Markov Regime Switching in Asset Prices

Definition

A modeling discipline in which the parameters of an asset-pricing model (drift, volatility, correlation, market price of risk, transition matrix) are deterministic functions of a latent discrete-state Markov chain $s_t\in$ {1, …, K}. The chain follows a transition matrix P (typically constant). No-arbitrage pricing is preserved by adding a regime-shift price of risk that compensates investors for unhedgeable jumps across regimes. Asset prices are computed conditional on a regime path, then aggregated by integrating over the regime distribution.

The latent regimes are typically interpreted as business-cycle states (expansion / recession), monetary policy stances (active / passive), volatility regimes (high / low), or liquidity states.

Intuition

Asset prices have persistent shifts in their conditional distribution that are not well captured by smooth time-varying parameters or by GARCH-style heteroskedasticity. A latent Markov chain captures these shifts as discrete jumps in regime, where each regime has its own constant parameters but the chain switches occasionally. This is dramatically more parsimonious than a fully time-varying model and yields economically interpretable regimes (e.g. recessions, monetary tightening cycles). The price for tractability is the no-arbitrage handling: investors face a non-diversifiable jump risk every time the regime can switch, and the model must price this risk to be consistent.

Formal notation

• Latent regime: $s_t\in$ {1, …, K}, transition matrix P with $P_{ij}=Pr($s_{t+1} = j | $s_t=i)$
• Within-regime dynamics: $y_{t+1}$ $=f(y_t,s_t)+g(y_t,s_t){\epsilon }_{t+1},{\epsilon }_{t+1}$ ~ N(0, I)
• Asset price within regime i: $p_t=h_i(y_t)$
• No-arbitrage with regime-shift premium: pricing kernel $M_{t+1}$ contains both a Brownian-shock price of risk and a regime-shift price of risk ${\lambda }^s$, satisfying $E_t^P[$M_{t+1} (1 $+R_{t+1})]=$ 1 across regimes.
• Conditional bond / asset price: $\mathbb{P}(t,T\mid$s_t = i)$=exp(A_i(\tau )-B_i(\tau {)}^{\prime }{y}_t)($often only approximate when chain has K > 1 regimes)

Variants

• Hamilton (1989) two-state regime model for GDP: original time-series formulation, no asset pricing — establishes the latent-state Markov filter.
• Bansal-Zhou (2002) regime-switching CIR: 2-state Markov chain governing CIR drift, volatility, and price of risk; estimated by EMM with an SNP auxiliary.
• Dai-Singleton-Yang (2007) preference-free affine regime switching: extends to multi-factor affine regime-switching with preference-free pricing.
• Markov-switching rational expectations (MSRE): forward-looking macro models with regime-dependent drift and policy rules; pricing requires simultaneous solution of the rational-expectations equilibrium and the regime-shift price of risk (Cho-Moreno, Farmer-Waggoner-Zha).
• Compound regime models: multiple independent Markov chains combined into a product chain (e.g. monetary-policy chain × wage-rigidity chain → 4 compound states), used in macro-finance applications such as the CRE asset pricing project.

Comparison

vs. GARCH / stochastic volatility: regime-switching models give discrete, interpretable, persistent shifts; GARCH gives smooth, continuous, less persistent variation. Empirically the two are complementary; Bansal-Zhou (2002) shows that for US Treasury yields, regime switching does much of the work that GARCH cannot do alone.

vs. Time-varying parameter models: regime switching is more parsimonious (constant parameters within K regimes vs. continuously drifting parameters everywhere) and admits cleaner interpretation, at the cost of restricting the variation to discrete jumps.

vs. Single-regime nonlinear models: regime switching gets state-dependent dynamics for free, without requiring custom nonlinear functional forms.

When to use

• When the data show persistent shifts in conditional moments that smooth models cannot capture.
• When you want economic interpretation of the latent state (business cycles, monetary regimes, crisis vs. normal).
• When the no-arbitrage discipline is essential and you can afford to compute regime-shift risk premia.
• When estimation tooling supports regime-switching likelihoods (Hamilton filter, particle filter, Rao-Blackwellized particle filter, or moment-based estimators like EMM).

Known limitations

• Identification of K: the number of regimes is rarely identifiable from the data; standard information criteria are unreliable.
• Constant transition matrix: most applications assume P is fixed, which means the model cannot endogenize regime onset (e.g. the model cannot say “monetary tightening triggers a recession regime”).
• Computational cost: with K regimes and T observations, the exact likelihood requires marginalizing over $K^{\top }$ regime paths. Filters (Hamilton, RBPF) reduce this but introduce approximation error or Monte Carlo noise.
• Approximation in pricing: closed-form bond / asset prices are usually not available across regimes; log-linear or quadratic approximations are typical and accumulate error with horizon.
• Regime path uncertainty: the latent regime is never directly observed; filtered probabilities are conditional on the structural model.

Open problems

• Endogenizing the transition matrix as a function of macro state (state-dependent transition probabilities).
• Joint estimation of the number of regimes K together with the structural parameters.
• Globally accurate non-affine pricing for regime-switching models with K > 2.
• Tractable likelihood evaluation for compound chains with many factor regimes (current state of the art is RBPF).

Key papers

• bansal-zhou-2002-term-structure-regime-shifts — establishes that Markov regime shifts inside a CIR no-arbitrage term-structure model jointly resolve conditional volatility, conditional correlation, and Campbell-Shiller expectations-hypothesis violations on US Treasury yields, with regimes interpretable as NBER business cycles.

My understanding

This is the architectural foundation for the entire CRE asset-pricing project. The choice to model monetary policy and wage rigidity as two independent binary Markov chains (4 compound regimes) is a direct application of this concept, and the Bansal-Zhou (2002) paper is the anchor precedent for the no-arbitrage discipline. Our key methodological contribution on top of the Bansal-Zhou template is the Rao-Blackwellized particle filter (RBPF) for likelihood evaluation: particles carry only the regime path, while the continuous state is integrated analytically (Kalman filter) and the asset prices are exponential-quadratic (Riccati). This sits in the same family as Bansal-Zhou’s simulation-based EMM approach but exploits the conditional Gaussianity of the within-regime model far more aggressively.