Gaussian Martingale Model

• Gaussian Martingale Models are stochastic processes with Gaussian dynamics and martingale properties, pivotal in credit risk and random field applications.
• They utilize explicit constructions (e.g., the Φ-martingale) to calibrate survival curves and ensure analytical tractability in derivative pricing.
• These models enable dynamic Gaussian copulas and precise martingale decompositions, supporting simulation, CDS option pricing, and risk management.

The Gaussian Martingale Model encompasses a family of stochastic processes with martingale properties, governed by Gaussian dynamics and employed in both probability theory and mathematical finance. In credit risk modelling, a prominent instantiation is the $\Phi$-martingale, a conic martingale based on Brownian diffusion with survival processes constrained to $[0,1]$ via the standard normal cumulative distribution. In random fields, martingale-type decompositions provide a robust characterization for Gaussian free fields (GFF) and their fractional analogues. These models serve both as theoretical foundations—embodying generalizations of Brownian motion—and as practical instruments for implementing arbitrage-free, analytically tractable frameworks in finance and probability (Mbaye et al., 2019, Aru et al., 2024).

1. Definition and Structure of Gaussian Martingale Models

A conic martingale is a family of $(\mathbb{Q}, \mathbb{F})$-martingales of the form

$$S_t(T) = F(Z_{t,T}), \quad 0 \leq t \leq T \leq T^*,$$

where $F:\mathbb{R}\to[0,1]$ is a $\mathcal{C}^2$-bijection, and $(Z_{t,T})_{0\leq t\leq T}$ solves a diffusion driven by a Brownian motion $B$, with dynamics

$$dZ_{t,T} = a(t,Z_{t,T})\,dt + n(t,Z_{t,T})\,dB_t, \quad Z_{0,T} = F^{-1}(G(T)),$$

and drift $a$ chosen so $S_t(T)$ is a martingale. For the Gaussian martingale (the $\Phi$-martingale), $F = \Phi$, the standard normal CDF, leading to

$$dZ_{t,T} = n(t)^2 Z_{t,T}\,dt + n(t)\,dB_t,\quad Z_{0,T} = \Phi^{-1}(G(T)),$$

and

$S_t(T) = \Phi(Z_{t,T})\in [0,1].$

This construction ensures $S_0(T) = G(T)$, providing automatic calibration to initial market survival curves (Mbaye et al., 2019).

2. Analytical Properties and Representations

With deterministic $n(t)$, the latent process admits the explicit solution:

$$Z_{t,T} = e^{\frac12\int_0^t n(u)^2 du} \Phi^{-1}(G(T)) + \int_0^t n(s) e^{\frac12\int_s^t n(u)^2 du} dB_s,$$

giving rise to

$$S_t(T) = \Phi\left(e^{\frac12\int_0^t n^2} \Phi^{-1}(G(T)) + \int_0^t n(s)e^{\frac12\int_s^t n^2} dB_s\right).$$

Setting $T=t$, the Azéma supermartingale $S_t := S_t(t) = \Phi(Z_{t,t})$ possesses a Doob–Meyer decomposition

$dS_t = -\alpha_t S_t dt + \sigma_t dB_t,$

where

$$\sigma_t = n(t)\,\varphi(Z_{t,t}), \qquad \alpha_t = e^{\int_0^t n^2} \frac{\varphi\left(\Phi^{-1}(G(t))\right) h(t) G(t)}{S_t},$$

and $h(t) = -G'(t)/G(t)$ is the hazard rate (Mbaye et al., 2019).

3. Dynamized Gaussian Copula: Joint Distributional Analysis

Because each $Z_{t,T}$ is Gaussian and driven by the same Brownian motion, vectors $(Z_{t,T_1}, \ldots, Z_{t,T_k})$ are jointly normal for $T_1 < \cdots < T_k$. Consequently, the joint law of $(S_t(T_1), \ldots, S_t(T_k))$ induces a Gaussian copula with time-varying correlations specified by the kernel

$K(s;t) = n(s) e^{\frac12\int_s^t n(u)^2 du}.$

This structure leads to the "dynamized" Gaussian copula (DGC) property: under partial information, the conditional law of default events is precisely that of a Gaussian copula modulated dynamically by $n(t)$. The $\Phi$-martingale is a DGC model if and only if $\int_0^\infty n^2(u)\,du = \infty$ (Mbaye et al., 2019). Arbitrage-freeness follows from the fact that under progressive filtration enlargement, $(\mathbb{Q}, \mathbb{F})$-martingales remain semimartingales ($\mathbb{G}$-adapted), implying no free lunch.

4. Default Time Construction and Azéma Supermartingales

The explicit construction scheme, per Crépey et al., is given by taking $f(t) = n(t) e^{-\int_0^t n^2}$ and $\ell(t) = -\Phi^{-1}(G(t))$, and defining the default time as

$$\tau = \inf\left\{t \geq 0: \int_0^t f(s) dB_s = \ell(t)\right\}.$$

The associated Azéma supermartingale is then $S_t = \Phi(Z_{t,t}) = \Phi(m_t - \ell(t))$ with $m_t = \int_0^t f(s) dB_s$. This construction provides a closed-form realization of the default time and its corresponding survival process, facilitating exact path simulation and semi-closed-form expressions in derivative pricing (Mbaye et al., 2019).

5. Application to Credit Risk: CVA and CDS Options

Within an intensity or $S$-process framework, Credit Valuation Adjustment (CVA) under wrong-way risk (WWR) is computed as

$$\text{CVA} = \mathbb{E}\left[\int_0^T (1 - R)\, V_u^+\, d(1_{\{\tau > u\}})\right] = (1 - R)\, \mathbb{E}\left[\int_0^T V_u^+\, (-dS_u)\right] = (1 - R)\, \mathbb{E}\left[\int_0^T V_u^+\, \alpha_u\, S_u\, du\right],$$

where $R$ is recovery, $V_u^+$ the discounted positive exposure, and $\alpha_u$ is as in the Azéma decomposition. For options on CDS, the payer price at time zero for strike $k$ and maturities $a, b$ is

$$\text{PSO}(0;a,b,k) = B(0)\, \mathbb{E}\left[S_a\left((1 - R)\int_a^b P_a(u) S_a(u) du - k\, C_a(a,b)\right)^+\right],$$

with annuity $C_a(a,b)$ and discount factor $P_a(u)$. Computations are expedited via the exact simulation of the $Z$ process (Mbaye et al., 2019).

6. Numerical Comparison with SSRJD and TC-JCIR Models

When calibrated to a common survival probability term structure $G(t)$, the $\Phi$-martingale, JCIR++ (SSRJD), and TC-JCIR each produce distinctive WWR and option pricing dynamics:

• Under zero correlation ($\rho=0$) between the driving Brownian motions of credit and exposure, all models coincide.
• As $\rho\rightarrow 1$, the $\Phi$-martingale achieves higher WWR effects, with endpoint CVA at $\rho=0.9$ reaching $\sim1.3\%$ for $n=0.75$, compared to $\sim1.0\%$ under TC-JCIR and $\sim0.9\%$ under JCIR++ (SSRJD).
• CDS option at-the-money implied volatilities: only the $\Phi$-martingale attains market-level Black vols (up to 15–25% for $n=0.2-0.5$ ), while SSRJD and TC-JCIR are constrained below 10% and 8%, respectively.
• The $\Phi$-martingale can exhibit a flexible strike skew in CDS option prices, lacking in the alternative models due to their calibration restrictions (Mbaye et al., 2019).

7. Martingale-Type Characterizations in Gaussian Random Fields

Separately, in the context of spatial stochastic processes, the Gaussian Martingale Model asserts that a field with local mean-zero martingale increments matching the scaling of the Laplacian (or fractional Laplacian) is uniquely a GFF or FGF. The decomposition property (MTD for GFF, FMTD for FGF) posits that, on each ball $B$, the field splits into an (α-)harmonic part and an orthogonal, independent noise. Under uniform moment bounds and zero boundary, this enforces the field’s law is Gaussian with the appropriate covariance structure. Proof employs Poisson resampling dynamics converging to the (fractional) stochastic heat equation, whose unique stationary law is GFF (or FGF). This characterization provides a metric, dynamical perspective analogous to Lévy’s for Brownian motion, enabling new approaches for scaling limits and dynamics preservation in statistical field theory (Aru et al., 2024).