Effective date: October 17, 2026. Revision: 1.
Several gov.house data pages show a projection after the latest official figure: a dashed line with a shaded range. This page explains how those projections are made, how we check them and what they are not.
§1What a projection is, and is not
A projection is a statistical extrapolation of a series from its own history. It does not know about upcoming legislation, Fed decisions, wars or storms, and it is not a gov.house view on what policy should be. It is not investment, financial or tax advice. Its job is narrower: to show where the series is heading if it keeps behaving as it has, and how wide the usual uncertainty is.
§2The models
Each series is first put on a regular calendar: weekly (last value of each week) for daily and weekly data, monthly for monthly data. The projection is then the average of standard, well-documented forecasting methods:
- Damped-trend exponential smoothing (Holt’s method with a damped trend), which follows the recent level and lets any trend fade out instead of extending it forever.
- An autoregressive model on monthly or weekly changes (ARIMA(p,1,0)), with the number of lags chosen by the Akaike information criterion.
- The Theta method, used only for series with a persistent trend, such as the national debt.
For CPI inflation we use a model built for how the 12-month rate is calculated. We project the monthly change in the price index, as the recent underlying pace (the last six months, seasonally adjusted, gradually fading towards the three-year pace) plus each calendar month’s typical seasonal pattern over the past five years, and then compute the 12-month rate from the projected index. This captures base effects: when an unusually large monthly increase from a year earlier drops out of the 12-month window, the rate falls, and the projection shows it. Months for which no data were published, such as October 2025, are filled by geometric interpolation of the index.
Series that move mainly on policy decisions, such as the federal funds target range and the Treasury cash balance, are not projected.
§3The benchmark and the backtest
Every time the data are refreshed, we re-run the models as if we had stood at each of the past several dozen weeks or months, using only the data available at that point, and compare their projections with what actually happened. We do the same for the simplest possible forecast: no change from the latest value. For many financial series this benchmark is very hard to beat.
The data page reports how much more (or less) accurate the models were than the no-change benchmark across all horizons, and the average miss at the longest horizon shown. If the models do not beat no change by at least 2%, the page shows the no-change projection instead, and says so. We would rather publish a flat line than a confident curve that has not earned it.
§4The 80% range
The shaded range is not a theoretical formula. At each horizon it is set by the 80th percentile of the absolute errors from the same backtest, so in past data about four outcomes in five fell inside it. It never narrows as the horizon lengthens. Ranges can be too narrow when conditions change sharply, which is exactly when projections are least reliable.
§5The live record
Each time a new official figure is published, gov.house stores the projection made at that moment. Once the projected periods arrive, the data page compares the stored projections with the published figures and reports the average miss next to the no-change benchmark. This record cannot be edited after the fact.
§6Revisions and corrections
Projections are recalculated automatically when the source agency publishes new or revised data. If we find an error in the method or the code, we correct it, note the change on this page with a new revision number, and follow our corrections policy.
§7Sources
All inputs are the official series credited on each data page: the U.S. Treasury (Fiscal Data, Daily Treasury Par Yield Curve Rates), the Bureau of Labor Statistics, Freddie Mac’s Primary Mortgage Market Survey and the Federal Reserve Bank of St. Louis (FRED). The calculations run on gov.house’s own server; no third-party forecasting service or AI model is involved.