This assessment is based on version 1 of this paper. Version 3 is now available on arXiv — the authors may have revised their methods, results, or conclusions.
Enhancing AI and Dynamical Subseasonal Forecasts with Probabilistic Bias Correction
Hannah Guan, Soukayna Mouatadid, Paulo Orenstein, Judah Cohen, Haiyu Dong, Zekun Ni, Jeremy Berman, Genevieve Flaspohler
Abstract
Decision-makers rely on weather forecasts to plant crops, manage wildfires, allocate water and energy, and prepare for weather extremes. Today, such forecasts enjoy unprecedented accuracy out to two weeks thanks to steady advances in physics-based dynamical models and data-driven artificial intelligence (AI) models. However, model skill drops precipitously at subseasonal timescales (2 - 6 weeks ahead), due to compounding errors and persistent biases. To counter this degradation, we introduce probabilistic bias correction (PBC), a machine learning framework that substantially reduces systematic error by learning to correct historical probabilistic forecasts. When applied to the leading dynamical and AI models from the European Centre for Medium-Range Weather Forecasts (ECMWF), PBC doubles the subseasonal skill of the AI Forecasting System and improves the skill of the operationally-debiased dynamical model for 91% of pressure, 92% of temperature, and 98% of precipitation targets. We designed PBC for operational deployment, and, in ECMWF's 2025 real-time forecasting competition, its global forecasts placed first for all weather variables and lead times, outperforming the dynamical models from six operational forecasting centers, an international dynamical multi-model ensemble, ECMWF's AI Forecasting System, and the forecasting systems of 34 teams worldwide. These probabilistic skill gains translate into more accurate prediction of extreme events and have the potential to improve agricultural planning, energy management, and disaster preparedness in vulnerable communities.
AI Impact Assessments
(3 models)Scientific Impact Assessment: Enhancing AI and Dynamical Subseasonal Forecasts with Probabilistic Bias Correction
1. Core Contribution
The paper introduces Probabilistic Bias Correction (PBC), a lightweight machine learning framework that learns to correct systematic errors in probabilistic subseasonal weather forecasts. Unlike approaches that attempt to build better atmospheric models from scratch, PBC operates as a post-processing layer that takes existing ensemble forecasts (from dynamical or AI models) and outputs corrected cumulative distribution functions. The framework combines two complementary corrections—Debias++ (location-, date-, and quantile-specific bias subtraction over adaptively selected training windows) and Persistence++ (regression-based blending of model forecasts with lagged observations and climatology)—followed by isotonic regression projection onto valid CDFs and simple averaging.
The key insight is elegant: while subseasonal models suffer from compounding systematic errors, these errors are themselves *predictable* and *correctable* with relatively simple statistical learning. This reframes the subseasonal "predictability desert" problem from one of fundamental atmospheric unpredictability to one of removable systematic bias.
2. Methodological Rigor
The experimental evaluation is comprehensive and convincing across multiple dimensions:
One methodological concern is the reliance on ERA5 reanalysis as ground truth, which has its own biases, particularly for precipitation in data-sparse regions. However, this is standard practice and the competition also used ERA5-T.
The algorithms themselves (Algorithms S1 and S2) are remarkably simple—essentially adaptive bias subtraction and linear regression with isotonic projection. This simplicity is both a strength (reproducibility, interpretability, low compute) and a potential limitation (ceiling on achievable improvement). The adaptive selection of training windows (spans of 14, 28, or 35 days) via 3-year lookback validation is a thoughtful design choice that balances bias-variance tradeoffs across different regimes.
3. Potential Impact
Immediate practical impact: The subseasonal timescale is critically important for agriculture, water management, energy planning, and disaster preparedness. The fact that PBC converts negative-skill forecasts into positive-skill forecasts for many targets—particularly precipitation, which is notoriously difficult—represents actionable improvement for decision-makers.
Extreme event forecasting: The flood forecasting analysis using GDACS events (543 events, 2022–2024) demonstrates that PBC transforms routinely negative BSS into consistently positive flood forecasting skill. Given recent work (reference [18]) quantifying how forecast improvements reduce mortality, these gains have clear humanitarian implications.
Operational adoption pathway: PBC's computational cheapness relative to dynamical ensembles or large AI models makes it immediately deployable, even at resource-constrained national meteorological services. The open-source release lowers adoption barriers further.
Paradigm influence: The work suggests that significant subseasonal skill may be "left on the table" by current operational debiasing protocols. This could redirect community effort toward better post-processing alongside model development.
4. Timeliness & Relevance
This paper arrives at a critical juncture. AI weather models (GenCast, Pangu-Weather, FuXi, AIFS) have demonstrated impressive medium-range skill but struggle at subseasonal timescales. Simultaneously, climate change is increasing the frequency and severity of extreme events, amplifying demand for extended-range forecasts. PBC directly addresses the gap between medium-range AI capabilities and subseasonal operational needs.
The ECMWF AI Weather Quest competition (launched 2025) signals institutional recognition that AI contributions to subseasonal forecasting need systematic benchmarking—PBC's dominance in this inaugural competition is well-timed for visibility and influence.
5. Strengths & Limitations
Key Strengths:
Notable Limitations:
Summary
This is a high-impact applied machine learning contribution that addresses a genuine operational bottleneck in weather forecasting. Its strength lies not in algorithmic novelty but in the systematic demonstration—across models, variables, regions, seasons, and a real-time competition—that lightweight statistical post-processing can yield substantial and reliable subseasonal skill gains. The work is well-positioned to influence operational forecasting practices globally.
Generated Apr 20, 2026
Comparison History (277)
Paper 1 demonstrates concrete, validated real-world impact: it doubles subseasonal AI forecast skill, won ECMWF's 2025 competition against 34 teams and six operational centers, and directly benefits agriculture, energy, and disaster preparedness. Its breadth of societal application and operational deployment give it broad, timely impact. Paper 2 offers rigorous theoretical unification of discrete diffusion models—elegant and clarifying—but its impact is narrower, primarily benefiting ML theorists and diffusion model designers rather than delivering broad cross-field or societal outcomes.
Paper 1 demonstrates immediate, massive real-world impact by fundamentally improving subseasonal weather forecasting, a notoriously difficult challenge. Its methodology was rigorously validated by winning the ECMWF 2025 real-time forecasting competition against top global operational centers and AI models. This proven operational superiority translates directly to critical applications in agriculture, disaster preparedness, and energy management. While Paper 2 presents an innovative generative AI approach for materials design, Paper 1's unprecedented empirical validation at a global scale and its direct, widespread societal utility give it a higher potential scientific and real-world impact.
While Paper 1 offers a highly relevant theoretical framework for AI copyright issues, Paper 2 demonstrates extraordinary, validated real-world impact. Subseasonal weather forecasting is notoriously difficult, yet Paper 2's PBC framework doubles AI forecasting skill and outperformed all major global dynamical models and AI forecasting systems to win the ECMWF 2025 real-time competition. Its immediate applications in disaster preparedness, agriculture, and energy management, combined with rigorous empirical validation against state-of-the-art operational baselines, give it a profoundly broader and more immediate scientific and societal impact across multiple disciplines.
Paper 1 addresses a critical global challenge—subseasonal forecasting—with demonstrated real-world validation (winning ECMWF's 2025 competition against 34 teams and six operational centers). Its impact spans agriculture, energy, and disaster preparedness for vulnerable communities. The results show dramatic, verified skill improvements. Paper 2 offers solid, timely engineering advances in quantization for diffusion models, but its impact is narrower and more incremental, competing in a crowded PTQ field with modest efficiency gains. Paper 1's breadth, rigor, and demonstrated operational superiority give it substantially higher scientific and societal impact.
Paper 1 has higher likely scientific impact: it delivers large, validated skill gains on societally critical subseasonal forecasting (weather extremes, agriculture, energy, disaster preparedness) using an operationally deployable probabilistic bias correction framework, with strong external evidence (ECMWF real-time competition performance). Its breadth spans climate science, ML for geophysics, and risk management, and is highly timely given growing extreme-event needs. Paper 2 is technically strong and useful for long-context LLM inference efficiency, but its impact is narrower (systems optimization) and primarily benefits a subset of ML deployment scenarios.
Paper 1 addresses a critical real-world problem (subseasonal forecasting) with demonstrated superior performance in a competitive operational benchmark, winning first place globally. Its impact spans agriculture, energy, and disaster preparedness—broad societal benefits, especially for vulnerable communities. While Paper 2 offers solid methodological innovation with deployment at scale, its impact is more confined to LLM post-training. Paper 1's combination of rigorous validation, timeliness (climate/extreme events), and breadth of downstream applications gives it higher potential scientific impact.
Paper 2 likely has higher scientific impact: it addresses a major, well-defined real-world problem (2–6 week subseasonal forecasting) with broad societal and economic applications, demonstrates large, measurable gains (doubling AI subseasonal skill; improvements across most targets), and is validated in an operationally relevant real-time ECMWF competition with top performance. The method (probabilistic bias correction) is deployable and broadly applicable to multiple forecast systems and variables. Paper 1 is novel for RL optimization theory/practice, but its impact is narrower and less externally validated.
Paper 1 demonstrates immediate, highly validated real-world impact by winning an international, real-time forecasting competition against leading operational centers. While Paper 2 offers an innovative approach to drug design, it currently relies on in-silico proxy evaluations. Paper 1's proven ability to drastically improve subseasonal forecasting skill provides broader, immediate, and concrete benefits for global agriculture, energy management, and disaster preparedness.
Paper 2 demonstrates higher scientific impact by addressing a grand challenge in meteorology: subseasonal weather forecasting. While Paper 1 provides valuable systems-level optimizations for edge LLMs, Paper 2 fundamentally advances earth sciences with transformative applications in disaster preparedness, agriculture, and energy. Its methodological rigor is exceptionally validated by winning the ECMWF 2025 real-time forecasting competition against top global operational centers. This proven ability to significantly improve forecasting skill guarantees broad, immediate impact across atmospheric science and crucial socio-economic sectors.
Paper 2 addresses a critical real-world problem (subseasonal weather forecasting) with demonstrated operational success, winning ECMWF's 2025 competition against 34 teams and six forecasting centers. Its impact spans agriculture, energy, and disaster preparedness with clear societal benefits. Paper 1 offers a valuable theoretical insight into low-rank optimizer training, but its impact is narrower, confined to ML systems research. Paper 2's combination of methodological rigor, validated real-world deployment, broad applicability, and timeliness gives it substantially higher estimated scientific impact.