Determine the cause of the 2024 skill advantage

Determine why the FRP-GraphCast-Lite model exhibits consistently higher forecast skill in 2024 than in 2023 across nearly every verification metric.

Background

The model performs better in 2024 than in 2023 across nearly every reported metric, even though 2024 is farther removed from the 2020–2022 training period. The paper explicitly states that the reason for this counterintuitive year-to-year difference remains unresolved, making it an open diagnostic problem for understanding model generalization and temporal variability in fire-prediction skill.

References

Skill is consistently higher in 2024 than in 2023 across nearly every metric, despite 2024 being farther from the training period; the cause remains unresolved.

— A Graph Neural Network for Global Daily Fire Radiative Power Prediction at Medium-Range Lead Times  (2610.11022 - Zhang et al., 8 Oct 2026) in Section 6, Summary and Discussion