Double Heston/Two-Factor SV Models

• Double Heston models are two-factor stochastic volatility frameworks that model asset dynamics using coupled slow and fast variance processes to capture persistent and transient market features.
• They retain analytical tractability via exponential-affine characteristic functions, facilitating semi-analytic pricing for both vanilla and exotic options.
• Efficient calibration and simulation methods, including Almost-Exact Simulation, reduce pricing errors and enhance risk management compared to single-factor models.

A double Heston or two-factor stochastic volatility (SV) model is a generalization of the original Heston SV framework, introducing two coupled variance processes to better capture the empirically observed features of implied volatility surfaces across different asset classes and maturities. These models retain analytical tractability for vanilla derivatives, can be calibrated to market data efficiently, and are widely used for pricing and risk management of both vanilla and exotic options.

1. Mathematical Structure and Model Definition

The double Heston model comprises a risk-neutral SDE system where the asset price $S_t$ evolves jointly with two independent or weakly correlated, non-negative variance processes, typically specified as Cox–Ingersoll–Ross (CIR) processes. The canonical form is: $$\begin{aligned} dS_t &= r S_t dt + S_t \sqrt{\nu_t^{(1)}}\, dW_t^{(1)} + S_t \sqrt{\nu_t^{(2)}}\, dW_t^{(2)}, \ d\nu_t^{(j)} &= \kappa_j (\bar{\nu}_j - \nu_t^{(j)}) dt + \gamma_{\nu_j} \sqrt{\nu_t^{(j)}}\, dW_t^{(j+2)}, \quad j=1,2, \end{aligned}$$ where $r$ is the risk-free rate; $\kappa_j$ (mean-reversion), $\bar{\nu}_j$ (long-run variance), and $\gamma_{\nu_j}$ (vol-of-vol) are positive parameters; Brownian motions $\{W_t^{(k)}\}_{k=1}^4$ have prescribed instantaneous correlations ($\langle W^{(1)},W^{(3)}\rangle_t = \rho_{1,3}t$, etc.), and Feller conditions ($2\kappa_j\bar{\nu}_j>\gamma_{\nu_j}^2$) ensure positivity of variances. The total instantaneous variance is $\nu_t = \nu_t^{(1)} + \nu_t^{(2)}$ (Dimitrov et al., 22 Dec 2025). Alternative multiscale forms appear in the literature, with a "slow" CIR-like process $$\begin{aligned} dS_t &= r S_t dt + S_t \sqrt{\nu_t^{(1)}}\, dW_t^{(1)} + S_t \sqrt{\nu_t^{(2)}}\, dW_t^{(2)}, \ d\nu_t^{(j)} &= \kappa_j (\bar{\nu}_j - \nu_t^{(j)}) dt + \gamma_{\nu_j} \sqrt{\nu_t^{(j)}}\, dW_t^{(j+2)}, \quad j=1,2, \end{aligned}$$0 and a "fast" mean-reverting process $$\begin{aligned} dS_t &= r S_t dt + S_t \sqrt{\nu_t^{(1)}}\, dW_t^{(1)} + S_t \sqrt{\nu_t^{(2)}}\, dW_t^{(2)}, \ d\nu_t^{(j)} &= \kappa_j (\bar{\nu}_j - \nu_t^{(j)}) dt + \gamma_{\nu_j} \sqrt{\nu_t^{(j)}}\, dW_t^{(j+2)}, \quad j=1,2, \end{aligned}$$1, possibly interacting (e.g., through $$\begin{aligned} dS_t &= r S_t dt + S_t \sqrt{\nu_t^{(1)}}\, dW_t^{(1)} + S_t \sqrt{\nu_t^{(2)}}\, dW_t^{(2)}, \ d\nu_t^{(j)} &= \kappa_j (\bar{\nu}_j - \nu_t^{(j)}) dt + \gamma_{\nu_j} \sqrt{\nu_t^{(j)}}\, dW_t^{(j+2)}, \quad j=1,2, \end{aligned}$$2 modulating instantaneous variance) (Malhotra et al., 2019).

This extension enables the model to superimpose distinct volatility factors—typically a smooth long-term (slow) component and a rougher (fast-reverting) one—capturing both persistent term-structure and transient features of observed implied volatility skews.

2. Analytical Tractability and Pricing Methodologies

Double Heston models are affine, such that the characteristic function of log-price $$\begin{aligned} dS_t &= r S_t dt + S_t \sqrt{\nu_t^{(1)}}\, dW_t^{(1)} + S_t \sqrt{\nu_t^{(2)}}\, dW_t^{(2)}, \ d\nu_t^{(j)} &= \kappa_j (\bar{\nu}_j - \nu_t^{(j)}) dt + \gamma_{\nu_j} \sqrt{\nu_t^{(j)}}\, dW_t^{(j+2)}, \quad j=1,2, \end{aligned}$$3 is exponential-affine in initial values of the stochastic variances, enabling semi-analytic pricing of European-style derivatives: $$\begin{aligned} dS_t &= r S_t dt + S_t \sqrt{\nu_t^{(1)}}\, dW_t^{(1)} + S_t \sqrt{\nu_t^{(2)}}\, dW_t^{(2)}, \ d\nu_t^{(j)} &= \kappa_j (\bar{\nu}_j - \nu_t^{(j)}) dt + \gamma_{\nu_j} \sqrt{\nu_t^{(j)}}\, dW_t^{(j+2)}, \quad j=1,2, \end{aligned}$$4 where $$\begin{aligned} dS_t &= r S_t dt + S_t \sqrt{\nu_t^{(1)}}\, dW_t^{(1)} + S_t \sqrt{\nu_t^{(2)}}\, dW_t^{(2)}, \ d\nu_t^{(j)} &= \kappa_j (\bar{\nu}_j - \nu_t^{(j)}) dt + \gamma_{\nu_j} \sqrt{\nu_t^{(j)}}\, dW_t^{(j+2)}, \quad j=1,2, \end{aligned}$$5 solve coupled Riccati differential equations parameterized by the correlations, mean-reversions, vol-of-vols, and allowing closed-form expression for vanilla options via inverse Fourier transforms (e.g., Carr–Madan, Lewis, or P1/P2 formulas). This property is retained—even when the spot/variance instantaneous correlation $$\begin{aligned} dS_t &= r S_t dt + S_t \sqrt{\nu_t^{(1)}}\, dW_t^{(1)} + S_t \sqrt{\nu_t^{(2)}}\, dW_t^{(2)}, \ d\nu_t^{(j)} &= \kappa_j (\bar{\nu}_j - \nu_t^{(j)}) dt + \gamma_{\nu_j} \sqrt{\nu_t^{(j)}}\, dW_t^{(j+2)}, \quad j=1,2, \end{aligned}$$6 is made stochastic through suitable parameterizations of the two variance factors (Higgins, 1 Feb 2026).

For path-dependent and early-exercise options (Bermudan, American), Monte Carlo approaches are necessary. The Almost-Exact Simulation (AES) scheme leverages the exact transition law of CIR processes (noncentral $$\begin{aligned} dS_t &= r S_t dt + S_t \sqrt{\nu_t^{(1)}}\, dW_t^{(1)} + S_t \sqrt{\nu_t^{(2)}}\, dW_t^{(2)}, \ d\nu_t^{(j)} &= \kappa_j (\bar{\nu}_j - \nu_t^{(j)}) dt + \gamma_{\nu_j} \sqrt{\nu_t^{(j)}}\, dW_t^{(j+2)}, \quad j=1,2, \end{aligned}$$7 distributions) for the variances and efficiently updates the log-price by treating certain stochastic integrals exactly, incurring only a small Euler-type bias (Dimitrov et al., 22 Dec 2025). This leads to efficient and accurate Monte Carlo pricing, including for high-dimensional and multi-factor setups.

3. Empirical Performance, Calibration, and Volatility Surface Fit

Double Heston models outperform single-factor SV and jump-diffusion models in fitting implied volatility surfaces, especially for assets where both short- and long-time market smiles must be captured simultaneously. Empirical calibration on SPX options demonstrates a marked reduction in mean relative pricing error across maturities. As reported:

• Mean relative error (MRE) for single-factor Heston: 6.97% (30d), 8.74% (90d), 2.84% (180d).
• MRE for two-factor model: 2.25% (30d), 4.56% (90d), 3.80% (180d) (Malhotra et al., 2019).

The multiscale (two-factor) framework enables this by locally roughening the variance process through a fast mean-reverting factor, accentuating the short-maturity skew and better matching the wings of the smile, while preserving the smoothness and term structure controlled by the slow factor (Malhotra et al., 2019). The addition of stochastic spot/volatility correlation—modeling the empirically observed "risk-reversal beta" in FX markets—further improves fit to observed exotic derivatives prices, such as knock-outs and one-touches (Higgins, 1 Feb 2026).

Calibration methodology generally combines global search (e.g., differential evolution over a coarse grid) with local solvers (Levenberg–Marquardt or BFGS), evaluating option prices via Fourier inversion or fast characteristic function methods. Key parameters separating fast/slow scales (mean reversion, vol-of-vol, fast factor intensity) are identifiable and interpretable in this framework.

4. Statistical Properties and Drift Estimation

Under subcritical ergodic regimes (i.e., positive mean-reversion coefficients and appropriate moment conditions), the double Heston process admits a unique stationary law, with exponential ergodicity verifiable via Foster–Lyapunov drift criteria and affine transform techniques. Explicit forms for the joint Laplace–Fourier transforms and the limiting distribution exist via associated Riccati ODEs (Alaya et al., 28 Jan 2025).

Drift parameter estimation can be performed via continuous-time maximum likelihood (MLE) or conditional least squares (CLS) estimators. For sufficiently regular sample paths:

• The MLE takes the form:

$$\begin{aligned} dS_t &= r S_t dt + S_t \sqrt{\nu_t^{(1)}}\, dW_t^{(1)} + S_t \sqrt{\nu_t^{(2)}}\, dW_t^{(2)}, \ d\nu_t^{(j)} &= \kappa_j (\bar{\nu}_j - \nu_t^{(j)}) dt + \gamma_{\nu_j} \sqrt{\nu_t^{(j)}}\, dW_t^{(j+2)}, \quad j=1,2, \end{aligned}$$8

where $$\begin{aligned} dS_t &= r S_t dt + S_t \sqrt{\nu_t^{(1)}}\, dW_t^{(1)} + S_t \sqrt{\nu_t^{(2)}}\, dW_t^{(2)}, \ d\nu_t^{(j)} &= \kappa_j (\bar{\nu}_j - \nu_t^{(j)}) dt + \gamma_{\nu_j} \sqrt{\nu_t^{(j)}}\, dW_t^{(j+2)}, \quad j=1,2, \end{aligned}$$9 are the drift regressors and $r$0 the diffusion matrix.

• Both MLE and CLS estimators are strongly consistent and asymptotically normal under exponential ergodicity and standard moment conditions:

$r$1

where $r$2 is the Fisher information-type matrix (Alaya et al., 28 Jan 2025). This suggests robust statistical inference for risk and pricing applications given sufficiently long time series of underlying and variance proxies.

5. Computational Schemes and Efficiency

The AES scheme for the double Heston model yields bias-free simulation of the variance factors, and a log-price update with error negligible in practical regimes. Compared to Euler–Maruyama discretizations:

• For American put options under double Heston, AES with $r$3 time steps and $r$4 paths yields $r$5 pricing error versus reference solutions; the corresponding Euler scheme error exceeds $r$6.
• AES achieves 15–25% reduction in CPU time compared to Euler schemes with the step count required for similar accuracy, as the variance process is integrated exactly and only a minor bias arises from the price process update (Dimitrov et al., 22 Dec 2025).
• For Bermudan options, relative errors are consistently 0.1–0.4% for modest step numbers, while Euler requires twice as many grid points to match.

For exotics (e.g., knock-outs and volatility swaps), Monte Carlo with pathwise variance simulation, Brownian-bridge crossing tests, and control variates is standard (Higgins, 1 Feb 2026).

6. Application Domains and Model Impact

Double Heston models are applied extensively in option pricing and risk management, wherever accurate modeling of implied volatility smiles and term structures is critical (e.g., equity index options, FX barriers, volatility swaps). The stochastic spot/volatility correlation extension enables modeling of risk-reversal dynamics (e.g., in FX), affecting both the valuation and hedging of barrier derivatives. Model risk is material: underpricing exotic payoffs under single-factor SV models compared to two-factor with stochastic correlation frequently exceeds interdealer bid-ask, implying practical calibration and pricing risk (Higgins, 1 Feb 2026).

The computational tractability afforded by the affine structure and efficient simulation schemes makes double Heston models feasible for calibration and production-level pricing of early-exercise and path-dependent instruments (Dimitrov et al., 22 Dec 2025).

7. Extensions, Limitations, and Future Directions

Current research continues to address limitations such as parameter identifiability across maturities and strikes, and full-surface calibration (possibly via time-dependent or piecewise-constant model coefficients). Affine extensions incorporating jumps, local volatility, or rough volatility processes are being studied to further improve fit and flexibility. Ongoing empirical comparisons to traded prices for higher-order exotics guide further refinement (Higgins, 1 Feb 2026). A plausible implication is that two-factor models, through their rich structure and advanced statistical properties, will remain foundational in volatility modeling for both theoretical exploration and practical derivative pricing.