Discovery of Diverse and Realistic Financial Tail-Risk Using Generative Flow Networks
Lists
Only you see this
Abstract
Accurate modeling of rare high-impact financial events is essential for robust risk management, but existing methods face notable challenges. Sampling-based techniques like Sequential Markov Chain Monte Carlo generate diverse scenarios but often lack focus on specific high-risk outcomes, leading to inefficient exploration. In contrast, Deep Reinforcement Learning methods optimize sequential decision-making effectively but tend to converge on a narrow set of high-reward scenarios, limiting diversity. To address these shortcomings, we propose using Generative Flow Networks (GFlowNets), which naturally learn to sample according to a predefined reward distribution. Our approach systematically generates diverse and realistic tail-risk scenarios by explicitly defining rewards that prioritize high-risk, impactful financial scenarios. Experiments on real-world financial data show that our GFlowNet-based method significantly enhances scenario diversity and realism, effectively capturing the complex nonlinear dependencies typical of financial markets. These findings demonstrate the potential of GFlowNets as a robust framework for financial risk modeling, offering actionable insights into rare but critical market events.
BibTeX
@misc{serruya2026discovery,
author = {Adir Serruya and Gilad Katz},
title = {{Discovery of Diverse and Realistic Financial Tail-Risk Using Generative Flow Networks}},
year = {2026},
eprint = {2610.38838},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.38838},
}