4-Factor Markov PDV Model

• The model is a finite-dimensional Markov reformulation that replaces full-path integrals with four latent factors representing short and long memory effects.
• It captures both trend (leverage) and activity features by employing two return-memory and two squared-volatility-memory factors driven by exponential kernels.
• Practical implementations using neural surrogates demonstrate efficient calibration for SPX/VIX smiles despite challenges in enforcing global volatility positivity.

The 4-factor Markov path-dependent volatility (PDV) model is a finite-dimensional Markov reformulation of the Guyon–Lekeufack path-dependent volatility specification in which instantaneous volatility is driven by two path features: a weighted average of past returns and a square-root weighted average of past squared volatility. In the Markovian version, each feature is represented by a short-memory and a long-memory exponential kernel, yielding two return-memory factors and two squared-volatility-memory factors, hence a coupled four-dimensional stochastic system driven by one Brownian motion (Nutz et al., 2023). In later applied work, the same architecture is used as a time-homogeneous, low-parametric model for joint SPX and VIX smile fitting, sometimes with an additional upside-volatility parameter that increases volatility when the trend factor is positive (Gazzani et al., 2024).

1. Origins and modeling objective

The model originates in the observation that equity-index volatility is strongly path-dependent. The central empirical ingredients are a trend or leverage feature, according to which recent negative returns tend to raise future volatility, and an activity feature, according to which large squared returns tend to be followed by large future volatility. In the original PDV formulation, these effects are encoded by kernels acting on the full past trajectory of returns and squared returns; the Markovian 4-factor version replaces those path integrals by a finite set of exponentially weighted state variables (Nutz et al., 2023).

This construction should be distinguished from models that are Markov in the asset price alone. The 4-factor PDV model is non-Markovian in the raw price history, but Markovian in an augmented state containing the four latent memory factors. In that sense, “4-factor” refers to the dimension of the latent memory state, not to four independent noise sources. The same viewpoint appears in later computational work, where the model is treated as a finite-dimensional state-space representation of path dependence and is used precisely because the Markov lift makes simulation, VIX learning, and calibration tractable (Baschetti et al., 12 Jul 2025).

The broader empirical motivation also extends beyond spot volatility. A related study on implied-volatility forecasting reports that a large part of the movements of the at-the-money-forward implied volatility for maturities from 1 to 24 months is explained by past returns and past squared returns, that the explanatory power decreases with maturity, and that up to roughly 4 years of past underlying evolution matter, especially through squared returns (Andrès et al., 2023). This situates the 4-factor Markov PDV model within a wider program in which volatility and implied-volatility dynamics are both treated as path-dependent.

2. Canonical state variables and stochastic dynamics

In the Markovian 4-factor specification studied by Nutz and Valdevenito, the instantaneous volatility is

$\sigma_t = B_0 + B_1 R_{1,t} + B_2 \sqrt{R_{2,t}},$

with aggregated memory variables

$$R_{1,t}=(1-\theta_1)R_{1,0,t}+\theta_1 R_{1,1,t}, \qquad R_{2,t}=(1-\theta_2)R_{2,0,t}+\theta_2 R_{2,1,t}.$$

The four latent factors satisfy

$$dR_{1,j,t}=\lambda_{1,j}\sigma_t\,dW_t-\lambda_{1,j}R_{1,j,t}\,dt,\qquad dR_{2,j,t}=\lambda_{2,j}(\sigma_t^2-R_{2,j,t})\,dt,\qquad j\in\{0,1\}.$$

Here $R_{1,j,t}$ are the return-memory factors and $R_{2,j,t}$ are the squared-volatility-memory factors. The parameter restrictions assumed in the wellposedness analysis are

$$B_0,\;B_2,\;\lambda_{1,j},\;\lambda_{2,j}\ge 0,\qquad B_1<0,\qquad \theta_1,\theta_2\in[0,1],$$

with initial conditions $R_{1,j,0}\in\mathbb{R}$ and $R_{2,j,0}>0$ (Nutz et al., 2023).

The four factors admit a direct interpretation. Each $R_{1,j,t}$ is an Ornstein–Uhlenbeck type factor driven by the volatility-weighted Brownian return $\sigma_t\,dW_t$, while each $$R_{1,t}=(1-\theta_1)R_{1,0,t}+\theta_1 R_{1,1,t}, \qquad R_{2,t}=(1-\theta_2)R_{2,0,t}+\theta_2 R_{2,1,t}.$$0 is an exponential moving average of the squared volatility $$R_{1,t}=(1-\theta_1)R_{1,0,t}+\theta_1 R_{1,1,t}, \qquad R_{2,t}=(1-\theta_2)R_{2,0,t}+\theta_2 R_{2,1,t}.$$1. The mixing parameters $$R_{1,t}=(1-\theta_1)R_{1,0,t}+\theta_1 R_{1,1,t}, \qquad R_{2,t}=(1-\theta_2)R_{2,0,t}+\theta_2 R_{2,1,t}.$$2 determine how strongly the model loads on the fast and slow memory scales.

In pricing applications, the same structure is often written with $$R_{1,t}=(1-\theta_1)R_{1,0,t}+\theta_1 R_{1,1,t}, \qquad R_{2,t}=(1-\theta_2)R_{2,0,t}+\theta_2 R_{2,1,t}.$$3-notation and embedded in a risk-neutral asset equation. One widely used version is

$$R_{1,t}=(1-\theta_1)R_{1,0,t}+\theta_1 R_{1,1,t}, \qquad R_{2,t}=(1-\theta_2)R_{2,0,t}+\theta_2 R_{2,1,t}.$$4

with

$$R_{1,t}=(1-\theta_1)R_{1,0,t}+\theta_1 R_{1,1,t}, \qquad R_{2,t}=(1-\theta_2)R_{2,0,t}+\theta_2 R_{2,1,t}.$$5

where $$R_{1,t}=(1-\theta_1)R_{1,0,t}+\theta_1 R_{1,1,t}, \qquad R_{2,t}=(1-\theta_2)R_{2,0,t}+\theta_2 R_{2,1,t}.$$6 is an additional parameter that “unlocks enough volatility on the upside to reproduce the implied volatility smiles of S{data}P 500 and VIX options” (Gazzani et al., 2024). When $$R_{1,t}=(1-\theta_1)R_{1,0,t}+\theta_1 R_{1,1,t}, \qquad R_{2,t}=(1-\theta_2)R_{2,0,t}+\theta_2 R_{2,1,t}.$$7, this reduces to the original linear-plus-square-root PDV form.

3. Exponential kernels and the Markov lift

The model is “path-dependent” because its basic observables are weighted functionals of the past return path. In the more general continuous-time PDV formulation, the two path features are

$$R_{1,t}=(1-\theta_1)R_{1,0,t}+\theta_1 R_{1,1,t}, \qquad R_{2,t}=(1-\theta_2)R_{2,0,t}+\theta_2 R_{2,1,t}.$$8

and the spot volatility is $$R_{1,t}=(1-\theta_1)R_{1,0,t}+\theta_1 R_{1,1,t}, \qquad R_{2,t}=(1-\theta_2)R_{2,0,t}+\theta_2 R_{2,1,t}.$$9 (Andrès et al., 2024). This formulation is genuinely non-Markovian for general kernels $$dR_{1,j,t}=\lambda_{1,j}\sigma_t\,dW_t-\lambda_{1,j}R_{1,j,t}\,dt,\qquad dR_{2,j,t}=\lambda_{2,j}(\sigma_t^2-R_{2,j,t})\,dt,\qquad j\in\{0,1\}.$$0.

The 4-factor Markov model arises when each kernel is approximated or specified as a convex combination of two exponentials. In the calibration-oriented notation,

$$dR_{1,j,t}=\lambda_{1,j}\sigma_t\,dW_t-\lambda_{1,j}R_{1,j,t}\,dt,\qquad dR_{2,j,t}=\lambda_{2,j}(\sigma_t^2-R_{2,j,t})\,dt,\qquad j\in\{0,1\}.$$1

so that each of the two path features decomposes into a short-memory and a long-memory component (Gazzani et al., 2024). In the more general kernel-analysis paper, the same idea is described as the recovery of the 4-factor Markovian PDV model when both kernels are convex combinations of two exponentials; then $$dR_{1,j,t}=\lambda_{1,j}\sigma_t\,dW_t-\lambda_{1,j}R_{1,j,t}\,dt,\qquad dR_{2,j,t}=\lambda_{2,j}(\sigma_t^2-R_{2,j,t})\,dt,\qquad j\in\{0,1\}.$$2 contributes two Markov factors and $$dR_{1,j,t}=\lambda_{1,j}\sigma_t\,dW_t-\lambda_{1,j}R_{1,j,t}\,dt,\qquad dR_{2,j,t}=\lambda_{2,j}(\sigma_t^2-R_{2,j,t})\,dt,\qquad j\in\{0,1\}.$$3 contributes two Markov factors (Andrès et al., 2024).

This Markovization is the tractability mechanism. The original non-Markovian model uses integrals over the entire past with kernel weights, whereas exponential kernels allow the convolutions to be rewritten as finite-dimensional SDEs. The resulting state vector

$$dR_{1,j,t}=\lambda_{1,j}\sigma_t\,dW_t-\lambda_{1,j}R_{1,j,t}\,dt,\qquad dR_{2,j,t}=\lambda_{2,j}(\sigma_t^2-R_{2,j,t})\,dt,\qquad j\in\{0,1\}.$$4

is sufficient to reconstruct $$dR_{1,j,t}=\lambda_{1,j}\sigma_t\,dW_t-\lambda_{1,j}R_{1,j,t}\,dt,\qquad dR_{2,j,t}=\lambda_{2,j}(\sigma_t^2-R_{2,j,t})\,dt,\qquad j\in\{0,1\}.$$5, $$dR_{1,j,t}=\lambda_{1,j}\sigma_t\,dW_t-\lambda_{1,j}R_{1,j,t}\,dt,\qquad dR_{2,j,t}=\lambda_{2,j}(\sigma_t^2-R_{2,j,t})\,dt,\qquad j\in\{0,1\}.$$6, and therefore $$dR_{1,j,t}=\lambda_{1,j}\sigma_t\,dW_t-\lambda_{1,j}R_{1,j,t}\,dt,\qquad dR_{2,j,t}=\lambda_{2,j}(\sigma_t^2-R_{2,j,t})\,dt,\qquad j\in\{0,1\}.$$7. Later computational work exploits exactly this property by learning quantities such as the VIX as measurable functions of the model parameters and these four factors, rather than of the full path (Gazzani et al., 2024).

4. Mathematical theory: wellposedness, non-explosion, and positivity

The main foundational result for the 4-factor Markov system is global strong wellposedness. The model has a unique strong solution and is non-explosive for all parameter values under the stated assumptions (Nutz et al., 2023). The proof proceeds by first establishing local wellposedness from standard SDE arguments and then ruling out finite-time explosion. The non-explosion step uses stopping times and estimates of the form

$$dR_{1,j,t}=\lambda_{1,j}\sigma_t\,dW_t-\lambda_{1,j}R_{1,j,t}\,dt,\qquad dR_{2,j,t}=\lambda_{2,j}(\sigma_t^2-R_{2,j,t})\,dt,\qquad j\in\{0,1\}.$$8

for

$$dR_{1,j,t}=\lambda_{1,j}\sigma_t\,dW_t-\lambda_{1,j}R_{1,j,t}\,dt,\qquad dR_{2,j,t}=\lambda_{2,j}(\sigma_t^2-R_{2,j,t})\,dt,\qquad j\in\{0,1\}.$$9

followed by Grönwall’s inequality (Nutz et al., 2023).

A central qualitative issue is the sign of the volatility process. In the 2-factor exponential model, a sufficient condition for positivity is

$R_{1,j,t}$0

The 4-factor model is materially different: an analogous global positivity statement fails in general, and it may happen that $R_{1,j,t}$1 but $R_{1,j,t}$2 for some $R_{1,j,t}$3, even if a formal analogue of the 2-factor condition holds (Nutz et al., 2023). One common misconception is therefore to treat the 4-factor extension as a routine multi-scale enrichment of the 2-factor case; mathematically, the positivity theory changes substantially.

The general-kernel analysis sharpens this point. For the non-Markovian stochastic Volterra formulation, existence and uniqueness of a continuous solution are proved under integrability and regularity assumptions on the kernels, together with a lower-bound condition on the activity feature. Under additional assumptions—most notably an exponential-type form for the return kernel $R_{1,j,t}$4 and the compatibility inequality

$R_{1,j,t}$5

—the paper proves a positive stochastic lower bound for $R_{1,j,t}$6, hence $R_{1,j,t}$7 a.s. if $R_{1,j,t}$8 (Andrès et al., 2024). In the pure exponential case, this condition reduces to $R_{1,j,t}$9, recovering the earlier 2-factor criterion. The same paper also stresses that the 4-factor Markov model should be viewed as a special finite-dimensional approximation of this richer non-Markovian theory.

A further unresolved point is the martingale property of the associated exponential price process. For the 2-factor PDV model, the exponential local martingale is shown to be a true martingale, but for the 4-factor model the problem remains open because the corresponding non-explosion estimate used in the 2-factor proof breaks down (Nutz et al., 2023).

5. Pricing, VIX representation, and calibration technology

Under the risk-neutral dynamics used in the SPX/VIX literature,

$R_{2,j,t}$0

SPX option prices are computed by Monte Carlo. The VIX is defined by the usual 30-day expected integrated variance relation

$R_{2,j,t}$1

In the 4-factor PDV model this conditional expectation is not available in closed form, because $R_{2,j,t}$2 contains nonlinear terms such as $R_{2,j,t}$3 and, in the parabolic version, the square-root term involving $R_{2,j,t}$4 (Gazzani et al., 2024).

This computational bottleneck motivated a sequence of neural-surrogate constructions. In the pathwise VIX-learning approach, the VIX is learned as a function

$R_{2,j,t}$5

where $R_{2,j,t}$6 collects the 10 model parameters and $R_{2,j,t}$7 is the 4-dimensional Markov state. The network therefore takes 14 inputs. The reported training setup uses a feed-forward fully connected architecture selected by Bayesian hyperparameter tuning, with training and validation sizes $R_{2,j,t}$8 and $R_{2,j,t}$9. Reported performance includes, for 99% of tested parameter configurations, a mean absolute error below about 0.55 VIX points, and for realistic calibrated parameters an average absolute error around 0.2 VIX points. The same framework yields SPX surface calibration up to 1 year in about 30 minutes on a GPU setup and joint SPX/VIX calibration in about 8 minutes (Gazzani et al., 2024).

A later deep-calibration paper removes not only the inner nested simulation for VIX, but also the expensive outer simulation inside the calibration loop. It learns SPX implied volatilities, VIX futures, and VIX call prices directly with two feed-forward networks, so that online pricing reduces to simple matrix–vector products. The reported calibration time then shrinks from about 12 minutes on a GPU in the half-neural/half-Monte-Carlo approach to about 5 seconds on a CPU in the fully neural approach (Baschetti et al., 12 Jul 2025). The same study reports that, on historical SPX surfaces from 2010–2023, the 4-factor Markov PDV model generally outperforms rough Heston and rough Bergomi in MSE/MAE for the surfaces studied, while also noting that the PDV specification becomes less flexible for very short-dated options and for full joint SPX/VIX calibration.

6. Extensions, reinterpretations, and related frameworks

The 4-factor Markov PDV model has been embedded into broader implied-volatility and stochastic-volatility frameworks. One direct extension combines a PDV-type underlying model with a parsimonious SSVI surface. In that construction, the parameters $$B_0,\;B_2,\;\lambda_{1,j},\;\lambda_{2,j}\ge 0,\qquad B_1<0,\qquad \theta_1,\theta_2\in[0,1],$$0 and $$B_0,\;B_2,\;\lambda_{1,j},\;\lambda_{2,j}\ge 0,\qquad B_1<0,\qquad \theta_1,\theta_2\in[0,1],$$1, which govern the at-the-money-forward implied volatility term structure through $$B_0,\;B_2,\;\lambda_{1,j},\;\lambda_{2,j}\ge 0,\qquad B_1<0,\qquad \theta_1,\theta_2\in[0,1],$$2, are regressed on PDV-style memory features built from past returns and squared returns, while the residual dynamics of $$B_0,\;B_2,\;\lambda_{1,j},\;\lambda_{2,j}\ge 0,\qquad B_1<0,\qquad \theta_1,\theta_2\in[0,1],$$3 are modeled by a hidden semi-Markov diffusion (Andrès et al., 2023). This shows that the PDV idea is not confined to spot-volatility forecasting; it also supplies a state representation for implied-volatility-surface dynamics.

A different line of work links stochastic volatility (SV) and PDV by filtering. Using assumed density filtering with an Inverse-Gamma approximation, a filtered SV model can be written as a PDV model in discrete time. In the Heston example, the correction step adds $$B_0,\;B_2,\;\lambda_{1,j},\;\lambda_{2,j}\ge 0,\qquad B_1<0,\qquad \theta_1,\theta_2\in[0,1],$$4 to the shape parameter and $$B_0,\;B_2,\;\lambda_{1,j},\;\lambda_{2,j}\ge 0,\qquad B_1<0,\qquad \theta_1,\theta_2\in[0,1],$$5 to the scale parameter, and under a constant-$$B_0,\;B_2,\;\lambda_{1,j},\;\lambda_{2,j}\ge 0,\qquad B_1<0,\qquad \theta_1,\theta_2\in[0,1],$$6 approximation the filtered-volatility recursion reduces to an explicit Guyon–Lekeufack-type PDV form with exponentially weighted past returns and past squared returns (Cohen et al., 2 Oct 2025). The paper states that this PDV representation can be viewed as a 4-factor Markov state process if one introduces a small set of sufficient statistics, such as the filtered volatility mean, exponentially weighted sums of returns and squared returns, and an auxiliary normalization variable. This suggests a conceptual bridge between finite-factor PDV models and filtered latent-volatility models.

The model also sits inside more general analytical envelopes. The signature-volatility framework represents volatility as a linear functional of the time-extended Brownian signature and explicitly treats finite-dimensional truncations as special cases; in that setting, a 4-factor Markov PDV model can be realized by choosing the coefficient process to load on four effective state components (Jaber et al., 2024). The path-dependent affine-process framework develops exponential-affine Fourier–Laplace formulas and generalized Riccati equations for path-dependent coefficients, and its rolling-window and delayed-volatility examples naturally suggest 4-factor Markov lifts through auxiliary lag and memory variables (Günther et al., 22 Jun 2026). These developments do not replace the 4-factor PDV model; rather, they provide broader mathematical contexts in which it appears as a tractable finite-dimensional specialization.

The principal open and delicate issues remain the same across these reinterpretations. Markovianity is achieved only in an augmented state space, not in the observed asset price alone; positivity is straightforward in some 2-factor or kernel-restricted settings but not in the general 4-factor model; and the true-martingale property of the exponential price process is unresolved in the canonical 4-factor formulation (Nutz et al., 2023). These features define the current technical frontier of the model as much as its empirical calibration success.