Academic and research sources
Stock and Watson (2002a) 'Forecasting Using Principal Components from a Large Number of Predictors': the theory paper (JASA) establishing consistency of principal-components factor estimation and factor-augmented forecasting in large panels.
Journal of the American Statistical Association · 2002
Stock and Watson (2002b) 'Macroeconomic Forecasting Using Diffusion Indexes': the empirical companion (JBES) demonstrating forecast gains from diffusion indexes estimated on a 215-predictor U.S. panel.
Journal of Business & Economic Statistics · 2002
Bernanke, Boivin, and Eliasz (2005) 'Measuring the Effects of Monetary Policy: A Factor-Augmented Vector Autoregressive (FAVAR) Approach': combined factor extraction with structural VAR identification.
The Quarterly Journal of Economics · 2005
Bai and Ng (2002) 'Determining the Number of Factors in Approximate Factor Models': proposed the ICp1, ICp2 and ICp3 criteria for selecting the number of factors.
Econometrica · 2002
Doz, Giannone, and Reichlin (2011) 'A Two-Step Estimator for Large Approximate Dynamic Factor Models Based on Kalman Filtering': provided the state-space estimation framework used by the ECB and NY Fed.
Journal of Econometrics · 2011
Research footing
Evidence and data
Check panel transformations, missing data, factor count, loading stability, and real-time forecast performance.
Calibration or measurement
Factor number, standardization, state dynamics, and release timing determine the signal extracted from the panel.
Boundaries
- Latent factors are rotations, not direct economic objects.
- Structural breaks can move loadings.
- A wide panel does not remove measurement error.
Use Guidance
When sufficient
Extracting common cyclical signals from a large panel of macro indicators when the object is a low-dimensional summary of real activity, such as a coincident index, a nowcast of quarterly GDP, or a real-activity factor for use in downstream regressions. Stock-Watson 2002 (JBES) diffusion indexes and their state-space extensions are the standard approach when the number of indicators is large but the common variation is concentrated in a few factors.
When sketch only
The extracted factor is a statistical object, not a structural shock. It captures whatever the common variation in the panel happens to be, which may conflate demand, supply, and financial shocks unless the panel is structured to isolate one dimension. Do not interpret the factor as a specific causal driver without an additional identification step.
When to switch
Switch to a FAVAR (Bernanke-Boivin-Eliasz 2005 QJE) when combining the factor with a small set of explicitly identified structural shocks. Switch to an SVAR (empirical:svar) when the variable count is small enough for direct identification without the dimensionality reduction. Switch to a nowcasting state-space model (empirical:nowcasting) when the goal is real-time tracking that explicitly accounts for the publication calendar.
Falsification signal
A factor that explains a small share of common variation across the panel, or that fails to predict the target series out of sample despite a large number of indicators, signals that the factor structure is weak or the panel is too heterogeneous for a low-dimensional factor to summarize it. If the first three factors together account for less than 30 percent of variance in a well-constructed activity panel, the diffusion-index assumption is not supported by the data.