Dynamic Factor Model

Definition

A dynamic factor model (DFM) represents a $q$ -dimensional time series $y_t$ as a linear combination of $k < q$ latent factor processes plus idiosyncratic noise: $y_t = \theta + X f_t + v_t$ , where $X$ is a $q \times k$ factor loadings matrix, $f_t \sim N(0, H_t)$ with time-varying diagonal $H_t$ , and $v_t \sim N(0, \Psi_t)$ with diagonal idiosyncratic variance matrix $\Psi_t$ . The conditional covariance decomposes as $\Sigma_t = X H_t X' + \Psi_t$ (low-rank-plus-diagonal). Dynamic models extend static factor analysis by allowing $H_t$ and $\Psi_t$ to evolve over time as stochastic processes.

Key Ideas

Dimensionality reduction: Reduces free covariance parameters from $O(q^2)$ to $O(qk)$ loadings plus $q$ idiosyncratic variances plus $k$ factor variance processes — critical for inference when $q$ is large.
Identification (Geweke-Zhou 1996): The loadings matrix $X$ is generically unidentified up to rotation ( $X f_t = (XR)(R^{-1}f_t)$ for orthogonal $R$ ). The standard convention fixes $X$ to be lower-triangular with unit diagonal, requiring a meaningful ordering of the $q$ series (e.g., placing the "dominant" series first).
Stochastic Volatility (SV) extension: Factor log-variances $\lambda_{jt} = \log h_{jt}$ follow a vector autoregression (VAR(1)): $\lambda_t = \mu + \Phi(\lambda_{t-1} - \mu) + w_t$ , $w_t \sim N(0, U)$ . Off-diagonal entries of $U$ capture co-movement in factor volatility processes. Idiosyncratic log-variances follow independent autoregressive (AR(1)) models (Aguilar-West 2000).
GARCH extension: Factor variances can follow GARCH (Generalized AutoRegressive Conditional Heteroskedasticity) processes (Engle-Ng-Rothschild 1992 factor-ARCH; Demos-Sentana 1998). Easier to estimate by Maximum Likelihood Estimation (MLE) but less flexible for Bayesian analysis.
Relationship to state-space: DFMs are special cases of state-space models with $f_t$ as the state vector; Forward Filtering Backward Sampling (FFBS, Carter-Kohn 1994) applies directly to draw $\{f_t\}$ in one block given the variance paths.
Connection to Cholesky discounting: The variance discounting baseline (West-Harrison 1997) recovers a $k=q$ factor structure via the Cholesky decomposition of the smoothed estimate $S_t$ , but with no parsimonious factor reduction.

How It Works

Set $k$ : Choose the number of factors guided by the Cholesky decomposition of the sample covariance or subject-matter knowledge. Excess factors cause Markov Chain Monte Carlo (MCMC) convergence failure (a useful diagnostic).
Order series: Place the "dominant" series first so that the lower-triangular constraint is substantively interpretable.
MCMC (Aguilar-West scheme):
- Draw latent factor paths $\{f_t\}$ via FFBS treating $H_t, \Psi_t$ as given.
- Draw log-volatility paths via Kim-Shephard-Chib (KSC, 1998) mixture-of-normals + FFBS for the multivariate dynamic linear model (DLM).
- Draw loadings $X$ equation-by-equation exploiting the lower-triangular constraint.
- Draw idiosyncratic SV parameters via univariate KSC smoother.
- Draw $U$ via Metropolis-Hastings with inverse-Wishart proposal.
Sequential updating: Fix parameters at MCMC posterior means from a training window; use the auxiliary particle filter (Pitt-Shephard 1999) to update volatility states in real time.

Why It Matters

Portfolio allocation: One-step-ahead $G_t = V(y_t|D_{t-1}) = X\hat{H}_t X' + \hat{\Psi}_t$ feeds directly into mean-variance optimization; the factor structure makes matrix inversion tractable via the Woodbury identity.
Interpretability: Factors may correspond to economically meaningful latent drivers (e.g., a "European currency factor" dominating DEM/FRF/ESP co-movements).
Selective intervention: Named factors allow targeted prior interventions when one component is expected to change (e.g., an anticipated central bank policy shift affects only the JPY factor, not DEM).
Volatility peak sharpness: Unlike the exponential-smoothing (discounting) baseline, the SV factor model captures sharp volatility spikes (e.g., Britain's 1992 Exchange Rate Mechanism (ERM) exit) without over-smoothing.

Open Questions

Number of factors: $k$ is typically fixed; Lopes-West (1998) developed MCMC for joint inference on $k$ in static models; extension to dynamic SV factors remains open.
Sequential parameter learning: Fixing parameters at MCMC means is a pragmatic approximation; simultaneous sequential learning on both parameters and states is computationally hard (Liu-West 2001 partially addressed this; particle MCMC methods provide a fuller solution).
Scaling: Chib-Nardari-Shephard (2006) extends to $p=50$ with a reduced blocking scheme; beyond $\sim 100$ series, full DFM MCMC becomes prohibitive.

Diffusion Indexes: Principal-Components Estimation (Stock-Watson 2002)

The dominant frequentist, large- $N$ approach is the diffusion-index method of Stock and Watson (2002). Writing the panel in static form $X_t = \Lambda F_t + e_t$ (an approximate factor model — some idiosyncratic cross-correlation in $e_t$ is allowed), the factors $F_t$ are estimated by principal components of the $N$ -series panel, which is consistent as $N,T\to\infty$ and essentially free to compute — no Kalman filter or MCMC. Forecasts of a target $y_{t+h}$ are then built by direct $h$ -step projection on the estimated factors and lags, $y_{t+h}=\alpha_h+\beta_h(L)\hat F_t+\gamma_h(L)y_t+\varepsilon_{t+h}$ , avoiding the error accumulation of iterating a factor VAR. Because $N$ is large the estimated factors behave like the true factors (negligible generated-regressor problem). Empirically, a handful of PC factors from 215 U.S. predictors beat AR, small-VAR, and leading-indicator forecasts by roughly a third in MSFE. This is the frequentist counterpart to the Bayesian state-space DFM above (PCA vs. Gibbs/Kalman) and the basis of the FAVAR and nowcasting literatures.

Bayesian shrinkage as an alternative to principal components (De Mol-Giannone-Reichlin 2008)

Instead of first extracting factors, one can forecast $y_{t+h}$ by Bayesian regression on all $N$ predictors with a shrinkage prior. De Mol, Giannone and Reichlin show that a Normal prior (ridge regression) and a double-exponential prior (lasso) both produce forecasts that are highly correlated with, and forecast as well as, the principal-component forecasts across a wide range of prior tightness. The reason is the same factor/collinearity structure: when a few large eigenvalues dominate the $N\times N$ design, ridge shrinkage projects onto those leading directions, so as $N\to\infty$ (with the prior tightened at the right rate) the ridge forecast converges to the PC forecast — an exact link between the principal-component and Bayesian-shrinkage views of large- $N$ forecasting. Lasso instead yields a sparse forecast (a few selected predictors), useful when one wants interpretability rather than a dense factor combination. The practical lesson: with a large, collinear panel the shrinkage is what matters, not whether it is implemented by factor extraction or by a prior.