Jesse Schmolze

Learning Time-Inhomogeneous Markov Dynamics in Financial Time Series via Neural Parameterization

with Jan Rovirosa  ·  advised by Prof. Alejandra Quintos

Preprint, arXiv:2605.04690. Presented at the UW–Madison Undergraduate Research Symposium (April 2026) and the ASA Midwest Regional Conference on Data Science and Statistics (May 2026).

This paper develops a neural network framework for modeling the dynamics of equity returns as a time-inhomogeneous Markov chain, estimating regime-dependent transition matrices as a function of macroeconomic and fundamental signals. The approach addresses the well-documented failure of the standard Markov property in financial time series, offering a flexible and interpretable alternative to static transition models.

arXiv:2605.04690 Read the paper (PDF)

Inspectable Neural Markov Models for Non-Stationary Time Series

with Jan Rovirosa

Working paper. Multi-asset extension of the paper above.

This paper extends the neural Markov framework to a 10-asset US bank panel covering 24 years and 6,183 trading days per asset, testing how the choice of conditioning variable affects the structural consistency of the learned transition matrices. Conditioning on realized volatility rather than returns yields a 5.6% improvement in internal model consistency and superior out-of-sample fit in 9 of 10 assets, with a universal cross-asset pattern showing that high-volatility regimes compress next-step distributions regardless of the current state.

Who Captures Trade-Network Growth? Institutional Capacity and the Absorption of Foreign Shocks

Final draft complete. Publication to follow.

This paper uses Double Machine Learning on a panel of 111 countries from 1996 to 2023 to estimate how a receiving country's political stability moderates its absorption of exogenous growth shocks transmitted through bilateral trade networks. Instrumenting via Leave-i-Out and commodity terms-of-trade shocks provides causal evidence that the absorption rate is near zero for politically unstable countries and approaches one-to-one for the most stable, with the baseline DML estimates shown to be a conservative lower bound.

Autonomous Race Strategy via Reinforcement Learning

Badger Solar Racing, UW–Madison

This project develops a real-time race strategy system for a competitive solar car in the American Solar Challenge, a ten-day, 1,000-mile cross-country race. A Proximal Policy Optimization agent is trained on a neural network surrogate of a high-fidelity MATLAB/Simulink vehicle simulation, learning to issue speed commands every five minutes that maximize distance while keeping the battery alive across a full race day.