Financial Mathematics Seminar

Department of Mathematics · Florida State University · Fall 2026

When & Where: Thursdays, 3:05–3:55 pm · Love Building (LOV), Room 0301
Talks are also streamed on Zoom — join link. All are welcome!

Upcoming Talks

Date Speaker Title & Abstract
Aug 27, 2026 — Organizational meeting
Sep 3, 2026 Ali Kara
Florida State University
Mean-Field Control with Decentralized Observations.
Abstract

TBA

Sep 10, 2026 Lingjiong Zhu
Florida State University
Optimal Portfolio Choice with Diffusion-Based Data Augmentation
Portfolio selection rules rely on estimated parameters, whose errors, driven by limited sample sizes, often erode realized performance. We propose augmenting observed data with synthetic samples generated by diffusion models. We show theoretically that the data-augmented rule dominates the standard plug-in rule in expected utility, and empirically, using simulated and real market data, find that augmentation substantially improves out-of-sample performance across every rule examined. This offers a practical tool for mitigating estimation risk, pointing to a promising role for generative AI in decision-making under uncertainty, including in predictions and sentiment measures.

TBA

Sep 17, 2026 Qijin Shi
UCSB
No-arbitrage with Rough Paths: A Collapse Toward the Classical Setting
Rough path theory extends stochastic calculus to irregular processes beyond the classical semimartingale framework. We investigate whether this broader calculus also permits a broader theory of frictionless financial markets without arbitrage. For markets with gains specified by rough integration, we formulate a No Controlled Free Lunch condition and use a Kreps–Yan-type theorem to characterize it through zero expected gains under an equivalent measure. This leads to the study of unbiased rough integrators. We establish classification results for one-dimensional rough paths under suitable regularity and integrability assumptions. As the permitted strategies expand from state-dependent rules to signature-based portfolios incorporating past information, the classification becomes increasingly restrictive: Gaussian marginals and a Hermite structure give way to a Brownian Itô lift with a deterministic change of time. Thus, within this framework, sufficiently rich continuous trading brings the model back to the classical semimartingale setting.

TBA

Sep 24, 2026 Qi Feng
Florida State University
Title: Continuous Policy and Value Iteration for Relaxed Stochastic Control, and beyond.
Abstract

We introduce a continuous policy-value iteration algorithm where the approximations of the value function of a stochastic control problem and the optimal control are simultaneously updated through Langevin-type dynamics. This paper focuses on the entropy-regularized relaxed control problem with infinite horizon. We establish policy improvement and demonstrate convergence to the optimal control under the monotonicity condition of the Hamiltonian. By utilizing Langevin-type stochastic differential equations for continuous updates along the policy iteration direction, our approach enables the use of distribution sampling techniques in machine learning to optimize the value function and identify the optimal control simultaneously. Numerical experiments are presented for LQ model with a non-concave reward, and a non-LQ example with a skewed-Laplace optimal control. Towards the end, I will present our new results on score-Hamiltonian policy iteration algorithms.

Oct 1, 2026 Brayton Link
Florida State University
Signature Kernel Learning for PDEs
Abstract

Partial differential equations (PDEs) provide a fundamental framework for modeling systems that evolve in both space and time, but finding their solution can become challenging when the underlying forcing or coefficient functions are known only through discrete or sequential observations. This work investigates the use of signature kernels as a machine learning framework for approximating solutions to such PDEs. Signature kernels arise from the path signature, which represents sequential data through a hierarchy of iterated integrals that capture increasingly complex interactions along a path. We focus on inhomogeneous PDEs in which the forcing function is observed along spatial and temporal trajectories. These trajectories are embedded into reproducing kernel Hilbert spaces through the signature kernel, allowing the unknown solution to be approximated using finite kernel expansions. In particular, we study the inhomogeneous heat and wave equations as a model problem and construct approximations of its solution using signature-kernel Gram matrices generated from sampled forcing data. The resulting framework combines the structure of the governing differential equation with data-driven representations of sequential observations, providing a connection between signature methods, kernel-based machine learning, and numerical PDE approximation.

Oct 8, 2026 David Herzog
Iowa State University
Title: Quantitative positivity of transition densities for random perturbations of Hamiltonian systems
Abstract: I will present recent results on a class of diffusion processes arising from random perturbations of conservative Hamiltonian systems. Under a set of abstract hypotheses---including basic structural assumptions on the Hamiltonian, a weak Lyapunov structure, and a quantitative notion of smoothing---we prove that transition densities satisfy a sharp, uniform pointwise lower bound over Hamiltonian sublevel sets in the small noise limit. By applying our general theorem, we obtain quantitative minorization estimates for a variety of models including Langevin dynamics, chains of oscillators coupled to heat bathes at different temperatures, and finite-dimensional fluid models such as stochastically forced Galerkin truncations of the Navier-Stokes equations and the Lorenz '96 system. As a corollary, assuming a stronger Lyapunov structure, our main result yields a sharp exponential rate of convergence to equilibrium in a weighted total variation norm. A central feature of our approach is that it does not require knowledge of the explicit form of the invariant measure, nor even its existence, and hence is broadly applicable to deduce minorization for physically relevant systems where invariant measures are inaccessible.

TBA

Oct 15, 2026 TBA —
Oct 22, 2026 Haoyang Cao
Johns Hopkins University
Title: Randomized Impulse Control Problem and Reinforcement Learning
Abstract

In this work, we study a randomized optimal impulse control problem. This framework is particularly suitable for devising reinforcement learning algorithms with exploration to solve the otherwise hard-to-explicitly-derive optimal impulse control policies. We provide a two-fold randomized impulse control framework and then analyze the associated HJB equation. We discuss its relations with (randomized) optimal stopping as well as the classical impulse control problem. We also present an exploratory reinforcement learning algorithm to learn the randomized impulse policy. This is a joint work with Prof. Yuchao Dong (Tongji University) and Zhouhao Yang (Johns Hopkins University).

Oct 29, 2026 Munawar Ali
Florida State University
Title: The Generalized Volterra Signature
Abstract

In this paper, we introduce the generalized Volterra signature \(\gvsig(\bX)\) of a path \(\bX\), extending the classical Volterra signature introduced in [Hager, Harang, Pelizzari \& Tindel, arXiv:2603.04525, 2026], which extends Lyons' path signature to the Volterra setting. We establish a new algebraic structure to encode the generalized Volterra signature. We also establish a generalized Chen's identity and the shuffle product of signature components for the geometric Volterra signature \(\geovsig(\bX)\). The shuffle product for the Volterra signature is the first of its kind and allows us to formulate a universal approximation theorem for geometric Volterra signatures. For a special class of kernel functions, we prove the existence of an extended path \(\Bar{\bX}\) and relate its classical signature to the geometric Volterra signature of the original path \(\bX\).

Nov 5, 2026 Daniel Michael Mazus
Florida State University
Title: Random Time Sampling for Probabilistic Methods in Stochastic Control
Abstract

Policy-gradient methods for continuous-time stochastic control estimate the cost of a policy by simulating the controlled state on a time grid. The accuracy of this estimate is limited by the time step, and every additional step is paid for again in each gradient evaluation. Walk-on-Shapes methods take a different approach: instead of following the whole path, they sample only where it leaves a sequence of spheres or space-time regions, so that a walk needs only a small random number of steps. This talk reviews time-grid methods for stochastic control and the Walk-on-Spheres and Walk-on-Heat-Balls algorithms, and then presents Walk-on-Cylinders (WoC), which solves the heat equation by sampling exits from space-time cylinders. Fixing the ratio of each cylinder’s height to its squared radius makes all exit laws rescalings of three fixed distributions, which are tabulated once before the walk starts. We discuss the expected number of steps of WoC and compare it with Walk-on-Heat-Balls on the test problems of Deaconu and Herrmann, where WoC reaches the same accuracy with fewer steps. Finally, we show that the cost of a fixed policy solves a linear PDE in which the running cost is a source term, and propose extending WoC to this source by sampling a space-time point inside each cylinder.

Nov 12, 2026 Distinguished Lecture —
Nov 19, 2026 Navid Bahadoran
Florida State University
Title: Targets for PCA in High-Dimension, Low-Sample-Size Factor Models
Abstract

When the number of assets far exceeds the number of observations, the leading principal components of returns estimate the factor directions poorly. This talk splits their error into two parts: a tilt out of the factor subspace, which the data can measure but only partly reduce, and a rotation inside it, which the data cannot detect at all. Targets such as market betas, firm characteristics or past principal directions can supply what the data lack. We show how to measure, from the data alone, how much a given target is worth. We correct the bias that arises when a target is built from the same returns. We also derive the best combination of sample directions and targets, with consistent estimates of its accuracy. Applications to covariance estimation and portfolio risk are discussed.

Nov 26, 2026 — No seminar — Thanksgiving break
Dec 3, 2026 Ye Chen
Florida State University
Title: Short-Horizon Portfolio Management with Diffusion-Based Data Augmentation
Abstract

Mean–variance portfolio optimization is sensitive to estimation error, particularly at short investment horizons, where noisy return estimates can generate unstable allocations and excessive trading. This study investigates diffusion-based data augmentation as a means of regularizing portfolio decisions. Under Gaussian returns, we characterize expected utility through the generator’s conditional moments and show that naive resampling recovers the classical plug-in portfolio in the infinite-sample limit. Appropriately tuned diffusion, with squared signal-decay factor proportional to the inverse sample size, can instead yield a positive second-order improvement in expected utility under the theoretical assumptions. For implementation, we exploit the affine Gaussian score to propagate the reverse Euler scheme’s mean and covariance deterministically, eliminating synthetic-sampling noise. We combine these moments with nested chronological validation of the diffusion horizon, constrained portfolio optimization, turnover penalties, and partial rebalancing. Development backtests on five large-cap U.S. equities compare the resulting strategies with equal-weight, classical mean–variance, and Ledoit–Wolf benchmarks across weekly to quarterly holding periods. Results favor diffusion at weekly and biweekly horizons, while covariance shrinkage performs better at longer horizons. Turnover control substantially reduces trading and improves net performance in the reported experiments. These findings support further evaluation of diffusion-based moment regularization combined with trading controls; confirmation on a locked holdout remains necessary.

Past Talks

Spring 2026

  • Apr 17Speaker Name (Affiliation) — Talk Title Here
  • Mar 27Speaker Name (Affiliation) — Talk Title Here
  • Feb 13Speaker Name (Affiliation) — Talk Title Here

Fall 2025

  • Nov 21Speaker Name (Affiliation) — Talk Title Here

Organizers & Contact

The seminar is organized on a rotating basis by Qi Feng (qfeng2@fsu.edu), Ali Kara (adk24b@fsu.edu), and Lingjiong Zhu (lzhu2@fsu.edu).

Fall 2026 and Spring 2027 are organized by Qi Feng.

To suggest a speaker or join the mailing list, email the current organizer. Directions to campus and parking information are available here.