Empirical forecasting models · Model guide
Dynamic factor model: question, structure, and use cases
Extracts a small number of latent factors from many macro indicators and lets those factors drive a current-quarter read.
How do you extract the common business-cycle signal from a large panel of economic indicators?
Background
Macroeconomic datasets contain hundreds of time series - output measures, employment, prices, financial indicators, surveys, trade flows - that all respond to a small number of underlying shocks: aggregate demand, aggregate supply, monetary policy, global risk appetite. The dynamic factor model (DFM) formalizes this observation. A small number of latent factors (typically 1-5) drive the co-movement across the entire panel. Each observed variable is a linear combination of these common factors plus an idiosyncratic (variable-specific) component. The factors are unobserved; they are inferred from the cross-sectional covariance structure of the data.
Geweke (1977) and Sargent and Sims (1977) introduced the static factor model to macroeconomics. Stock and Watson (1989, 2002a, 2002b) developed the modern dynamic factor framework used in applied work today. Their key innovation: when the cross-section n is large (n > 20), the factors can be consistently estimated by principal components of the data matrix, even without specifying the factor dynamics or the idiosyncratic error structure. This 'large n, large T' asymptotic framework made factor models practical for panels of 100-200 variables without requiring full maximum likelihood estimation.
The state-space representation (Forni, Hallin, Lippi, Reichlin 2000; Doz, Giannone, Reichlin 2011) casts the DFM as a Kalman filter problem: factors follow a VAR in the state equation, and observed variables are noisy linear projections of the factors in the observation equation. The Kalman filter handles missing data, mixed frequencies, and ragged edges naturally. The New York Fed's nowcasting model and the ECB's real-time factor model both use this state-space approach.
DFMs dominate macroeconomic forecasting competitions. Stock and Watson (2002b) showed that factor-augmented regressions beat univariate benchmarks for 215 U.S. macro series at horizons of 1-24 months. The factor-augmented VAR (FAVAR, Bernanke, Boivin, Eliasz 2005) combines extracted factors with policy variables in a VAR to identify monetary policy effects in a data-rich environment. Central banks worldwide - the Fed, ECB, Bank of Canada, Reserve Bank of Australia - maintain DFM-based forecasting and monitoring systems.
How the Parts Fit Together
The input is a T x n panel of stationary time series, typically standardized to zero mean and unit variance, with rows as dates and columns as variables. The panel contains variables from multiple economic categories: output/production, employment/labor, prices/wages, interest rates, exchange rates, money/credit, surveys, trade. The cross-section dimension n ranges from 20 (small DFM) to 200+ (large DFM). Each variable is classified by category and transformation (level, first difference, log difference) to ensure stationarity.
The model separates each variable into two components: a common component driven by r latent factors, and an idiosyncratic component specific to that variable. The common factors follow a VAR(p) in the state equation: ... + eta_t. The observation equation links each variable to the factors: lambda_i' , where lambda_i is the factor loading for variable i. The idiosyncratic errors can be serially correlated (dynamic idiosyncratic) or white noise (exact factor model).
Estimation follows one of two paths. (A) Principal components: with X stored as T x n, compute X'X/T, take the first r eigenvectors as the factor-loading space, and extract factors as f_hat_t = Lambda_hat' . Equivalently, the factor scores are the first r principal components of the T x n data matrix. (B) State-space/Kalman filter: specify the full state-space model (state equation + observation equation), estimate parameters by EM algorithm or quasi-maximum likelihood, and extract factors by the Kalman smoother. Path (A) is fast and requires no distributional assumptions; path (B) handles missing data and mixed frequencies but requires iterative estimation.
Applications
The New York Fed's Weekly Economic Index (WEI) and its predecessor, the dynamic factor model used for nowcasting (Bok et al. 2018), extract common factors from a panel of 37 monthly and quarterly indicators covering output, labor, housing, consumption, surveys, and financial conditions. The model operates in state-space form with the Kalman filter, handling mixed frequencies (monthly and quarterly data) and ragged edges (different publication dates) within a unified framework. The extracted factor becomes a coincident economic index and a direct input to the GDP nowcast.
The FAVAR (Bernanke, Boivin, Eliasz 2005) is the dominant tool for studying monetary policy transmission in a data-rich environment. The idea: extract factors from a large panel (100+ macro and financial variables), then include the extracted factors alongside the federal funds rate in a VAR. Structural identification (Cholesky, sign restrictions) applied to the VAR yields impulse responses of every variable in the panel to a monetary policy shock. The FAVAR avoids the dimensionality problem of including all 100+ variables directly in the VAR.
The ECB's area-wide model and the Eurosystem's real-time forecasting infrastructure use DFMs to handle the heterogeneous data availability across euro area member states. Some countries publish GDP estimates earlier than others; some indicators are monthly in one country and quarterly in another. The state-space DFM imputes missing values, aggregates across countries, and produces a euro area factor that tracks aggregate economic activity. The Banbura and Modugno (2014) implementation is the methodological backbone.
DFMs struggle when the data-generating process changes structurally: the COVID-19 pandemic produced factor estimates that were orders of magnitude larger than any historical observation, rendering the Gaussian state-space model's forecast intervals meaningless. The linear factor structure also misses sector-specific dynamics: a shock concentrated in one sector (e.g., oil prices affecting energy production) may not be captured by factors estimated from the full panel. Factor-augmented models with sector-specific factors (Kose, Otrok, Whiteman 2003) partially address this at the cost of additional complexity.
Components
The r x 1 vector of unobserved common factors at time t. These capture the co-movement across the n observed variables.
The n x r matrix of factor loadings. Row i (lambda_i') measures how variable i loads on each of the r factors.
The n x 1 vector of variable-specific errors. : idiosyncratic and common components are orthogonal.
The dimension of the latent factor space. Selected by information criteria (Bai and Ng 2002) or scree plot of eigenvalues.
The r x r matrices governing the dynamics of the factors in the state equation. Capture the persistence and cross-dynamics of the common factors.
The r x 1 vector of shocks to the factors. . These are the primitive common shocks before rotation/identification.
The n x n covariance matrix of . Diagonal in the exact factor model; block-diagonal or sparse in the approximate factor model.
Assumptions
As n grows, the r largest eigenvalues of the n x n covariance matrix diverge (they grow with n), while all remaining eigenvalues stay bounded. The common component explains a non-vanishing fraction of the total variance in the cross-section.
If violated: If the factor structure is weak (eigenvalues grow slowly), principal component estimates of the factors are inconsistent. A weak factor loads on only a subset of variables; it can be missed by methods that assume pervasive factors.
All variables in the panel are covariance stationary after appropriate transformations (differencing, log transformation). The factor dynamics are stable (companion matrix eigenvalues inside the unit circle).
If violated: Non-stationary variables create spurious factor estimates: unit-root-driven trends dominate the first principal component regardless of economic content.
The idiosyncratic errors are weakly cross-sectionally correlated. Strong clustering (e.g., all energy variables have correlated idiosyncratic shocks) violates the approximate factor model assumptions.
If violated: Strong idiosyncratic cross-correlation inflates the eigenvalues of the covariance matrix, leading the number-of-factors criteria (Bai-Ng) to overestimate r.
The number of common factors r is correctly specified. Bai-Ng IC criteria, the Onatski (2010) test, or the Ahn-Horenstein (2013) eigenvalue ratio test provide data-driven guidance.
If violated: Too few factors: the omitted factor's variance is absorbed into the idiosyncratic errors, creating cross-correlated residuals and biased factor-augmented regressions. Too many factors: noise is elevated to factor status, adding estimation variance without improving the common component.
Each variable is a linear function of the common factors. No threshold effects, no nonlinear factor loadings.
If violated: Nonlinear dependence on the factors (e.g., financial variables respond asymmetrically to positive vs. negative factor realizations) is not captured. Nonlinear factor models exist but are computationally demanding and rarely used in macro practice.
The panel is either balanced (no missing observations) or missing data is handled by the EM algorithm / Kalman filter. Dropping variables with missing data instead of imputing can bias factor estimates if missingness is non-random.
If violated: Listwise deletion of variables with missing data reduces the effective cross-section and can eliminate informative series. The EM algorithm iteratively imputes missing values using the current factor estimates, then re-estimates factors.
Continue reading
Concepts, data, and nearby models
Open the concept, data series, policy setting, or neighboring model that anchors this page.