Skip to content

DeepLévy: Learning Heavy-Tailed Uncertainty in Highly Volatile Time Series

Home / Papers / DeepLévy: Learning Heavy-Tailed Uncertainty in Highly Volatile Time Series

DeepLévy: Learning Heavy-Tailed Uncertainty in Highly Volatile Time Series

Authors: Yang Yang, Du Yin, Hao Xue, Flora D. Salim Date: 2026-05-11 Paper ID: openalex:2605.10364

Summary

DeepLévy is a novel probabilistic forecasting framework designed to handle heavy-tailed uncertainty in highly volatile time series. By moving away from likelihood-based inference, it instead learns mixtures of Lévy stable distributions through the minimization of discrepancy between empirical and parametric characteristic functions. This approach allows the model to flexibly capture extreme events and non-Gaussian behaviors over multiple forecasting horizons. Experiments confirm that DeepLévy consistently outperforms existing state-of-the-art models in tail risk assessment.

Key Contributions

  • Introduces DeepLévy, a neural framework for heavy-tailed uncertainty estimation in time series using mixtures of Lévy stable distributions.
  • Replaces intractable likelihood-based inference with characteristic function matching to enable stable learning of non-Gaussian distributions.
  • Demonstrates superior tail risk metric performance over state-of-the-art probabilistic models under high volatility.

Open Questions & Future Work

Key Concepts

  • deeplevy: A neural framework for probabilistic forecasting that models heavy-tailed uncertainty by learning mixtures of Lévy stable distributions via characteristic functions.

Archivist Review

I approved DeepLévy as it offers a distinct, theoretically grounded approach to heavy-tailed uncertainty estimation by moving from density-based to characteristic-function-based optimization. The identified open question is approved because it addresses a clear technical bottleneck in moving from univariate to multivariate risk modeling for heavy-tailed phenomena. No other candidates were proposed.

Approved Concepts

  • DeepLévy: It introduces a novel framework for modeling heavy-tailed uncertainty in volatile time series by bypassing intractable likelihoods using characteristic function matching.

Approved Open Questions

  • Multivariate Stable Distributions and Temporal Tail Dynamics: Multivariate modeling is essential for capturing inter-asset correlations during market shocks, and temporal tail dynamics would allow for more accurate dynamic risk assessment in non-stationary series.

Metadata & Links

url
https://arxiv.org/abs/2605.10364
paper_id
2605.10364
paper_source
openalex
domain
time-series
tags
architectures
datasets
concept_slugs
deeplevy
dataset_slugs
skill
TimeSeriesSkill
created_at
2026-05-14T06:08:57Z