Embracing Heteroscedasticity for Probabilistic Time Series Forecasting

Authors: Yijun Wang, Qiyuan Zhuang, Xiu-Shen Wei Date: 2026-03-25 Paper ID: openalex:2603.24254

Summary

This paper addresses the limitation of existing non-autoregressive probabilistic time series forecasting (PTSF) models, which often implicitly assume homoscedasticity due to MSE-based objectives. The authors propose the Location-Scale Gaussian VAE (LSG-VAE), which explicitly models time-varying conditional variance using a location-scale likelihood formulation, enabling faithful capture of aleatoric uncertainty. LSG-VAE also incorporates an adaptive attenuation mechanism to robustly handle high-volatility observations during training. Experiments across nine benchmarks show LSG-VAE significantly outperforms fifteen strong generative baselines while maintaining efficiency.

Key Contributions

• Proposed Location-Scale Gaussian VAE (LSG-VAE), a generative PTSF framework that explicitly parameterizes time-dependent variance via a location-scale likelihood, overcoming the homoscedastic bias of models like TimeVAE.
• Introduced an adaptive attenuation mechanism to automatically down-weight highly volatile observations during training, enhancing robustness in trend prediction.
• Demonstrated consistent outperformance against fifteen strong generative baselines across nine benchmark datasets in probabilistic forecasting tasks.

Limitations

The abstract does not specify any explicit limitations or future work directions beyond the scope of the main proposal.

Key Concepts

• location-scale-gaussian-vae-lsg-vae: A VAE framework for probabilistic time series forecasting that explicitly models time-dependent conditional variance using a location-scale likelihood formulation.
• adaptive-attenuation-mechanism-volatility-downweighting: A mechanism within LSG-VAE that automatically reduces the influence of highly volatile training observations to improve trend prediction robustness.

Archivist Review

Two primary concepts were approved: the Location-Scale Gaussian VAE (LSG-VAE) as the core modeling framework, and the Adaptive Attenuation Mechanism as a reusable training stability technique for heteroscedastic models. No specific open questions were identified, and the single dataset mentioned (ETTh1) is a generic benchmark that does not require a unique vault entry. The review focused on approving mechanisms central to handling time-varying uncertainty in sequence generation.

Approved Concepts

• Location-Scale Gaussian VAE (LSG-VAE): LSG-VAE is the core proposed method that explicitly models heteroscedasticity in probabilistic forecasting using a location-scale likelihood.
• Adaptive Attenuation Mechanism: This is a novel training regularization technique specifically designed to improve robustness in heteroscedastic modeling by managing the impact of high-variance points.

Rejected Candidates

• [dataset] ETTh1 (ETTh1) - low_impact: ETTh1 is a common benchmark dataset that does not warrant a standalone vault entry unless the paper introduced it or showed a novel technique uniquely applicable to it.

Links

• Abstract
• PDF