Regime-Switching Jump-Diffusion Models

Updated 26 March 2026

RSJD models are hybrid stochastic systems combining continuous diffusion with abrupt jumps, modulated by a regime-switching process.
They provide a robust framework for analyzing volatility clustering, structural changes, and risk dynamics in finance, insurance, and engineering.
Advanced inference and numerical methods enable precise parameter estimation, option pricing, and stability analysis under regime-dependent conditions.

A regime-switching jump-diffusion (RSJD) model is a hybrid stochastic process in which a continuous-time diffusion with jumps is modulated by an underlying regime process, typically a finite- or countable-state Markov or semi-Markov chain. The resulting process captures intermittent structural changes in system dynamics and accommodates both diffusive variability and abrupt spikes. RSJD models have become foundational across stochastic control, financial mathematics, insurance, and engineering due to their ability to encode random jumps, volatility clustering, and regime-dependent transition structures.

1. Mathematical Formulation

Formally, an RSJD system consists of two processes: a continuous state variable $X(t) \in \mathbb{R}^d$ and a discrete environment or regime process $\Lambda(t)$ taking values in a (finite or countable) set $S$ . The joint evolution is governed by a Markovian or semi-Markovian switching mechanism. The stochastic differential equation is

$dX(t) = b(X(t), \Lambda(t))\,dt + \sigma(X(t), \Lambda(t))\,dW(t) + \int_U c(X(t^-), \Lambda(t^-), u)\,\tilde N(dt, du),$

where $b$ is regime-dependent drift, $\sigma$ is regime-dependent diffusion, $N(dt, du)$ is a Poisson random measure, and the compensated measure is $\tilde N(dt, du)$ . The regime process $\Lambda(t)$ may be a continuous-time Markov chain with intensity matrix $Q(x) = (q_{ij}(x))$ , possibly dependent on the current state $X(t)$ (Xi et al., 2018, Zlotchevski et al., 8 Nov 2025, Xi et al., 2017, 2002.01422).

Extensions include semi-Markov modulated coefficients and jump measures, randomization of regime parameters, and random or deterministic switching times (Wolf et al., 2024, Das et al., 2016). The jump mechanism can be of Merton type with Gaussian increments, Kou type with double-exponential jumps, or more general Lévy kernels (Zlotchevski et al., 8 Nov 2025).

2. Stochastic Analysis: Existence, Uniqueness, and Stability

Existence and Uniqueness

Under appropriate Lyapunov-type growth control and local regularity conditions on $b, \sigma, c$ , there exists a unique strong solution to the coupled RSJD SDE (Xi et al., 2018, 2002.01422, Xi et al., 2017). For non-Lipschitz coefficients, tractability relies on a concave modulus $p$ and superlinear Lyapunov weight $\varphi$ , ensuring control of explosion and pathwise uniqueness even with superlinear growth (e.g., Cox–Ingersoll–Ross or branching diffusions as examples).

Feller and Strong Feller Properties

With local continuity and ellipticity, the RSJD semigroup $P_t f(x, k) = \mathbb{E}_{x, k}[f(X(t), \Lambda(t))]$ is Feller, enabling continuous dependence on initial values. The strong Feller property, under additional local smoothness and ellipticity, implies transition semigroups regularize bounded Borel functions to continuous functions (Xi et al., 2018, 2002.01422, Xi et al., 2017). This is crucial for the uniqueness of invariant measures and ergodic analysis.

Almost Sure and Moment Exponential Stability

Stability results are established using Lyapunov functions $V(x, i)$ . For linear RSJD,

$dx(t) = a_{\Lambda(t)} x(t) dt + b_{\Lambda(t)} x(t) dW(t) + \int c_{\Lambda(t^-)}(z) x(t^-) \tilde N(dt, dz),$

the necessary and sufficient condition for almost sure exponential stability is

$\sum_{i} \pi_i \left[ a_i - \frac{1}{2} b_i^2 + \int (\ln|1 + c_i(z)| - c_i(z)) \nu(dz) \right] < 0,$

where $\pi$ is the stationary law of the regime process (Chao et al., 2017, Yang et al., 2014).

3. Infinitesimal Generator and Ergodic Properties

The extended generator for RSJD acting on $f \in C^2_b(\mathbb{R}^d \times S)$ is

$\mathcal{A} f(x, i) = b(x, i) \cdot \nabla_x f(x, i) + \frac{1}{2} \mathrm{Tr}[\sigma \sigma^T(x, i) D^2_x f(x, i)] + \int_U \bigl[ f(x + c(x, i, u), i) - f(x, i) \bigr] \nu(du) + \sum_{j \neq i} q_{ij}(x) [f(x, j) - f(x, i)].$

Ergodicity (existence of a unique invariant measure and convergence to stationarity) is typically established via Foster–Lyapunov drift criteria, irreducibility (all regimes communicate), and strong Feller property (Chen et al., 2018, Xi et al., 2017, 2002.01422). For systems with countably many regimes, Lyapunov control in both continuous and discrete components is essential.

4. Control, Games, and Mean-Field Theory

Mean-Field Control and Master Equations

In large-population systems, agents’ RSJD states interact through the empirical measure. The mean-field controlled McKean–Vlasov RSJD,

$dX_t = b(t, X_t, \mu_t, \alpha_t, \Lambda_{t-})\,dt + \sigma(t, X_t, \Lambda_{t-})\,dW_t + \int_E \gamma(t, X_{t-}, \Lambda_{t-}, e)\,\tilde N(dt, de),$

leads to a master equation for the value function $V(t, \mu, i)$ on the space of probability measures,

$-\partial_t V + H(\mu, i, D_m V, D^2_{mm} V) + \sum_j q_{ij} V(t, \mu, j) = 0, \quad V(T, \mu, i) = \int h(T, x, \mu, i) \mu(dx).$

Finite-agent systems converge to the mean-field solution with explicit rates, and controlled propagation of chaos applies to optimal trajectories (Bayraktar et al., 2021).

Risk-Sensitive and Zero-Sum Linear-Quadratic Games

RSJD models with random coefficients and cone constraints admit explicit solutions in quadratic control and game problems. Optimal feedback is constructed via BSDEs or Riccati equations with jumps, with the closed-loop law characterized as a solution to backward stochastic Riccati equations on each regime (Shi et al., 2024, Tang et al., 8 Mar 2026). In zero-sum games, existence and explicit feedback for the saddle point are established via a combination of FBSDEs and completion-of-the-square (Tang et al., 8 Mar 2026).

Numerical Methods

Euler–Maruyama discretization schemes are analyzed for exponential stability, and Markov chain approximation (MCAM) methods are developed for integro-differential HJB systems arising in RSJD control and games. The MCAM retains local consistency, ensures convergence in value functions, and scalability for high-dimensional problems (Yang et al., 2014, Bui et al., 2018).

5. Inference, Estimation, and Applications

Parameter Estimation and State Inference

Bayesian inference and MCMC, often coupled with Hamilton filtering, are widely used for parameter estimation in RSJD with latent regime sequences. Efficient jump detection and volatility estimation can be achieved via maximal-threshold methods and surrogate data testing, allowing discrimination between multiple noise and regime sources (Das et al., 2019, Persio et al., 2016). Regime-duration distributions, jump intensities, and regime-specific volatilities are consistently estimated even under non-Lipschitz or semi-Markov switching (Das et al., 2019, Das et al., 2016).

Option Pricing

RSJD processes provide tractable yet flexible asset pricing models, especially for European and forward-starting derivatives. Option pricing relies on explicitly computable regime-dependent characteristic functions: $\varphi(u;T) = \bm p \exp\left[(Q - A(u))T\right] \bm 1,$ where $A(u)$ contains the characteristic exponents for jump and diffusion in each regime (Wolf et al., 2024, Ramponi, 2011). Fourier transform approaches (Carr–Madan, COS, FFT) and PDE/integro-partial differential equation methods accommodate regime-dependency in drift, volatility, jump frequency, and distribution (Goswami et al., 2018, Ramponi, 2011). Local volatility representations for regime mixtures facilitate calibration to market implied volatility surfaces (Wolf et al., 2024).

Applications

RSJD models underpin regime-dependent risk management, asset allocation, optimal reinsurance/investment games, and system control under uncertainty (Andruszkiewicz et al., 2014, Bui et al., 2018, Das et al., 2016). They capture tail risk, volatility clustering, and "business cycle"-style periodicity via periodic solutions, and are suited for inference in high-frequency data and financial time series modeling (Guo et al., 2019, Persio et al., 2016, Das et al., 2019).

6. Advanced Topics and Open Directions

Schrödinger Bridges and Large Deviation Theory

The Schrödinger bridge problem over RSJD path space leads to new regimes of transport and entropy minimization, with analytic and stochastic calculus formulations for path measures featured in the RSJD class (Zlotchevski et al., 8 Nov 2025). The minimal entropy martingale measure (MEMM), characterized in regime-switching contexts, generally differs from the Esscher transform, requiring explicit adjustment of regime-modulated jump intensities and Brownian drift (Crescenzo et al., 2015).

Non-Lipschitz, Countable Regimes, and Ergodicity

Expansions to countably infinite regime spaces are accompanied by technical difficulties such as loss of compactness and more complex pathwise Lyapunov and coupling arguments. Exponential and polynomial ergodicity are analyzed using Lyapunov drift, irreducibility, and strong Feller properties, ensuring unique invariant measures and mixing properties (Chen et al., 2018, Xi et al., 2017, 2002.01422, Xi et al., 2018).

Mean Field, Propagation of Chaos, and Numerical Analysis

Propagation of chaos phenomena hold with explicit error rates for optimally controlled RSJD systems as the agent count increases, underlying the scalability of mean-field control (Bayraktar et al., 2021). Legitimate and stable numerical solutions require new analysis as Poisson jumps invalidate high-order expansions; stochastic approximation and MCAM emerge as robust discretization schemes (Yang et al., 2014, Bui et al., 2018).