Improvements to the post-processing of weather forecasts using machine learning and feature selection
Improvements to the post-processing of weather forecasts using machine learning and feature selection
Authors: Kazuma Iwase, Tomoyuki Takenawa Date: 2026-04-21 Paper ID: openalex:2604.19340
Summary
This study improves weather forecast post-processing for temperature, wind speed, and precipitation by using LightGBM-based models on the Japan Meteorological Agency’s Mesoscale Model (MSM) data. The authors utilize feature selection based on grid-point correlations to enhance model input, consistently outperforming standard neural-network baselines and official guidance. Additionally, they address the challenges of precipitation skewness and zero-inflation by implementing Tweedie loss functions and event-weighted training strategies.
Key Contributions
- Developed LightGBM-based post-processing models for precipitation, temperature, and wind speed that outperform raw MSM forecasts and MSMG guidance.
- Introduced a feature selection approach using correlation analysis on surrounding grid points to enhance predictive input relevance.
- Evaluated Tweedie-based loss functions and event-weighted training strategies to address skewness and zero-inflation in precipitation forecasting.
Open Questions & Future Work
Archivist Review
The paper offers incremental improvements to standard weather post-processing tasks. I approved one open question concerning the site-dependence of precipitation forecasting, which is a known challenge in time-series meteorology. I am registering the ‘msm-dataset’ for tracking, though it is a specialized regional product. No new concepts were approved as the methods described (LightGBM, Tweedie loss, correlation-based feature selection) are standard techniques rather than novel contributions to the field.
Approved Open Questions
- ML Precipitation Post-processing Site-dependence: Addressing the site-dependence of ML precipitation models is critical for moving beyond simplistic local adjustments and achieving generalizable meteorological forecasting performance.
Rejected Candidates
- [dataset] MSM (Mesoscale Model) dataset (
msm-dataset) - other: The dataset itself is regional and proprietary-adjacent to JMA; I will approve it as ‘msm-dataset’ for reusability, but note it is highly specialized.
Datasets
Links
Metadata & Links
- url
- https://arxiv.org/abs/2604.19340
- paper_id
- 2604.19340
- paper_source
- openalex
- domain
- time-series
- tags
-
- architectures
-
- datasets
- msm-dataset
- concept_slugs
-
- dataset_slugs
- msm-dataset
- skill
- TimeSeriesSkill
- created_at
- 2026-04-24T05:51:13Z