Regime-Switching Dynamic Factor Model

Reduces a large panel of N time series (e.g., hundreds of asset returns) into a small set of K unobserved common factors, where K << N. This dimensionality reduction isolates the systematic drivers of co-movement. The observation equation is typically:

X_t = Λ_{S_t} F_t + ε_t

• X_t: Vector of observed variables at time t
• F_t: Vector of latent factors
• Λ_{S_t}: Factor loading matrix dependent on the regime S_t
• ε_t: Idiosyncratic errors, often assumed Gaussian

Unlike static factor models, the sensitivity of each observed series to the latent factors changes with the market state. A stock's beta to the 'market' factor can be fundamentally different in a bull regime versus a bear regime.

• Interpretation: A loading of 1.2 in a low-volatility regime might drop to 0.8 in a crisis, reflecting a structural shift in correlation.
• Mechanism: The matrix Λ_{S_t} is selected from a finite set based on the current state of a hidden Markov chain.

The evolution of the unobservable regime S_t is governed by a first-order Markov process. The probability of switching from a contraction regime to an expansion regime is constant and defined by a Transition Probability Matrix.

• Persistence: Diagonal elements near 1.0 imply sticky, long-lasting regimes.
• Inference: The Hamilton filter is used to compute filtered and smoothed probabilities of being in each regime at every point in time, given the observed data.

The latent factors themselves follow a Vector Autoregression (VAR) whose coefficients and covariance matrix switch regimes. This captures how the dynamics of the macroeconomy change.

F_t = μ_{S_t} + Φ_{S_t} F_{t-1} + η_t

• μ_{S_t}: Regime-specific drift (e.g., negative drift in a recession).
• Φ_{S_t}: Regime-specific propagation matrix.
• η_t ~ N(0, Σ_{S_t}): Shocks with state-dependent volatility, capturing heteroskedasticity.

Estimation is non-trivial due to the latent nature of both the factors and the regimes. The Expectation-Maximization (EM) algorithm is the standard workhorse.

• E-Step: Estimate the latent factors and regime probabilities using the Kalman filter and Hamilton filter.
• M-Step: Update the model parameters (Λ, Φ, Σ, transition probabilities) to maximize the expected log-likelihood.
• Iteration continues until convergence, yielding a local maximum of the likelihood function.

RS-DFMs excel at nowcasting—predicting the current quarter's GDP growth before official release. The model automatically adjusts its weighting scheme based on the inferred regime.

• In a high-uncertainty regime, the model down-weights noisy survey data and relies more heavily on hard production indicators.
• This dynamic re-weighting produces more accurate real-time estimates than static factor models during turning points in the business cycle.

The foundational framework that reduces a large panel of time series to a small number of unobserved common factors. In finance, these latent factors capture co-movement across hundreds of asset returns, interest rates, or macroeconomic indicators. The 'dynamic' aspect means factors evolve according to a vector autoregression (VAR), allowing shocks to propagate through time. Without regime-switching, a standard DFM assumes constant factor loadings and transition dynamics—an assumption that breaks during structural breaks like the 2008 financial crisis or COVID-19.

The engine that allows factor dynamics to shift across regimes. An unobservable discrete state variable follows a first-order Markov process, meaning the probability of being in a given regime depends only on the previous period's state. Key components:

• Transition probability matrix: Defines persistence and switching likelihood
• Ergodic probabilities: Long-run proportion of time spent in each regime
• Filtered vs. smoothed probabilities: Real-time inference vs. full-sample hindsight This mechanism enables the model to endogenously identify bull, bear, and crisis regimes without pre-specified break dates.

A defining feature where the sensitivity of observed variables to latent factors changes across regimes. For example:

• In low-volatility regimes, sector ETFs may load weakly on a common market factor
• In crisis regimes, correlations spike and all assets load heavily on a single 'panic' factor This captures correlation breakdown phenomena where diversification benefits evaporate precisely when needed most. Estimation requires the Expectation-Maximization algorithm or Bayesian MCMC methods to jointly infer factors and regime states.

The workhorse estimation algorithm that combines state-space filtering with regime inference. At each time step, the filter:

• Predicts the latent factor vector using regime-conditional transition equations
• Updates estimates based on new observations via the Kalman gain
• Weights predictions across all possible regime paths using Hamilton's filter For a model with 2 regimes and 200 time periods, naive evaluation requires 2^200 paths. The collapsing procedure approximates this by merging Gaussian mixtures at each step, making estimation computationally tractable.

Unlike linear DFMs that produce a single forecast, this model generates regime-specific predictions weighted by inferred state probabilities. Practical outputs include:

• Density forecasts: Full predictive distributions capturing fat tails
• Scenario analysis: 'What if we remain in a high-volatility regime for 6 months?'
• Turning point signals: When the smoothed probability of a recession regime crosses 50% Portfolio managers use these to implement regime-responsive asset allocation, reducing equity exposure when crisis regime probabilities spike.

The standard maximum likelihood approach for calibrating regime-switching DFMs when both factors and regimes are latent. The algorithm iterates between:

• E-step: Compute expected values of latent factors and regime states given current parameter estimates, using the Kalman smoother and Hamilton filter
• M-step: Maximize the complete-data log-likelihood to update factor loadings, VAR coefficients, and transition probabilities Convergence is monitored via the observed log-likelihood, though EM can converge to local optima. Multiple random restarts or Bayesian MCMC alternatives address this.