores change. The score for 2010 was not published in 2010; it was calculated today using today's methodology applied to 2010 data.
Data revisions from BLS and BEA can retroactively change historical values. Preliminary estimates are often revised 1-3 months later.
Design Decisions
Key choices, with rationale. Open to revision as the model evolves.
Why is the Productivity-Wage Gap weighted highest?
It directly measures the central thesis. Every other component is either a cause of or a consequence of this gap. Bivens & Mishel (2015) show this divergence is the defining structural change of the modern U.S. economy. If you could only track one number, this would be it.
Should the model distinguish AI-driven productivity from cyclical productivity gains?
The Dashboard plots major AI milestones (Transformer 2017, GPT-3 2020, ChatGPT 2022, GPT-4/Claude/Gemini 2023) on the Productivity vs. Wages chart for visual inspection. The score itself does not yet separate AI-driven from cyclical productivity because no reliable decomposition exists in public data. As BLS develops AI-specific productivity measures, we plan to incorporate them.
Why use COE/GDP instead of the canonical BLS labor share?
The BLS labor share series (PRS85006173) is an index (2017=100), not a percentage. Converting it back to a meaningful percentage requires assumptions about the base level. COE/GDP gives a clean, interpretable percentage from two well-understood quarterly series. The tradeoff: it understates labor share by 5-8pp because GDP includes taxes and subsidies.
Why invert the score (higher = worse)?
The site asks 'How close are we to The Upper Limit?' A score that directly answers that question (44 = 44% of the way there) is more intuitive than a health score that requires mental inversion. It also makes trends clearer: a rising line on the score history chart means things are getting worse, which matches natural reading.
Academic Grounding
The papers that inform the indicators, weights, and conceptual model.
Understanding the Historic Divergence Between Productivity and a Typical Worker's Pay
Bivens, J. & Mishel, L. (2015). Economic Policy Institute
Primary evidence for the productivity-wage gap component (35% weight). Documents the post-1973 wedge and decomposes its causes.
The Global Decline of the Labor Share
Karabarbounis, L. & Neiman, B. (2014). Quarterly Journal of Economics, 129(1), 61-103
Foundational for the compensation share component (15% weight). Attributes global labor share decline to falling relative investment prices.
The Fall of the Labor Share and the Rise of Superstar Firms
Autor, D., Dorn, D., Katz, L. F., Patterson, C. & Van Reenen, J. (2020). Quarterly Journal of Economics, 135(2), 645-709
Shows market concentration drives labor share decline. Supports the structural (not cyclical) interpretation of falling compensation share.
Automation and New Tasks: How Technology Displaces and Reinstates Labor
Acemoglu, D. & Restrepo, P. (2019). Journal of Economic Perspectives, 33(2), 3-30
Models automation displacement and task creation. Informs LFPR and unemployment components. Supports the premise that AI could reduce labor demand structurally.
U.S. Economic Prospects: Secular Stagnation, Hysteresis, and the Zero Lower Bound
Summers, L. H. (2014). Business Economics, 49(2), 65-73
The secular stagnation hypothesis. The Upper Limit concept echoes this framework: structural demand shortfalls constrain growth. Informs the exclusion of GDP growth from the score.
Imperialism: A Study
Hobson, J. A. (1902). James Pott & Co.
Early underconsumption theory. The intellectual ancestor of The Upper Limit's core premise: insufficient purchasing power constrains aggregate demand.
Critique This Methodology
If you see a flaw in the model, a missing variable, or a better weighting approach, we want to hear it. The Upper Limit is designed to evolve.
Explore the Data