The industry models
Forecasts of monthly indicators a few months ahead: for 13 industries, from steel output to freight volumes and exports, and for the economy as a whole (industrial production, jobs, unemployment, prices).
Latest forecasts, series by series · industry models
Each industry model's forecast for every series it targets, next to the AR benchmark's.
The idea: a few hidden drivers behind many series
Hundreds of economic statistics rise and fall together, because they respond to the same few underlying forces: the business cycle, inflation pressure, interest rates, and conditions in a particular industry. The models on this page treat these forces as hidden variables, called factors, and estimate them from the data. Each series is explained as a weighted mix of the factors plus a part of its own. Forecasting the factors forward then gives a forecast of every series.
How this relates to the GDP nowcast. The GDP nowcast is built on the same idea, but it is a separate model with a different job. It estimates a single number, this quarter's GDP growth, from 35 hand-picked national series, and reports its uncertainty. The models on this page forecast many monthly series, 0 to 12 months ahead, from much larger data sets. They test whether detailed industry data improve forecasts.
Three models: one baseline, two ways to add industry data
| Model | Data | Factors | The question it answers | Results |
|---|---|---|---|---|
| General model IND-G the baseline · reference |
The 126 national series (FRED-MD) | 4 economy-wide factors | Do economy-wide factors forecast better than a simple benchmark? | National indicators |
| Industry models IND-<INDUSTRY>, IND-<INDUSTRY>-P5 13 separate models, one per industry, each in an activity and a price version · production |
The 126 national series plus 9–19 series from one industry | The 4 economy-wide factors plus 2 for the industry: its activity and its prices | Does information about an industry improve forecasts for that industry? | Industry by industry · One page per model |
| Economy model IND-ECON one model, all 13 industries at once · reference |
The 126 national series plus the activity series of all 13 industries (243 series) | 4 economy-wide factors plus 13 industry factors | Do industry factors improve the national forecasts? | National indicators · Factors over time |
Industry by industry · industry models
Each industry model forecasts its own industry's series. Each dot below is one of those series.
Each dot is one forecast target: % smaller errors than the AR benchmark. Right of the line = better than the AR. Hover a dot for the series.
Gains are largest where an industry has several timely series that move together, such as trade flows, or freight, tonnage and rail carloads. They are smallest where the target is a broad payroll total that its own past already predicts well (health care, leisure).
One page per model
Each industry model has two published versions, activity and prices, and each version has its own page: its latest forecasts, its accuracy target by target, what its factors explain, and its data.
National indicators · general and economy models
Here the general model and the economy model forecast the main national series, without and with the 13 industry factors. On national indicators the general model alone is about level with the AR benchmark. Its forecasts help in recessions but lose a little in long expansions, and 2013–2026 is mostly expansion. Averaging a model's forecast with the AR's ("+ AR") usually beats both. With all 13 industry factors, the economy model combined with the AR beats the AR on industrial production at every horizon.
Error relative to the AR benchmark (below 1 = better); 2000–2026, real-time data vintages. Green: at least 2% better than the AR. DI is the Stock–Watson diffusion-index forecast, a standard benchmark.
The model in detail
For economistsEquations
Each series \(i\), transformed to be stationary and standardized, is a common part driven by the factors plus a series-specific part:
$$x_{it} = \lambda_i^{g\prime} g_t + \lambda_i^{s}\, s_{k(i),t} + e_{it}$$ $$f_t = A_1 f_{t-1} + A_2 f_{t-2} + u_t, \qquad f_t = (g_t', s_t')', \qquad u_t \sim N(0, Q)$$ $$e_{it} = \rho_i\, e_{i,t-1} + \varepsilon_{it}, \qquad \varepsilon_{it} \sim N(0, \sigma_i^2)$$- \(g_t\): the 4 economy-wide factors; \(s_{k,t}\): the factor of industry \(k\).
- Loadings are restricted: national series load only on the economy-wide factors; an industry series loads on those and on its own industry's factor.
- The industry models have the same form with 6 factors: the 4 general ones plus an industry activity factor and an industry price factor.
For economistsSize, estimation and testing
- Economy model: 243 series (126 FRED-MD, 117 industry activity), monthly from Mar 1959 (about 17% of data points missing), 17 factors with 2 lags, a 277-element state and 2,306 estimated parameters.
- Estimation: maximum likelihood by the EM algorithm with the Kalman filter and smoother (Bańbura and Modugno 2014), which handles missing values and ragged data edges directly.
- Data handling: FRED-MD's transformation codes and the McCracken–Ng outlier rule (values more than 10 interquartile ranges from the median treated as missing).
- Choices: 4 economy-wide factors and 2 lags were chosen by forecast accuracy over 2000–2012 and then kept fixed, so the later test is not tuned to its own results. Industries get one factor each by design.
- Real-time test: every month from Jan 2000 to the latest FRED-MD vintage, the model is fitted on the FRED-MD vintage of that month and forecasts 0 to 12 months ahead. Industry series use their ALFRED archives, which begin between 2010 and 2017. Industry scores on this page count only forecast dates after an industry's archives begin (for the economy model, after all 13 have begun, Nov 2017), so every forecast uses data as first published. Industry models are re-fitted every third month.
- Scores: root mean squared error relative to the AR, on cumulative changes; Diebold–Mariano tests with the Harvey–Leybourne–Newbold correction for p-values.
- Prices: with FRED-MD's default transformation (change in the growth rate), industry price forecasts mostly lose to the AR. The "price t-code 5" comparison uses the plain growth rate for both the model and the AR.