The research, led by Muhammed Ahmed Ibrahim with Tamirat Temesgen Dufera and Sileshi Mebrate, focuses on the fractional Fokker–Planck equation, or FFPE, governing the joint probability density of an asset price and its variance in a Heston-type stochastic volatility model. Unlike the classical Heston framework of 1993, which assumes fixed parameters and memoryless Brownian drivers, the new formulation allows every key coefficient—the drift, the mean-reversion rate, the long-run variance, and the volatility of volatility—to vary explicitly with time. Crucially, the model replaces standard Brownian motion with fractional Brownian motion, a stochastic process that carries long-range dependence and memory through a single tunable quantity known as the Hurst exponent.

The mathematical heart of the work lies in the connection between the fractional order of the governing equation and market memory itself. The authors tie the Caputo fractional derivative order alpha directly to the Hurst exponent through the relation alpha equals two times H, restricted to H between zero and one-half. This constraint is not arbitrary: it ensures the equation is well posed under a single initial condition and guarantees that when H equals one-half—the memoryless case—the formulation collapses exactly onto the classical Heston Fokker–Planck equation. In other words, the fractional model is a genuine generalization that recovers the textbook result as a limiting case, rather than an ad hoc extension detached from stochastic foundations.

Solving such an equation numerically is notoriously difficult. The Caputo derivative is nonlocal, meaning the rate of change at any instant depends on the entire history of the solution, and conventional finite difference or finite element schemes struggle to handle fractional operators efficiently, particularly as dimensionality grows. The authors’ answer is a physics-informed neural network, or PINN, an architecture that embeds the governing equation directly into the training loss of a deep neural network. Instead of discretizing the domain into a mesh, the network takes asset price, variance, and time as inputs and learns to output the full probability density as a smooth, continuous function across the entire spatiotemporal domain.

The PINN itself is a fully connected network with six hidden layers of eighty neurons each, using tanh activations in the interior and a Softplus output layer that enforces the positivity every probability density demands. Training minimizes a composite loss combining four penalties: the residual of the fractional Fokker–Planck equation evaluated at thousands of collocation points, deviations from the initial Gaussian-kernel approximation of a point mass, boundary attenuation at the edges of the truncated computational domain, and a normalization constraint enforcing conservation of total probability. The fractional derivative is approximated with an L1-type quadrature over two hundred temporal nodes, and all spatial derivatives are computed exactly through automatic differentiation. Optimization proceeds in two stages, with twenty thousand iterations of the Adam optimizer followed by quasi-Newton refinement under L-BFGS.

The results are striking. The mean absolute residual of the governing equation fell to roughly 4.5 times ten to the minus six, with the overall PDE loss stabilizing below ten to the minus four after about eight thousand epochs. Independent training runs with five different random seeds produced tightly clustered errors, with a mean relative error of about 0.2 percent and small standard deviations, demonstrating that the framework is reproducible rather than lucky. When benchmarked against finite difference, spectral, and Monte Carlo methods at matched accuracy, the PINN achieved the lowest mean absolute error and root-mean-square error of the four approaches, with unit correlation against reference solutions throughout the price domain.

The physics the network uncovers is equally revealing. Sensitivity experiments across fractional orders from 0.1 to 0.9 showed that stronger memory—smaller values of alpha—produces broader, flatter, increasingly non-Gaussian probability distributions, confirming that the fractional operator plays a genuine dynamical role rather than serving as a cosmetic correction. Time-dependent coefficients amplified this effect further: allowing the mean-reversion and volatility-of-volatility parameters to evolve in time increased the kurtosis of the variance distribution by roughly 72 percent compared with the constant-parameter case, a signature of stronger volatility clustering. Notably, the time-dependent specification mainly reshaped the variance dynamics, while the asset-price distribution changed more modestly, suggesting that where markets store and release uncertainty depends heavily on how their parameters drift through time.

The framework’s practical payoff comes through option pricing. By integrating the discounted call payoff against the learned density, the authors priced European options across a range of strikes and compared the results with the Black–Scholes benchmark calibrated at the money. The fractional time-dependent model produced systematically higher prices than Black–Scholes, with the gap peaking at about 0.112 near the at-the-money strike, and it generated a smooth, downward-sloping implied volatility skew—falling from roughly 0.83 for deep in-the-money options to about 0.45 for out-of-the-money contracts—precisely the kind of structure observed in real markets but invisible to constant-volatility models. The implied volatility term structure remained stable and monotone across maturities from a quarter to more than a year.

The authors are careful about the limits of their claims. The study was validated on synthetic, literature-parameterized settings rather than calibrated market data, the efficiency comparison against finite difference solvers is hardware-dependent and specific to the two-dimensional problem, and the memory cost of the nonlocal Caputo quadrature grows with temporal resolution. Extensions to Hurst exponents above one-half, multi-asset systems, and rough volatility models remain open. Still, the demonstration that a physics-informed network can solve a fully time-dependent, memory-bearing stochastic volatility equation with residuals near ten to the minus six—on an ordinary laptop—marks a meaningful step toward financial models that treat market memory not as a nuisance but as a computable, physically interpretable quantity. As quantitative finance increasingly borrows the tools of machine learning, this work suggests the future of risk modeling may belong to networks that are taught the laws of physics before they are asked to predict anything at all.

Subject of Research: Physics-informed neural networks for solving the time-fractional Fokker–Planck equation in a Heston-type stochastic volatility model with time-dependent parameters.

Article Title: Physics-informed neural networks for the fractional Fokker–Planck equation in a Heston-type stochastic volatility model with time-dependent parameters

Article References: Ibrahim, M. A., Dufera, T. T., & Mebrate, S. (2026). Physics-informed neural networks for the fractional Fokker–Planck equation in a Heston-type stochastic volatility model with time-dependent parameters. Machine Learning with Applications, 26, Article 101007. https://doi.org/10.1016/j.mlwa.2026.101007

Image Credits: AI Generated

DOI: 10.1016/j.mlwa.2026.101007

Keywords: physics-informed neural networks, fractional Fokker–Planck equation, stochastic volatility, Heston model, fractional Brownian motion, Caputo derivative, option pricing, implied volatility skew, anomalous diffusion, computational finance, machine learning, deep learning