How Our Models Are Built

Fourteen monthly series span four blocks. Labor: nonfarm payrolls, average weekly hours, the JOLTS quits rate, and initial jobless claims. Credit: bank lending standards from the Federal Reserve's senior loan officer survey, investment-grade and high-yield option-adjusted spreads, and inflation-adjusted bank loan growth. Rates and the curve: the policy rate, the ten-year minus two-year Treasury spread, and the real ten-year yield. Activity and demand: industrial production, retail sales deflated to real terms, and manufacturing new orders. Leading series such as claims, spreads, the curve, and orders capture directional change; coincident series anchor the level. Exact series identifiers, the retail-sales deflator, and the new-orders source are specified in the production standard.

Each series is converted to a standardized score against a trailing ten-year window, expressing the current observation in standard deviations from its historical mean so that measures reported in different units share one scale. Values are winsorized at the 1st and 99th percentiles, which caps the influence of extreme prints while keeping them in the dataset. Scores aggregate into the four block readings and then into one composite. Weights are held fixed between scheduled methodology reviews and published whenever they change.

Economic series arrive at different frequencies and with different publication delays. Observations are aligned to a common information date before aggregation, so the composite reflects the evidence as of one point in time and never mixes stale and fresh prints mechanically. The production standard governs the treatment of first releases, later revisions, and historical data vintages.

The standardized composite maps onto an index scale anchored at 100 in January 2019. A reading above or below 100 shows the model's position relative to that reference point, and the technical specification discloses how many index points correspond to one historical standard deviation. The level is a model reading; it does not represent a percentage change in output, income, employment, or market value.

From the latest reading the model projects an eighteen-month path under three conditional scenarios. The base path carries current momentum forward. The upside path assumes improving credit conditions and stronger real-income growth. The downside path assumes tighter credit and weaker demand. These are conditional scenarios and carry no probability weights. The scenario band widens with the forecast horizon and with disagreement among the fourteen inputs, so a conflicted signal reads as a wider band. The statistical procedure that will set the band's width in production is documented in the production standard.

The composite maps to a regime band spanning recovery, expansion, late cycle, and contraction.

Signal agreement measures how consistently the fourteen inputs point in the same direction. It is a breadth measure; it does not state the probability that the model is correct. That question belongs to validation.

The Index also displays a twelve-month recession reading in the tradition of Estrella and Mishkin and the New York Fed's yield-curve model. The displayed figure is illustrative. It becomes a live calibrated probability when the target definition, estimation method and sample, recession chronology, data-vintage policy, and out-of-sample results are published under the production standard.

Rate differentials compare U.S. short-term rates and real yields with those of selected trading partners, an uncovered-interest-parity lens. Positioning measures speculative dollar exposure against its own history, drawn from CFTC futures data and survey evidence. Terms of trade incorporate trade balances and commodity exposure in the commodity-currency tradition of Chen and Rogoff. Valuation compares each currency with a slow-moving fair-value anchor built from purchasing-power-parity and behavioral equilibrium estimates. The instruments, contract coverage, currency universe, treatment of pegged and managed currencies, and anchor specification are documented in the production standard.

Each driver converts to a standardized score on a common scale, and the four scores combine into a directional reading for the broad dollar and, where data permit, for individual currencies. Rate differentials and positioning carry more weight at short horizons; valuation and terms of trade carry more weight as the horizon lengthens. The horizon weights are fixed between reviews and published, so any call decomposes back into its four drivers.

The conviction score runs from 0 to 10 and measures how strongly the four drivers align behind the current call. A score near 10 means rates, positioning, terms of trade, and valuation lean the same way, so the direction is well supported even where the magnitude is uncertain. A low score flags a divided setup in which the dollar is more likely to range. The score carries no information about the expected size of a move or the probability that the drivers are jointly right.

In the Index, the model signal summarizes direction and driver agreement for the broad dollar. The map's other lenses supply descriptive context: each economy's currency performance against the dollar, local-currency equity performance, and central-bank policy rates. Those observed measures sit alongside the model and are labeled accordingly. The broad-dollar reading is benchmarked to the Federal Reserve's nominal broad dollar index from the H.10 release, the Fed's trade-weighted measure across twenty-six partner currencies; where a real, inflation-adjusted reading appears, it is identified as the Fed's monthly real broad dollar index.

Three layers. Hazard draws on federal and specialist datasets covering flood, wind, wildfire, heat, and severe storms. Exposure estimates the property value sitting within relevant hazard areas, with the valuation basis defined in the production standard. Market tracks homeowner premium growth and coverage pressure: insurer withdrawals, nonrenewals, underinsurance, FAIR plan growth, and migration into other residual markets, with uninsured and underinsured exposure reported separately where the data permit.

The climate-risk score combines hazard with the value exposed to it, on the principle that financial risk depends on the assets in a hazard's path as well as on the hazard itself. A moderate hazard over dense, high-value property can therefore create more balance-sheet pressure than a severe hazard over sparsely developed land. The three published lenses share the same state spine and remain distinct measures: the climate-risk score is the hazard-by-exposure composite, while premium growth and coverage pressure are read from insurance-market filings and residual-market data. The production standard sets the hierarchy that keeps overlapping hazard and loss datasets from double counting.

State rankings name the measure being ranked: fastest premium growth, largest increase in coverage pressure, highest hazard-by-exposure score, fastest growth in residual-market exposure. The premium-growth lens compares homeowner premium growth with shelter inflation, benchmarked to the CPI owners' equivalent rent series; because owners' equivalent rent measures shelter inflation generally, a state reads as repricing when insurance costs outrun housing costs at large. Coverage pressure tracks the weakening or substitution of private capacity through insurer retreat and residual-market growth. A full estimate of uninsurable or uninsured value across jurisdictions is a separate calculation, published only when the underlying data support it.