Problem

Subseasonal weather forecasts, spanning two to six weeks ahead, are vital for critical decision-making in sectors like agriculture, water management, energy, and disaster preparedness. While current weather models, both physics-based dynamical and data-driven artificial intelligence (AI), offer high accuracy for predictions up to two weeks, their skill diminishes rapidly at these longer subseasonal timescales. This decline is primarily due to the accumulation of errors over many timesteps and persistent systemic biases, creating what has been termed a "predictability desert." Furthermore, traditional deterministic forecasts struggle to capture the inherent uncertainty at these horizons, highlighting the need for robust probabilistic forecasting methods that predict distributions over future weather states.

Method

This research introduces Probabilistic Bias Correction (PBC), a novel machine learning framework designed to substantially reduce systematic errors in subseasonal forecasts by learning from historical probabilistic predictions. Unlike approaches that correct raw observations, PBC operates directly on the space of probability distributions, allowing for optimization of not just the forecast center but also its spread and shape. The framework converts an input ensemble into an initial probabilistic forecast using climatological quantile bins. It then employs two computationally efficient machine learning models in parallel: Debias++, which applies location-, date-, and quantile-specific corrections learned by minimizing probabilistic forecasting error over adaptively selected training periods, and Persistence++, which incorporates recent weather trends and climatological shifts by blending the forecast with lagged observations. The outputs from both Debias++ and Persistence++ are then projected onto the space of valid cumulative distribution functions (CDFs) using isotonic regression to ensure monotonicity, and finally averaged to form the robust PBC prediction. PBC is designed for practical operational deployment, featuring incremental retraining daily with observable historical data.

Results

The application of PBC consistently yielded significant skill gains across various state-of-the-art subseasonal forecasting models. When applied to the ECMWF dynamical ensemble, PBC tripled positive precipitation skill and transformed previously unskillful temperature and mean sea level pressure forecasts into accurate predictions, demonstrating superiority over existing operational debiasing methods. For instance, skill gains were observed in 98% of grid cells for precipitation. PBC also notably improved the skill of leading AI models, such as ECMWF’s AI Forecasting System (AIFS-SUBS), tripling its positive skill for temperature in week 4 and precipitation in week 3, and doubling it for mean sea level pressure in week 4. Furthermore, it enhanced the performance of the hybrid PoET model, which combines AI and dynamical forecasts, with RPSS gains of 70–96% for precipitation. In the 2025 ECMWF AI Weather Quest, an international real-time forecasting competition, an ensemble combining PBC-ECMWF and PBC-PoET (named MicroDuet) secured first place for all weather variables and lead times, outperforming numerous operational centers and competitors worldwide. Crucially, PBC also significantly boosted the accuracy of extreme weather event predictions, leading to over 100% gains in Brier Skill Score for precipitation extremes and consistently improved flood forecasting skill, demonstrating its utility in early warning systems.

Implications

The Probabilistic Bias Correction (PBC) framework represents a significant advancement in subseasonal weather forecasting by consistently improving both leading dynamical and AI models. This establishes a new state of the art for hybrid and fully data-driven forecasting systems. The enhanced ability to detect extreme weather events early is particularly impactful, offering the potential to mitigate needless deaths and economic losses in the face of increasing climate change impacts. PBC is highly suitable for operational deployment due to its cost-effectiveness and adaptability, allowing meteorological centers, including those with limited resources, to upgrade their existing forecast products with minimal overhead. By targeting predictable errors rather than relearning atmospheric dynamics, PBC serves as an effective post-processing tool. The release of its open-source code aims to provide the community with a standardized instrument to bridge the gap between raw model output and actionable intelligence for various critical downstream applications, including agriculture, energy production, healthcare, and disaster response.