State-Dependent Regime Transition Probabilities

Definition

A state-dependent regime transition probability is a Markov-chain transition matrix ${\Pi }_{ij}(X_t)=P(s_{t+1}=j|s_t=i,X_t)$ whose entries are explicit, typically smooth, functions of the contemporaneous (latent or observed) state vector $X_t$. This generalizes the standard Markov-switching model — where $\Pi$ is a constant — by letting expected regime durations and switching hazards respond to the macro / financial state.

Intuition

In a constant-$\Pi$ Markov-switching model, the expected duration of regime $i$ is $1/(1-{\Pi }_{ii})$, a single number for the entire sample. This is empirically implausible for monetary policy regimes: a tightening cycle is more likely to end when inflation has cooled and the curve has flattened, not at a fixed average rate. State-dependent $\Pi (X_t)$ lets the data speak: when the level or slope of the yield curve (or, equivalently, the macro factors driving it) points one way, the conditional probability of switching shifts.

The economic content is that the macro state forecasts regime change. The identification works because regime transitions are observed (probabilistically, through the smoother) and you can ask whether they cluster in particular regions of the state space.

Formal notation

Two common parametrizations:

1. Multinomial logit (used in DSY 2007): ${\Pi }_{ij}(X_t)=exp({\gamma }_{ij,0}+{\gamma }_{ij,1}^{\prime }{X}_t)/{\Sigma }_{k}exp({\gamma }_{ik,0}+{\gamma }_{ik,1}^{\prime }{X}_t)$

   with one row of identifying restrictions (typically set the parameters of $j=i$ to zero).

2. Probit / standard normal: ${\Pi }_{ij}(X_t)=\Phi ({\gamma }_{ij,0}+{\gamma }_{ij,1}^{\prime }{X}_t)$

   for the case $M=2$, where $\Phi$ is the standard normal CDF.

The risk-neutral counterpart ${\Pi }^Q(X_t)$ is obtained by twisting ${\Pi }^P(X_t)$ with a regime-shift Radon–Nikodym density ${\xi }_{ij}(X_t)$ whose log is itself affine in $X_t$:

${\Pi }_{ij}^Q(X_t)\propto {\Pi }_{ij}^P(X_t)\cdot exp({\eta }_{ij,0}+{\eta }_{ij,1}^{\prime }{X}_t)$

normalized so each row sums to one.

Variants

• Filardo (1994): state-dependent transition probabilities with observed macro indicators in a non-pricing Markov-switching regression.
• Diebold–Lee–Weinbach (1994): time-varying-transition Markov regression, early reference for the multinomial-logit form.
• DSY 2007 (dai-singleton-yang-2007-regime-shifts-term-structure): state-dependent transitions wired into a no-arbitrage Gaussian DTSM, with closed-form bond prices via a log-linear approximation of $log{\Pi }^Q(X_t)$.
• Endogenous-volatility variants: $\Pi (X_t)$ made a function of realized volatility rather than the level of factors.

Comparison

Parametrization	Closed-form bond pricing	Identification difficulty	Expected duration time-varying
Constant $\Pi$	Yes (exact)	Easy	No
Logistic $\Pi (X_t)$	Yes (after log-linear approx)	Moderate (needs many regime transitions)	Yes
Probit $\Pi (X_t)$	Yes (after log-linear approx)	Moderate	Yes
Free-form $\Pi (X_t)$	Numerical only	Hard (curse of dimensionality)	Yes

When to use

• The Markov-switching model is being used to capture regimes whose duration visibly correlates with macro state (monetary policy cycles, recession phases, ZLB episodes).
• You have enough in-sample regime transitions to identify the parameters of $\Pi (X_t)$ — usually requires at least 4–6 switches and a sample with meaningful state-space variation.
• You need a model where the term premium or long-yield expectations respond to the probability of an upcoming regime change, not just the fact of the current regime.

Known limitations

• Identification is thin when there are few historical regime transitions; the parameters of $\Pi (X_t)$ can be very weakly pinned down even when the reduced-form fit improves.
• Closed-form bond pricing under state-dependent $\Pi$ requires a log-linear approximation that is non-uniform across the state space (worst exactly at the regime boundary, where the model matters most).
• Coupled with priced regime-shift risk, the model has many parameters and estimation can settle into shallow local optima — each new degree of freedom in ${\gamma }_{ij,1}$ widens the basin of admissible parameter vectors.
• Latent vs observed factors: when $X_t$ is latent, the identification of $\Pi (X_t)$ becomes a joint problem with the factor extraction, which can mask state-dependence in finite samples.

Open problems

• A principled way to choose between logistic, probit, and free-form parametrizations of $\Pi (X_t)$ other than in-sample BIC.
• Particle-filter / Rao–Blackwellized particle filter implementations of state-dependent transitions: the regime path is harder to integrate out analytically when $\Pi$ depends on the same $X_t$ that the filter is updating.
• Extension to compound Markov chains (multiple independent sub-chains) where each sub-chain has its own state-dependent transition function — does identification break, or can the product structure be exploited?

Key papers

• dai-singleton-yang-2007-regime-shifts-term-structure — embeds state-dependent transitions in a no-arbitrage Gaussian DTSM and shows state-dependence improves yield fit and identifies the regime-shift risk premium more sharply than constant-$\Pi$ baselines.

My understanding

For the current CRE asset pricing project, the choice to use constant $\Pi$ is a deliberate simplification driven by RBPF tractability — the regime path is analytically integrable conditional on the continuous state precisely because $\Pi$ doesn’t depend on the continuous state. Adopting state-dependent $\Pi (X_t)$ would couple the regime sampler back to the Kalman update and likely break Rao–Blackwellization (or at least require a much heavier within-particle correction). DSY 2007 is the canonical evidence that this simplification is costly on yields data; whether it is costly on cap rates / CRE prices is an open empirical question for the project.