| 10 Feb 2026
Subseasonal Precipitation Prediction in the Yangtze River Delta Region Using an Adaptive Bias Correction Method
Abstract. Accurate and reliable prediction of subseasonal precipitation is crucial for disaster prevention and mitigation, particularly in the densely populated and economically developed Yangtze River Delta (YRD) region. Despite substantial progress in subseasonal-to-seasonal prediction (S2S) since the launch of the World Meteorological Organization’s S2S program, global models still exhibit considerable biases in the YRD region. To address these systematic errors in 3 to 6-week precipitation forecasts, this study employs an Adaptive Bias Correction (ABC) machine learning correction method to post-process biweekly accumulated precipitation predictions from the Climate Forecast System version 2 (CFSv2) and European Centre for Medium-Range Weather Forecasts Integrated Forecasting System (ECMWF) models. Evaluation against the 2021 CLDAS high-resolution gridded dataset reveals substantial improvements after ABC correction. For the CFSv2 model, the root mean square error (RMSE) is reduced by 3.79 mm (5.5 %) for weeks 3–4 and by 8.32 mm (12.6 %) for weeks 5–6. The ECMWF model shows an even larger RMSE reduction of 10.26 mm (15 %) across weeks 3–6, with wintertime errors decreasing by up to 35.7 %. Spatial pattern correlation and overall forecast skill are also markedly enhanced. For high-impact low-temperature rain/snow and freezing rain events, the ABC method reduces absolute bias in the CFSv2 model by 21.1 %–33.3 % and in the ECMWF model by 10.4 %–20.5 %, effectively eliminating the pronounced wet bias over the Hunan–Jiangxi–Fujian area. These improvements provide critical technical support for operational early warning of freezing rain disasters in the YRD region.
General Assessment
This manuscript presents an application of the Adaptive Bias Correction (ABC) method—a machine learning–based post-processing framework—to improve subseasonal (3–6 week) precipitation forecasts from two global models (CFSv2 and ECMWF) over the Yangtze River Delta (YRD) region. The authors evaluate the method using high-resolution CLDAS observations and report substantial improvements in RMSE, anomaly correlation, and forecast skill, particularly for winter and freezing rain events.
While the study is relevant and the ABC method is reasonably well implemented, the manuscript suffers from several critical weaknesses in terms of methodological clarity, validation design, and alignment with the scope of AMT. These issues must be addressed before the manuscript can be considered for publication.
Major Concerns
Reframe the manuscript as a validation and intercomparison study of bias correction techniques, or
Clearly articulate what novel methodological insight is contributed beyond the original ABC paper.
Resolution mismatch: Forecasts at 1.5° are bilinearly interpolated to 0.0625° to match CLDAS. This is statistically problematic and artificially inflates spatial detail. The authors acknowledge this as a limitation but do not assess its impact on skill metrics. A scale-consistent verification (e.g., upscaling observations to forecast resolution) should be performed as a sensitivity test.
Evaluation period: The study evaluates only one year (2021). This is insufficient to demonstrate robustness, especially given interannual variability in monsoon and winter precipitation regimes. The authors should either extend the evaluation period or explicitly frame the study as a proof-of-concept with clear caveats.
Metrics: The definition of "skill" in Eq. (1) is unconventional and not clearly distinguished from ACC in Eq. (3). The use of non-centered anomaly correlation without standardization is not standard practice in forecast verification. The authors should adopt widely used metrics (e.g., anomaly correlation coefficient, Brier score, ranked probability score) and justify any deviations.
The three components (Dynamical++, Climatology++, Persistence++) are described at a highly conceptual level. Critical details—such as hyperparameters, training window selection, feature engineering, and ensemble weighting—are missing.
No cross-validation or out-of-sample testing strategy is described. Given the adaptive nature of the method, it is unclear how overfitting is avoided.
The authors state that models are trained on "past precipitation and forecast data" but do not specify the training period, data partitioning, or how the "rolling" update is implemented.
Statements such as “the ABC method improved the spatial correlation coefficient by 60%” are misleading when baseline values are low (e.g., from 0.25 to 0.40). Relative improvements should be interpreted with caution, and absolute improvements should be emphasized.
No statistical significance testing is performed. Given the short evaluation period (2021), it is impossible to determine whether the improvements are robust or due to chance. The authors should apply bootstrapping or permutation tests to assess significance.
Table 1 is unreadable in its current format. The columns are not clearly labeled, and the structure is confusing. A reformatted table with clear separation of lead times, models, and seasons is urgently needed.
Figures 7–9 show case studies but lack quantitative performance metrics for those specific events (e.g., bias, RMSE, correlation). Subjective visual comparison is insufficient.
Figure 2 is a flowchart that adds little value. It should either be expanded to include methodological details or removed.
Minor Issues