Policy Risk Index · Methodology · PRI-V1
How the Index Is Built
An index you can't audit is a marketing number. This page is the audit surface: the components, the weights, the transforms, the error bars, and the rules that keep anyone — including us — from quietly changing the number.
What It Measures — and What It Doesn't
The index is a barometer, not a forecast. Higher readings mean current law is under more pressure to change in ways that cost households money — benefit cuts, premium growth, tax resets. It does not predict which change comes, or when. It measures how much pressure is on the system, the way a barometer measures the weather without naming the storm.
It is a national gauge. The personalized PRIA Score is the opposite — computed for your household, where higher means more resilient. One reads the climate; the other reads your house.
Five Components, Weighted by What Shapes Decades
The index is a weighted composite of five components, each scored 0–1 from live data. The weights are deliberate priors, not curve-fit coefficients — ranked by how much each force shapes multi-decade household outcomes, and we say so rather than pretending otherwise. A sensitivity check is part of the build: shifting any weight by ±5 points does not change the story the index tells at any point in its history.
Debt-to-GDP, federal interest burden, and the Treasury yield curve.
Debt service compounds over decades and constrains every other policy lever — the dominant structural risk once past a debt-service threshold.
Live congressional bill activity, scored per bill across inflation, fiscal, and stagnation pressure.
Legislation is the mechanism through which risk becomes policy that touches households. Deficit reduction counts as relief — the scale is signed.
Social Security and Medicare trust-fund health from the annual Trustees Reports: depletion timing, projected benefit cuts, actuarial balance.
Directly threatens retiree income, but the timing and shape of any change is a political choice — so it rides behind fiscal and legislative.
Demographics (age dependency, fertility) and news-based economic policy uncertainty.
Sets the long-run baseline but moves too slowly to dominate a 20–30 year planning horizon.
Inflation versus target, unemployment, GDP growth, money supply, Fed policy stance.
The noisiest, most mean-reverting signals — deliberately the smallest weight.
From Raw Data to Scores: Percentiles, Not Opinions
Signals with deep history are scored as percentiles of their own full historical distribution — debt-to-GDP against every quarter since 1966, the fed funds rate against every month since 1954, policy uncertainty against every day since 1985. No hand-picked cutoffs, no ceiling the data can hit. A reading is a plain-English claim — “fiscal pressure is at its 94th percentile since 1962” — checkable by anyone with the same public data.
Signals with natural units keep them. Years until trust-fund depletion is a countdown, not a distribution. Fertility is measured as distance below the 2.1 replacement rate. The rule: percentiles for signals that cycle, natural scales for countdowns and thresholds.
The Legislative Layer — AI With Receipts
Congressional activity can't be scored with a spreadsheet, so the legislative component uses language models under a fixed protocol: models and scoring rubrics are version-pinned, every bill's score is cached and re-scored only when its status or the rubric version changes, and each reading publishes its receipts — the specific bills driving the number, so you can check our reading of a bill against the bill.
The scale is signed. A bill that cuts deficits, reforms entitlements, or raises revenue scores as relief, not as zero. An index that can't register Congress doing the right thing isn't measuring Congress. The aggregation on top of the model output is deterministic arithmetic — status weights, bipartisanship, rank decay — published, not vibes.
Two Series: the Reading and the Gauge
The headline index is the slow series — it moves on structural forces, not market spasms. A separate shock gauge watches the fast signals: equity volatility, credit spreads, oil, geopolitical risk, and prediction-market pricing of fiscal and political events. The gauge has trigger floors by design — a VIX of 18 is not a shock — and can only push the effective risk level up, never down. When the level is shock-elevated, we say so explicitly rather than letting a market spasm masquerade as structural change.
Reading the Number
The index runs 0–100, higher is more pressure, in five bands: Low (under 20), Moderate (20–39), Elevated (40–59), High (60–79), and Severe (80+). Band changes carry hysteresis: the published band only changes when the score clears a boundary by more than the index's own noise, so a score hovering at 59–61 reads as one regime, not a weekly flip-flop.
The Margin of Error — Published, Like a Poll
We measured the pipeline's own noise by running the full calculation repeatedly on identical data. The deterministic components reproduce to four decimal places; the legislative layer varies with which bills the discovery pass surfaces. Net result: the published index carries a margin of error of ±4 points. Moves smaller than that are within methodology noise, and we won't narrate them as news. We re-measure after any model or rubric change. To our knowledge no other index in this genre publishes its error bars. They all have them.
Data Sources
All sources are public. The pipeline fails loudly when data is missing — it never fabricates a value to fill a gap.
Governance: Nobody Edits the Number
Published readings cannot be edited — by anyone, including us. There is no override. If a reading is wrong (a corrupt upstream data point, a source outage), the only lever is retraction: the reading is unpublished with a stated reason, the reason goes in the public corrections log, and the corrected reading comes from re-running the pipeline — never from a form.
The methodology itself is versioned. Changes to components, weights, transforms, or the legislative protocol ship as a new version, documented here, with the level impact stated. This is version PRI-V1.
What This Index Doesn't Do
- It doesn't predict. A barometer at 68 doesn't tell you which storm or which week — it tells you the pressure.
- The weights are judgment, stated as judgment. Empirically fitting them would require a target variable for household distress and decades of component data; that research program is running, and results will publish when they exist.
- The legislative layer depends on language models. The protocol pins versions and publishes receipts, but a model's reading of a 900-page bill is a reading, not an audit.
- Percentile scores use an expanding window — each new observation shifts history's percentiles microscopically. The methodology version, not the window, is what's frozen.
- The historical series before live publication is a reconstruction, labeled as such wherever it appears.