Jump Diffusion Model

• Jump Diffusion Model is a stochastic process that combines continuous diffusion with sudden Poisson or Lévy jumps to capture abrupt changes in asset prices.
• It decomposes the dynamics into local (diffusion) and nonlocal (jump) components, resulting in partial integro-differential equations that facilitate closed-form and moment-based solutions.
• Applications span asset pricing, risk management, and stochastic control, leveraging advanced numerical and statistical methods to accurately estimate model parameters.

A jump diffusion model is a class of stochastic processes that generalizes standard diffusion (Brownian) dynamics by superimposing discontinuous jumps, typically generated by Poisson or Lévy processes, onto otherwise continuous paths. Such models are foundational across mathematical finance, stochastic control, and risk-sensitive portfolio theory, allowing for realistic representation of sudden, large-magnitude events observed empirically in asset returns, economic indicators, and engineered systems.

1. Canonical Jump Diffusion Dynamics

The archetypal jump diffusion process for an asset price $S_t$ is governed—in the risk-neutral measure—by the stochastic differential equation (SDE): $$dS_t = S_{t^-}[(r - d - \lambda\bar{j})\,dt + \sigma\,dW_t + J\,dN_t],$$ where $r$ is the risk-free rate, $d$ is the dividend yield, $\sigma$ is diffusion volatility, $W_t$ is a standard Brownian motion, $N_t$ is a Poisson process with intensity $\lambda$, $J$ is the random jump size (often lognormal), and $\bar{j} = \mathbb{E}[J]$ is the compensator ensuring the martingale property. The explicit solution reads: $$S_t = S_0 \exp\Big[(r-d-\lambda \bar{j} - \frac{1}{2}\sigma^2)t + \sigma W_t\Big] \prod_{i=1}^{N(t)} (1 + J_i),$$ where $\{J_i\}$ are i.i.d. jump sizes (Gaudenzi et al., 2017).

Extensions cover stochastic volatility, regime-switching, multidimensional factor jumps, and the use of more general jump distributions, including Lévy and double-exponential laws (Liang et al., 2010, Budhathoki et al., 20 Nov 2025, Filipović et al., 2017).

2. Mathematical Representation and Infinitesimal Generator

For a general $d$-dimensional process $X_t$, the infinitesimal generator $\mathcal{G}$ acting on $C^2$ test functions $f : \mathbb{R}^d \to \mathbb{R}$ is: $$\mathcal{G}f(x) = \frac{1}{2}\operatorname{tr}(a(x)\nabla^2 f(x)) + b(x)^\top \nabla f(x) + \int_{\mathbb{R}^d} \left[f(x+\xi) - f(x) - \xi^\top \nabla f(x)\right] \nu(x, d\xi),$$ where $a(x)$ is the diffusion matrix, $b(x)$ is the drift, and $\nu(x, d\xi)$ is the jump kernel (Filipović et al., 2017). This generator defines the partial integro-differential equations (PIDEs) that arise as generalized Kolmogorov or Hamilton–Jacobi–Bellman (HJB) equations for control and pricing problems (Davis et al., 2011, Patel et al., 2018).

The key object is the decomposition into local (continuous) and nonlocal (jump) operators, making jump-diffusion models fundamentally nonlocal and infinitely dimensional even in finite state spaces.

3. Distinct Model Classes: Affine, Polynomial, and Lévy-Based Variants

Affine Jump Diffusions

Affine jump diffusion models satisfy

$$\mathcal{G}(e^{u^\top x}) = [F(u) + R(u)^\top x] e^{u^\top x},$$

with $a(x)$, $b(x)$, and $\nu(x, d\xi)$ linear in $x$. They admit closed-form solutions for characteristic functions via Riccati equations and subsume classical models like Merton and Heston with jumps (Filipović et al., 2017).

Polynomial Jump Diffusions

Polynomial jump diffusions extend affine models by allowing the generator to map polynomials of degree $n$ to polynomials of at most degree $n$ (possibly higher in certain jump moments): $$b \in \operatorname{Pol}_1(E), \quad a + \int \xi\xi^\top \nu(\cdot, d\xi) \in \operatorname{Pol}_2(E), \quad \int \xi^\alpha \nu(\cdot, d\xi) \in \operatorname{Pol}_{|\alpha|}(E), \ |\alpha|\geq3,$$ thus supporting broader classes of tractable nonlinearity and closed-form computation of moments (Filipović et al., 2017). Polynomial jump diffusions are preserved under polynomial transformations and under time changes by independent Lévy subordinators.

Non-Affine/State-Dependent Extensions

Recent advances include jump intensities and magnitudes directly modulated by contemporaneous diffusion movements (diffusion-dependent jumps), resulting in explicit no-arbitrage drift conditions under Girsanov/Esscher transforms (Virk et al., 17 Dec 2025).

4. Numerical and Option Pricing Methodologies

Jump-diffusion PIDEs admit a spectrum of numerical methods:

• Infinite series (Merton): European option prices as Poisson mixtures of Black–Scholes terms, requiring infinite summation for high jump intensities (Gaudenzi et al., 2017, Lau et al., 2017).
• Bivariate/Tree methods: Lattice approximations (e.g., seven-node trees), leveraging pruning strategies to reduce complexity to $O(n\log n)$ (European) and $O(n^2\log n)$ (American) with rigorous truncation error control (Gaudenzi et al., 2017).
• Finite-difference/FFT methods: High-order compact schemes employing Crank–Nicolson leap-frog time discretization, Pade-based spatial stencils, and FFT-accelerated treatment of the jump integral (Patel et al., 2018).
• Monte Carlo and Path Integrals: Applied for complex payoffs or stochastic volatility, sometimes in conjunction with Fourier inversion techniques (Liang et al., 2010).
• Moment Expansions and Polynomial Methods: Systematic expansion in polynomials of state and jump variables, enabling closed-form moments for a wide class of payoffs and facilitating efficient option pricing through orthogonal polynomial bases and moment-matching (Filipović et al., 2017).

5. Advanced Econometric and Statistical Inference

Parameter estimation in jump-diffusion frameworks relies on methodologies such as:

• EM algorithm for hidden Markov and regime-modulated models: Handling latent regime-switches and unobserved jumps with Laplace or Gaussian jump-size laws, as in the Markov-modulated jump-diffusion and semi-Markov extensions (Eslava et al., 2022, Goswami et al., 2018).
• Bayesian approaches (Gibbs samplers): Simultaneous inference on drift, volatility, jump-intensity, and jump-size parameters; these approaches provide robust identification of jump features in returns data (Lau et al., 2017).
• Deterministic Nonlinear Filtering (DNF): Grid-based (Kitagawa-type) state space filters for high-dimensional, non-Gaussian models with stochastic jump intensities, outperforming particle methods for likelihood evaluation (Bégin et al., 2019).
• Kramers–Moyal moment regression: Adaptive, zone-binned extraction of conditional moments up to sixth order, facilitating model discrimination between diffusion and jump-diffusion regimes in empirical time series (Budhathoki et al., 20 Nov 2025).

6. Applications and Model Extensions

The jump diffusion framework underpins quantitative approaches in:

• Asset Pricing and Options: Realistic option valuation, volatility surface calibration, and risk assessment under heavy-tailed and asymmetric innovations (Davis et al., 2011, Liang et al., 2010).
• Risk-Sensitive and Constrained Control: Feedback controls under stochastic volatility, nonlocal risk, and jump-driven factor models, with existence and uniqueness guaranteed by viscosity and PDE comparison techniques (Davis et al., 2011).
• Regime-Switching and Semi-Markov Models: Integrating explicit memory, regime duration, and nonlocal jumps in the joint representation of asset price, interest rate, and volatility evolution (Goswami et al., 2018, Eslava et al., 2022).
• Stochastic Volatility with Jump Intensity Clustering: Modeling crises and volatility spikes via stochastic Cox-process-driven jump rates, with deterministic filtering for empirical estimation (Bégin et al., 2019).
• Emission Market Modeling: Capturing information shocks and abrupt price changes in emission allowance certificates, with nonlinear PIDEs for optimal pricing and hedging (Borovkov et al., 2010).

7. Empirical Regularities, Model Selection, and Controversies

Jump-diffusion models are empirically validated via improved fit to fat tails, volatility clustering, and jump clustering observed in financial returns (Gaudenzi et al., 2017, Budhathoki et al., 20 Nov 2025). Statistical model selection hinges on metrics such as the ratio of fourth to second Kramers–Moyal coefficients ($R = D^{(4)}/D^{(2)}$), with $R \ll 1$ diagnostically pure diffusion and $R \gg 1$ signifying jumps (Budhathoki et al., 20 Nov 2025). In regime-switching settings, hands-on thresholding is required for jump detection in discrete-time data (Eslava et al., 2022). A plausible implication based on recent literature is that proper modeling of diffusion–jump dependence is critical for arbitrage-free pricing and robust risk management, especially in environments with feedback between volatility and jump events (Virk et al., 17 Dec 2025).

---

Jump-diffusion models remain the principal paradigm for modeling rare, discontinuous shocks in continuous-time stochastic settings, bridging classical Gaussian frameworks and general Lévy processes. Their tractability, structural robustness under transformations, and flexibility in empirical calibration drive ongoing methodological and applied research across mathematical finance, econometrics, and stochastic control.