Skip to content

QuantWeather: Quantile-Aware Probabilistic Forecasting for Subseasonal Precipitation

Home / Papers / QuantWeather: Quantile-Aware Probabilistic Forecasting for Subseasonal Precipitation

QuantWeather: Quantile-Aware Probabilistic Forecasting for Subseasonal Precipitation

Authors: Lei Chen, Xinyu Su, Xiaohui Zhong, Hao Li Date: 2026-05-11 Paper ID: openalex:2605.10297

Summary

QuantWeather is a novel end-to-end probabilistic forecasting framework designed for subseasonal precipitation that replaces traditional ensemble post-processing with a dual-head architecture. By jointly optimizing deterministic and probabilistic heads, the model provides reliable uncertainty estimates while allowing for stochastic sampling in a single forward pass. This approach significantly reduces the computational and memory burdens associated with generating and maintaining large reforecast datasets for post-hoc calibration. Evaluations show that the framework achieves higher forecasting skill than standard ensemble methods across key performance metrics.

Key Contributions

  • QuantWeather, a dual-head probabilistic forecasting framework, achieves superior uncertainty calibration compared to standard ensemble-based methods for subseasonal precipitation.
  • The model supports efficient stochastic sampling, enabling high-fidelity probabilistic outputs with a single forward pass, significantly reducing computational and storage overhead.
  • Experimental results demonstrate that QuantWeather outperforms conventional ensemble methods in probabilistic skill while eliminating the need for expensive post-hoc reforecast calibration.

Open Questions & Future Work

Key Concepts

Archivist Review

I have approved a generalized concept for the dual-head architecture used in the paper, as it represents a broader, reusable mechanism for uncertainty-aware forecasting. I have also approved a refined open question regarding the fundamental trade-offs between probabilistic calibration and ensemble dispersion in multi-objective architectures. Other candidates were rejected as near-duplicates or application-specific variants.

Approved Concepts

  • Dual-Head Probabilistic Forecasting Framework: This architecture provides a scalable alternative to computationally expensive post-hoc ensemble calibration by embedding probabilistic estimation directly into the model training objective.

Approved Open Questions

  • Quantile Calibration Dispersion Trade-off: Understanding this trade-off is essential for developing models that simultaneously deliver high-accuracy point forecasts and robust, well-calibrated uncertainty estimates without compromising the benefits of ensemble averaging.

Rejected Candidates

  • [concept] QuantWeather Framework (quantweather-framework) - duplicate_existing: This is a specific application instance of a dual-head framework, which is better captured by the generalized concept slug.
  • [open_question] Balancing Quantile Calibration and Dispersion (balance-quantile-calibration-ensemble-dispersion) - other: The title was rewritten for clarity and conciseness.

Metadata & Links

url
https://arxiv.org/abs/2605.10297
paper_id
2605.10297
paper_source
openalex
domain
time-series
tags
probabilistic-forecastinguncertainty-estimationtime-series-forecasting
architectures
datasets
concept_slugs
dual-head-probabilistic-forecasting-framework
dataset_slugs
skill
TimeSeriesSkill
created_at
2026-05-14T06:09:40Z