K.E.Y.N.E.S.Knowledge Engine Yielding Nonstop Economic Scenarios

Models

Every model on this site has a name. The models in production publish forecasts, and each has its own page with an overview, its latest forecasts and its accuracy.

Names read FAMILY-VERSION-MODIFIER: GDP for the GDP nowcast, IND for the industry models; P5 marks the price version of an industry model, which forecasts prices as plain growth rates.

In production

ModelWhat it forecastsPage
GDP-2 This quarter's real GDP growth, updated weekly, from 35 national series; our replica of the New York Fed's Staff Nowcast 2.0GDP nowcast →

The 13 industry models forecast each industry's own monthly series 0 to 12 months ahead. Each industry has one model with two published versions: activity, and prices measured as plain growth rates.

Reference models

These models are not published as forecasts. They are the baselines and building blocks the production models are measured against; their accuracy is on the Industry models page.

IND-G · general model

The 126 national series (FRED-MD) and 4 economy-wide factors. Every industry model is this model plus one industry's series and two industry factors, so IND-G is the baseline for the question "does information about an industry help?" Accuracy on national indicators →

IND-ECON · economy model

All 13 industries at once: 243 monthly series (126 national series and 117 industry activity series) and 17 factors, 4 economy-wide and one for each industry. Its forecasts of industry activity series are compared with each industry model's on that model's page, and it is tested on national indicators on the Industry models page. Its factors are shown below.

Benchmarks

NameWhat it is
ARAutoregression: forecasts a series from its own recent past. Every model is scored against it ("% better than the AR").
RWRandom walk: forecasts no change from the latest value.
DIDiffusion index (Stock and Watson): forecasts a series from principal-component factors.
BRBridge regression: nowcasts a series from its own past and the month's weekly indicators; scored for the general model only.
NYFEDNew York Fed nowcasts: the Staff Nowcasts it published weekly in 2016–2019; scored against GDP-2.

Factors: how the models work

Every model here explains many data series with a few hidden factors. Think of a factor as a hidden index, like an unpublished "state of the economy" number. Each data series is modeled as a mix of a few factors plus noise of its own. A series' loading on a factor is how strongly it moves with that factor, and the models only allow the loadings that make economic sense: a jobs series can load on the labor factor, a steel series on the steel factor. Which series drive which factor is shown on each model's page (GDP-2, for example) and, for the economy model IND-ECON, below.

IND-ECON: what one factor explains

Share of each series' month-to-month variation explained by this factor (top 20 series)

Estimated factor value each month, standard deviations; grey bands are recessions

Estimated on the latest data vintage (1959 on). The factor's sign is arbitrary: it is fixed so that a chosen anchor series loads positively.

IND-ECON: how much the factors explain, series by series

Share of each series' monthly variation explained by the economy-wide factors and by its own industry's factor. The rest is noise specific to the series.

IND-ECON: how the factors relate to each other

Industry factors are nearly uncorrelated with each other

Correlation of monthly factor values, 2000–2026 excluding March–December 2020

Blue: move together; orange: move oppositely; pale: unrelated. The few strong pairs make economic sense: energy and chemicals (oil and gas are chemical feedstocks), energy and retail (gas-station sales are part of retail), autos and retail (car dealers), transport and leisure (travel). So each industry factor carries information the others do not.