Comparative Evaluation of ecPoint and Local EMOS for CMA-GEPS Precipitation Forecast over Eastern China

Comparative Evaluation of ecPoint and Local EMOS for CMA-GEPS Precipitation Forecast over Eastern China | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Short Report Comparative Evaluation of ecPoint and Local EMOS for CMA-GEPS Precipitation Forecast over Eastern China Stejik Sonum, Pu Liu, Wang Jialing, Wang Yong This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8883798/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Post-processing of numerical weather prediction (NWP) models constitutes a pivotal link in enhancing forecast performance. Despite their recognition as cutting-edge point-based post-processing techniques, systematic comparative evaluations of ecPoint (ECWMF for point forecasts) and Ensemble Model Output Statistics (EMOS)—particularly assessments of their applicability outside Europe and to Chinese ensemble forecasting systems—remain insufficient. In this study, we evaluate two advanced post-processing techniques—Ensemble Model Output Statistics (EMOS) and the point-based European Centre for ECMWF statistical ensemble method (ecPoint)—for calibrating ensemble precipitation forecasts. A comprehensive assessment of the performance of these ensemble post-processing methods is conducted using the China Meteorological Administration’s Global Ensemble Forecasting System (CMA-GEPS) forecast over eastern China. The results demonstrate that both methods significantly reduce systematic biases and improve the reliability and dispersion of ensemble forecasts. Notably, improvement in forecast accuracy is observed even under convective weather conditions, and early warnings of extreme precipitation events. Overall, while both methods show comparable performance, they exhibit distinct behaviors across different regions. The ecPoint method slightly outperforms EMOS in terms of Continuous Ranked Probability Score (CRPS) and provides improved resolution and early warning capabilities at various precipitation thresholds. NWP Post-processing Ensemble forecast Probability prediction ecPoint EMOS Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 1 Introduction Precipitation forecasts play a crucial role in weather forecasting, yet improving their accuracy remains a major challenge in both operational meteorology and scientific research. Numerical Weather Prediction (NWP) models are essential for precipitation forecasts. Due to the inherently chaotic nature of the atmosphere and errors in the initial conditions and model physics in NWP, NWP forecasts are subject to uncertainties. Consequently, ensemble forecasting is necessary to provide uncertainty estimates in precipitation forecasts (Lorenz 1963; Astakhova 2021). Over recent decades, precipitation forecasting has gradually shifted from deterministic, single-model predictions toward ensemble-based probabilistic approaches to quantify uncertainty (Gneiting 2005; Hemri 2014; Dai et al. 2018; Palmer 2002). Nowadays, the most advanced probabilistic forecasts rely on ensemble NWP systems (Hemri et al. 2022), which form the core of modern probabilistic precipitation and hydrometeorological forecasting (Andrade et al. 2024; Liu et al. 2022; Yadav et al. 2022; Li et al. 2017). With advances in ensemble forecasting theory and technology, substantial progress has been made in operational ensemble prediction worldwide. Nevertheless, raw ensemble outputs often exhibit systematic biases and insufficient spread, making statistical post-processing necessary to improve forecast accuracy and reliability (Chen et al. 2024; Hamill et al.1998). As a result, post-processing techniques have become an integral component of operational workflows in many national meteorological services, providing more accurate, automated, and seamless forecast products for users and the public (Vannitsem et al. 2021; Whan et al.2018). Numerous post-processing methods have been proposed in recent years and can be broadly classified as parametric or non-parametric (Vannitsem et al. 2021). Key parametric approaches include Bayesian Model Averaging (Sloughter et al. 2010; Baran et al. 2014) and Ensemble Model Output Statistics (EMOS) (Gneiting et al. 2005; Baran and Lerch 2016; Scheuerer and Hamill 2015), while common non-parametric methods include Quantile Mapping (Madadgar et al. 2014) and Quantile Regression Forests (Meinshausen et al. 2006). EMOS and BMA generally show comparable skill for probabilistic precipitation forecasts, with each having distinct advantages (Javanshiri et al. 2021; Han et al. 2018). Several studies have explored EMOS formulations tailored to precipitation. Baran et al. (2016) applied the Censored-Shifted Gamma (CSG) distribution within the EMOS framework and found it outperformed Generalised Extreme Value (GEV)-based EMOS and Gamma-based BMA. More recently, Angus et al. (2024) evaluated EMOS_CSG using ECMWF dual-resolution 24-hour precipitation forecasts across Europe and showed that EMOS_CSG substantially enhanced forecast skill, approaching the performance of more advanced quantile-mapping techniques while requiring no additional historical data. However, EMOS implicitly maps grid-cell average precipitation to point locations (Hemri S 2022), which may limit performance under highly sub-grid variability conditions (Hewson and Pillosu 2021). To address the limitations of previous post-processing techniques, the ECMWF developed ecPoint, an operational post-processing system specifically designed for point precipitation forecasting (Hewson and Pillosu 2021). ecPoint is a non-parametric technique that accounts for sub-grid weather variability, grid-scale biases, and regime-dependent relationships. Notably, it requires only one year of global historical data to generate reliable point-wise probabilistic forecasts (Hewson and Pillosu 2021). Hemri and Hewson (2022) compared ecPoint with EMOS for 12-hour precipitation over Switzerland and surrounding areas and found that ecPoint outperformed EMOS at longer lead times and for heavy precipitation events. Gascón et al. (2024) demonstrated that ecPoint improves ECMWF maximum-precipitation forecasts and better identifies severely affected areas across Europe. Pillosu et al. (2024) validated ecPoint against one year of observations in Ecuador, showing strong capability for identifying flash-flood-prone areas associated with small-scale convective systems. While Hemri et al. (2022) demon strated the advantages of ecPoint over Global EMOS, it remains an open question whether these advantages hold against a locally optimized EMOS (Local EMOS) in regions with topographical variance. Local EMOS is capable of correcting stationary local biases; however, we hypothesize that it lacks the physical mechanism to represent the transient, sub-grid convective extremes characteristic of the East Asian Monsoon. Therefore, this study compares ecPoint against a strict Local EMOS benchmark to determine if statistical localization is sufficient, or if the physics-based sub-grid mapping of ecPoint is required for Eastern China. Through comparative evaluation, this study provides a theoretical foundation for the further development of ensemble forecast post-processing techniques and their promotion in practical applications. The remainder of this paper is organized as follows: Section 2 describes the study area, data sources, and the principles of the ecPoint and EMOS methods, along with the verification methodology. Section 3 presents a comprehensive evaluation of the baseline GEPS forecasts and the two post-processing techniques, including probabilistic scores, deterministic measures, and representative case studies. Section 3 provides the conclusions and discussion. 2 Data and Method 2.1 Research Area and datas The study area encompasses eastern China, bounded by latitudes 25.25°–36.25°N and longitudes 114.0°–124.0°E (Fig. 1 ), and includes the Yangtze River Delta region. The region’s topography is highly varied, including both coastal and inland areas. It experiences pronounced seasonal variations and is characterized by a subtropical monsoon climate, with additional influences from temperate monsoons, and is situated at the intersection of northern and southern climatic zones and at the interface between terrestrial and marine environments (Ding 2013). Precipitation is concentrated in the summer months, while winters may bring substantial snowfall. The core factors influencing summer weather in eastern China—the Western Pacific Subtropical High (WPSH), the Meiyu front, and frequent typhoon landfalls—frequently produce severe weather, causing plum-rain floods and typhoon-induced inundations. Prominent administrative provinces within the study area include Jiangsu, Zhejiang, Shanghai, and Anhui, collectively known as the “Region of Rivers and Lakes.” Beyond its complex climate and geography, this area is one of the most densely populated in China, amplifying the socio-economic consequences of extreme weather events. The China Meteorological Administration's Global Ensemble Forecast System (CMA-GEPS) is built on the domestically developed GRAPES (Global/Regional Assimilation and Prediction System) model (Huo et al.2020; Gao et al. 2022; Chen et al. 2003; Shen et al. 2020). Initial perturbations are generated using singular vector techniques (Li et al. 2019), and uncertainty in the model's physical tendencies is represented through the Stochastically Perturbed Parameterization Tendencies (SPPT) and Stochastic Kinetic Energy Backscatter (SKEB) schemes (Yuan et al. 2016; Peng et al. 2019). 2.2 Description of the dataset This study employs the GEPS system, which reports data starting at 12:00 UTC. This system, with 30 ensemble members and features a horizontal resolution of 50×50 km. The training dataset for model calibration spans May 19, 2020, to May 18, 2023 (a total of 3 years). The test dataset for precipitation forecasts covers June 1 to August 31, 2023, with forecast lead times ranging from 12 to 84 hours, utilizing cumulative precipitation forecast products at 12-hour intervals. Hourly precipitation data from 15,670 meteorological stations within the study area were selected as observational data. For the ecPoint method, the following control forecasts were chosen as forecast factors: total precipitation (tp), convective precipitation (cp), the U-component of the 700 hPa wind field (u700), 700 hPa wind field V-component (v700), and convective available potential energy (cape) as forecast factors. The training process employed a 12-hour forecast lead time. The EMOS_CSG technique utilizes ensemble forecasts of TP as forecast factors, with training data covering seven forecast lead times ranging from 12 to 84 hours. 2.3 Post-processing method In this study, we explore the ecPoint and EMOS methods for precipitation post- processing. A brief overview of these two techniques is presented below, with detailed formulations provided in the supplementary material. The ecPoint method is a station-based statistical post-processing system developed at ECMWF (Pillosu and Hewson 2017; Hewson and Pillosu 2021). It is also a point-scale scheme based on decision-tree modelling that explicitly accounts for sub grid-scale variability in NWP precipitation fields. Rather than treating grid-box mean precipitation as representative of point measurements, ecPoint conditions on a set of flow-dependent predictors—such as model resolution, orography, convective environment and precipitation regime—to derive statistical transfer functions between grid-scale precipitation and its conditional distribution at observation sites within each grid cell. These transfer functions are obtained by estimating conditional probability distributions from large training archives of co-located model forecasts and rain-gauge observations, and are then applied in real time to transform raw model output into calibrated, point-scale probabilistic precipitation forecasts (Hewson and Pillosu 2021). The Ensemble Model Output Statistics (EMOS) method is an extension of the traditional Model Output Statistics (MOS) approach for quantitative forecasting. Building on MOS, EMOS provides a parameterized post-processing framework that links ensemble forecasts to a chosen predictive probability distribution. The method first specifies an appropriate probability distribution for the forecast variable and then uses a link function to relate the distribution parameters to the ensemble forecasts. The parameters are estimated by an optimization procedure and used for correcting forecast biases (Gneiting et al., 2005; Baran et al., 2016). For continuous variables such as temperature or wind speed, EMOS typically assumes a normal or truncated normal distribution. However, for precipitation—which is non-negative, highly skewed, and has a point mass at zero—the normal distribution is unsuitable. Instead, EMOS adopts a left-Censored, Shifted Gamma (CSG) distribution to model precipitation. The CSG distribution accommodates continuous values that can be positive or zero, with left-censoring at zero (Scheuerer and Hamill 2015). This formulation underpins the CSG-based EMOS model proposed by Baran and Nemoda (2016) (Scheuerer 2015). 3 Results 3.1 Verification This study assesses the probabilistic performance of the raw ensemble and its post-processed counterparts using ecPoint and EMOS through a systematic comparative analysis. For continuous variables, overall probabilistic skill is quantified using the Continuous Ranked Probability Score (CRPS) and the Continuous Ranked Probability Skill Score (CRPSS). For event-based forecasts, the Brier Score (BS) and receiver operating characteristic (ROC) curves, summarized by the area under the curve (AUC), are employed to evaluate reliability and discrimination. Additionally, reliability diagrams and sharpness histograms are used to diagnose forecast calibration and spread. The training dataset spans three years. For scoring, we used forecasts from the validation dataset initialized at 12:00 UTC during summer 2023 with lead times of 12–84 hours; post-processed outputs from all stations within the study domain were included. Boxplots of CRPS show ecPoint and EMOS (Figure S1 ) significantly outperform the raw GEPS at all lead times. ecPoint has the lowest mean CRPS, while EMOS yields the most stable (low median) errors; the raw GEPS shows the highest means and extremes, indicating larger biases and ensemble spread. All methods exhibit a 12‑hour periodicity, and CRPS variability increases with lead time. To explore the spatial heterogeneity of correction performance between the two post-processing techniques across stations, the Continuous Ranked Probability Skill Score (CRPSS) was computed for each station using the raw GEPS as the reference. On average, EMOS and ecPoint improve forecast skill by approximately 10% and 26%, respectively (Figs. 2 a–b). EMOS outperforms the raw model at nearly all stations except parts of southern China and localized areas in Shandong. ecPoint exhibits positive CRPSS values at nearly all stations, demonstrating robust and widespread improvement. The relative CRPSS using EMOS as the baseline (Fig. 2 c) highlights clear spatial contrasts: EMOS performs better in central Anhui, southern Zhejiang, and mountainous/coastal regions of Fujian and Jiangxi, while ecPoint is more effective in transition-terrain zones, valleys (e.g., the Fuzhou Basin), plains, and high-elevation areas of southern southern Anhui and Shandong. Lead-time-dependent CRPSS analyses (Fig. 2 d) show positive medians for both methods, with ecPoint exhibiting increasing skill at longer lead times. For precipitation thresholds of > 0.0 mm, 0.1 mm, and 5.0 mm (Figs. 3 a–c), both post-processing techniques yield substantially lower Brier Scores (BS) relative to GEPS, and ecPoint shows the greatest improvement. At higher thresholds (≥ 10.0–30.0 mm), the performance of ecPoint and EMOS gradually converges. EMOS slightly outperforms ecPoint at isolated lead times. Across all thresholds, the most substantial BS reduction—indicating maximum improvement—occurs at the 84-hour lead time, implying that post-processing effectively extends the useful forecast horizon. Aggregated across all lead times (Fig. 3 d), BS decreases monotonically with increasing threshold for all methods, and both ecPoint and EMOS consistently outperform GEPS. However, at extreme thresholds (> 25.0–100.0 mm), improvement diminishes and may approach zero due to limited sample size. The reliability diagrams (Fig. 4 ) show that ecPoint and EMOS greatly improve reliability across thresholds. ecPoint generally lies closest to the diagonal, while GEPS deviates most strongly, especially at high thresholds. EMOS typically performs between the other two. At thresholds > 0.1 mm low-information intervals (e.g., 0–0.2 mm), both corrected forecasts tend to be under-confident because of influence from the raw ensemble. At medium-to-high confidence intervals, EMOS often becomes over-confident. At thresholds ≥ 10 mm, both methods exhibit reduced reliability, with EMOS deviating more than ecPoint. GEPS becomes increasingly conservative at thresholds > 15.0 m. Sharpness plots (Fig. 4 , upper left) show a U-shaped distribution for GEPS, while EMOS demonstrates higher sharpness than ecPoint at lower thresholds but poorer reliability, explaining its higher BS. As thresholds increase, sharpness profiles of all models converge; at thresholds > 15.0–70.0 mm, all three cluster near zero. Together with the reliability results, this indicates strong performance of both post-processing methods for high-threshold precipitation (details in Figure S3). AUC-threshold plots (Fig. 5 ) confirm that all methods perform reasonably at low thresholds (≤ 5.0 mm/12h). However, GEPS deteriorates rapidly at thresholds above 25.0 mm. EMOS shows noticeable degradation above 50.0 mm, whereas ecPoint maintains the highest AUC across most thresholds, despite reduced curve smoothness beyond 70.0 mm. For distinguishing heavy and extreme precipitation (e.g., torrential rain), ecPoint or EMOS should be prioritized. ROC curves and AUC values (Figure S2) show that both ecPoint and EMOS enhance event discrimination across thresholds. At thresholds above 0.1 mm, both methods outperform GEPS. As threshold increases, GEPS progressively approaches the no-skill line, while EMOS and ecPoint maintain clear positive curvature. EMOS’s ROC smoothness declines beyond 50.0 mm, and ecPoint’s smoothness weakens past 70.0 mm, due to sparse samples. AUC values remain relatively stable with lead time, showing limited sensitivity from 24 to 84 hours. Combining AUC and reliability, both post-processing techniques significantly improve prediction of heavy and extreme precipitation. 3.2 Case studies Typhoon Doksuri (26–28 July 2023) Figure 6 compares 12-hourly 85th-percentile forecasts against observations. GEPS reasonably captures precipitation occurrence during 0–24 hours, but misplaces heavy rainfall cores too far east. Both post-processing techniques markedly improve location and intensity forecasts. ecPoint provides the most accurate representation of extreme precipitation (> 200.0 mm/12 h) in Fujian during 36–48 hours, whereas GEPS and EMOS underestimate peak magnitudes at several stations. Rainstorm event associated with a quasi-stationary Meiyu front(19 Jun 2023) From 00:00 to 12:00 on 19 Jun 2023, under a quasi-stationary Meiyu front, moderate to heavy rainfall occurred along the river and in the southern areas of Jiangsu Province and in northern Jiangxi Province, with localized amounts reaching rainstorm intensity or above. The 36–48 h precipitation forecasts initialized at 12:00 on 17 June 2023 (Fig. 7 ) show that, at the 50th percentile, the raw model predicts an overly broad precipitation area, whereas the ecPoint and EMOS post-processed forecasts produce spatial patterns more consistent with observations. At the 90th percentile, ecPoint and EMOS better reproduce the location and intensity of rainstorms along the Yangtze River in Jiangsu and in the border region between northern Jiangxi and southern Anhui, while GEPS gives more diffuse and indistinct rainstorm forecasts there. Both post-processing methods, however, show a marked positive bias in precipitation amounts over northern Jiangsu and Shandong, which is more pronounced for ecPoint. Across diverse precipitation types and events, both post-processing techniques significantly enhance forecast accuracy, spatial continuity, and ensemble dispersion. ecPoint performs particularly well in capturing high-precipitation areas and convective events at high quantiles (e.g., ≥ 95th) but may overpredict heavy rainfall in some regions (e.g., Figure S4). EMOS compensates for light-precipitation under-forecasting but is prone to false positives over broader areas. Overall, both methods yield substantial improvement over the raw GEPS ensemble, with complementary strengths depending on threshold, elevation, and weather regime. 4 Conclusions and Discussion This study evaluates two post-processing techniques—ecPoint and EMOS—applied to the CMA-GEPS precipitation ensemble forecast system for Eastern China. Overall, both methods substantially improved the reliability and accuracy of the forecasts, while reducing the ensemble forecast’s systematic bias and enhancing the dispersion of the ensemble forecasts. ecPoint demonstrates the highest overall skill, particularly in terms of CRPS, and produces the most substantial improvement for light-precipitation forecasts. For heavy precipitation and above, the accuracy of ecPoint and EMOS becomes comparable, as reflected by their Brier Scores. Both methods substantially improve the calibration and resolution of forecast probabilities. ecPoint exhibits consistently stronger reliability across thresholds, while EMOS performs more stable at moderate thresholds but less reliable under heavy-rainfall conditions. Both post-processing methods improve performance for convective cases. ecPoint, in particular, exhibits strong early-warning skill for extreme precipitation, with enhanced discrimination at high quantiles. Spatial analysis indicates that the EMOS model performs better in certain high-elevation areas—particularly in the highly influenced terrain zones of southeastern and central East China, as well as most coastal regions. In contrast, the ecPoint model excels in river valleys, sloping terrain, and most plains (excluding Shandong Mountains and southern Anhui). Considering the limited topographical variability within the study area, the validation results pertaining to different geographical altitudes were not statistically significant and are therefore not discussed further in this paper. Moreover, the variations in scoring metrics across different forecast lead times were also marginal. Future research is warranted to explore these aspects in diverse geographical regions and with extended forecast lead times. The EMOS formulation used here employs a single forecast factor (total precipitation, tp) over the 12- to 84-hour lead times. This minimal factor set—combined with EMOS’s monotonic relationship between predictors and predictands—likely limits its performance and leaves room for further optimization. Conversely, ecPoint requires decision-tree training and thus uses only one time step (12-hour forecasts) but incorporates five forecast factors. The reduction in temporal depth may influence temporal sensitivity but enhances its capability to represent spatial variability and extreme-event structure. To ensure comparability with EMOS, ecPoint forecasts are bilinearly interpolated from grid to station. This interpolation increases spatial dispersion and can lead to more apparent false positives—especially at high thresholds with inherently low event frequency. As a result, ecPoint’s probability-based accuracy metrics (e.g., BS) may be degraded despite its strong reliability and resolution. In recent years, Artificial Intelligence techniques, have become a key focus in model post-processing technologies, achieving significant advancements in meteorology. These methods automatically learn complex data patterns, offering innovative solutions for ensemble forecast post-processing (Reichstein et al. 2019). To align with practical operational requirements, AI-based approaches are not within the scope of this study; however, future research will consider incorporating AI methods to further expand the research content based on the findings of this paper. Declarations Competing interests The authors declare that they have no competing interests. Supplements Supplement: The supplementary materials for this paper consist of three sections: the first part provides a detailed explanation of the post-processing methods and validation procedures; the second part contains supplementary result discussions; and the third part includes supplementary figures and explanations. Author Contribution CRediT authorship contribution-statementeSonum Steiik Data-curation, Formal analysis, Investigation, Software, Validation, Visualization,Writing-original-draft, Writing-eview & editing. LIU Pu: Data-curation, Investigation, Software, Writing-review-& editing.Wang-Jialing: Supervision, Writing-review-& editing. WANG Yong: Conceptualization, Data-curation, Funding acquisition, Supervision, Writing-review-& editing, Project administration. Acknowledgement The first author special thanks to Pillosu Fatima and Hewson Timothy of the ECWMF for sharing the ecPoint software and offering assistance. We are grateful to Ms Chen Jieyu for her guidance and support in revising this paper. We also extend our appreciation to Zhu Yanwei, Huo Ziqiang, Cao Bufan, Song Qianqian and Wang Mingming for their participation in discussions and valuable suggestions throughout the research process. In addition, the authors are grateful to the reviewers for their patient responses and thoughtful suggestions throughout the review process. References Andrade FS, Arsenault R, Poulin A, Troin M, Armstrong W. (2024) Application of weather post-processing methods for operational ensemble hydrological forecasting on multiple catchments in Canada. Journal of Hydrology 642: 131861. Angus M, Widmann M, Orr A, Ashrit R, Leckebusch GC, Mitra A. (2024) A comparison of two statistical postprocessing methods for heavy-precipitation forecasts over India during the summer monsoon. Quarterly Journal of the Royal Meteorological Society 150(761): 1865-83. Astakhova E, Alferov D, Alferov Y, Bundel A. (2021) Ensemble approach to weather forecasting. In Journal of Physics: Conference Series p012070. Bottazzi M, Rodríguez-Muñoz L, Chiavarini B, Caroli C, Trotta G, Dellacasa C, Marras GF, Montanari M, Santini M, Mancini M, D'Anca A. (2024) High performance computing to support land, climate, and user‐oriented services: The HIGHLANDER Data Portal. Meteorological Applications 31(2): e2166. Baran S, Lerch S. (2016) Mixture EMOS model for calibrating ensemble forecasts of wind speed. Environmetrics 27(2): 116 − 30. Baran S, Nemoda D. (2016) Censored and shifted gamma distribution based EMOS model for probabilistic quantitative precipitation forecasting. Environmetrics 27(5): 280–292. Baran S, Horanyi A, Nemoda D. (2014) Comparison of the BMA and EMOS statistical methods in calibrating temperature and wind speed forecast ensembles. Idojárás 118(3): 217–241. Chen J, Janke T, Steinke F, Lerch S. (2024) Generative machine learning methods for multivariate ensemble postprocessing. The Annals of Applied Statistics 18(1): 159–183. Dai K, Zhu YJ, Bi BG. (2018) The review of statistical post-process technologies for quantitative precipitation forecast of ensemble prediction system. Acta Meteorologica Sinica 76(4): 493–510(in Chinese). Gneiting T, Raftery A E. (2005) Weather forecasting with ensemble methods. Science 310(5746): 248–249. Gascón E, Montani A, Hewson T D. (2024) Post-processing output from ensembles with and without parametrised convection, to create accurate, blended, high‐fidelity rainfall forecasts. Quarterly Journal of the Royal Meteorological Society 150(762): 3117–3145. Grönquist P, Yao C, Ben-Nun T, et al. (2021) Deep learning for post-processing ensemble weather forecasts. Philosophical Transactions of the Royal Society A 379(2194): 20200092. Hewson T D, Pillosu F M. (2021) A low-cost post-processing technique improves weather forecasts around the world. Communications Earth & Environment 2(1): 132. Han K, Choi J T, Kim C. (2018) Comparison of statistical post-processing methods for probabilistic wind speed forecasting. Asia-Pacific Journal of Atmospheric Sciences 54(1): 91–101. Hamill T M, Colucci S J. (1998) Evaluation of Eta–RSM ensemble probabilistic precipitation forecasts. Monthly Weather Review 126(3): 711–724. Hemri S, Hewson T, Gascón E, et al. (2022) How do ecPoint precipitation forecasts compare with postprocessed multi-model ensemble predictions over Switzerland? European Centre for Medium Range Weather Forecasts. Hemri S, Scheuerer M, Pappenberger F, et al. (2014) Trends in the predictive performance of raw ensemble weather forecasts. Geophysical Research Letters 41(24): 9197–9205. Hersbach H, Bell B, Berrisford P, et al. (2020) The ERA5 global reanalysis. Quarterly journal of the royal meteorological society 146(730): 1999–2049. Javanshiri Z, Fathi M, Mohammadi S A. (2021) Comparison of the BMA and EMOS statistical methods for probabilistic quantitative precipitation forecasting. Meteorological Applications 28(1): e1974. Liu X, Zhang L, She D, et al. (2022) Postprocessing of hydrometeorological ensemble forecasts based on multisource precipitation in Ganjiang River basin, China. Journal of Hydrology 605: 127323. Li W, Duan Q, Miao C, et al. (2017) A review on statistical postprocessing methods for hydrometeorological ensemble forecasting. Wiley Interdisciplinary Reviews: Water 4(6): e1246. Lorenz E N. (1963) The mechanics of vacillation. Journal of Atmospheric Sciences 20(5): 448–465. Meinshausen N, Ridgeway G. (2006) Quantile regression forests. Journal of machine learning research 7(6). Madadgar S, Moradkhani H, Garen D. (2014) Towards improved post-processing of hydrologic forecast ensembles. Hydrological Processes28(1): 104–122. Pillosu F M, Bucherie A, Kruczkiewicz A, et al. (2024) Can global rainfall forecasts identify areas at flash flood risk? Proof of concept for Ecuador. ECMWF Technical Memo (917). Pillosu F, Hewson T. (2017) New point-rainfall forecasts for flash flood prediction. ECMWF Newsletter 153: 2–3. Palmer, T. N. (2002) The economic value of ensemble forecasts as a tool for risk assessment: From days to decades. Quarterly Journal of the Royal Meteorological Society: A journal of the atmospheric sciences, applied meteorology and physical oceanography 128(581): 747–774. Reichstein M, Camps-Valls G, Stevens B, et al. (2019) Deep learning and process understanding for data-driven Earth system science. Nature 566(7743): 195–204. Scheuerer M, T. M. Hamill. (2015). Statistical Postprocessing of Ensemble Precipitation Forecasts by Fitting Censored, Shifted Gamma Distributions. Mon. Wea. Rev 143: 4578–4596. Raftery A E, Gneiting T, Balabdaoui F, et al. (2005) Using Bayesian model averaging to calibrate forecast ensembles. Monthly weather review 133(5): 1155–1174. Sloughter J M L, Gneiting T, Raftery A E. (2010) Probabilistic wind speed forecasting using ensembles and Bayesian model averaging[J]. Journal of the american statistical association 105(489): 25–35. Vannitsem S, Bremnes J B, Demaeyer J, et al. (2021) Statistical postprocessing for weather forecasts: Review, challenges, and avenues in a big data world. Bulletin of the American Meteorological Society 102(3): E681-E699. Wilks D S. (2011) Statistical methods in the atmospheric sciences. Academic press, USA Whan K, Schmeits M. (2018) Comparing area probability forecasts of (extreme) local precipitation using parametric and machine learning statistical postprocessing methods. Monthly Weather Review 146(11): 3651–3673. Yadav R, Yadav S M. (2022) Review on Statistical Post-processing of Ensemble Forecasts. In International Conference Innovation in Smart and Sustainable Infrastructure. pp469-476. Shen Xueshun, Wang Jianjie, Li Zechun, Chen Dehui, Gong Jiandong. (2020) China's independent and innovative development of numerical weather prediction. Acta Meteorologica Sinica 78(3): 451–476(in Chinese). Chen Jing, Zhu Yuejian, Duan Wansuo, Zhi Xiefei, Min Jinzhong, Li Xiaoli, Deng Guo, Yuan Huiling, Feng Jie, Du Jun, Li Qiaoping, Gong Jiandong, Shen Xueshun, Mu Mu. (2025) A review on development, challenges, and future perspectives of ensemble forecast. Acta Meteorologica Sinica 83(3): 480–502(in Chinese). Chen Jing, Xue Jishan, Yan Hong. (2003) Uncertainty and Ensemble Forecasting Experiment of Mesoscale Rainstorm Numerical Forecast in South China. Acta Meteorologica Sinica (04): 432–446(in Chinese). Xiaoli LI, Yongzhu LIU. (2019) The improvement of GRAPES global extratropical singular vectors and experimental study[J]. Acta Meteorologica Sinica 77(3): 552–562(in Chinese). Yuan Y, Li X L, Chen J, et al. (2016) Stochastic parameterization toward model uncertainty for the GRAPES mesoscale ensemble prediction system. Meteor Mon 42(10): 161–1175(in Chinese). Peng F, Li X L, Chen J et al. (2019) A stochastic kinetic energy backscatter scheme for model perturbations in the GRAPES global ensemble prediction system. Acta Meteor Sinica 77(2): 180–195(in Chinese). GAO Li, ZHENG Jiawen, ZHAO Zuosen, LUO Yuelin, REN Pengfei, YAO Guohua. (2022) Research, Development, and Application of the Unified Post-Processing System for the CMA-GEPS/REPS Ensemble Prediction. Advances in Earth Science[J] 37(12): 1211–1222(in Chinese). HUO Zhenhua, LIU Yongzhu, CHEN Jing, LI Xiaoli, QU Anxiang. (2020) The preliminary appliation of tropical cyclone targeted singular vectors in the GRAPES global ensemble forecasts[J]. Acta Meteorologica Sinica 78(1): 48–59 (in Chinese). Ding Y H. (2013) Climate of China[M]. Beijing: China Meteorological Press(in Chinese). Additional Declarations No competing interests reported. Supplementary Files SupplementforEvaluatingecpointandEMOS.pdf Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8883798","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Short Report","associatedPublications":[],"authors":[{"id":599832700,"identity":"b33aa2cb-c2c3-4473-92da-2daf65995967","order_by":0,"name":"Stejik Sonum","email":"","orcid":"","institution":"Nanjing University of Information Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Stejik","middleName":"","lastName":"Sonum","suffix":""},{"id":599832701,"identity":"77c8c35b-36ce-410f-a4ae-5d3def5e8f62","order_by":1,"name":"Pu Liu","email":"","orcid":"","institution":"Nanjing Meteorological Bureau","correspondingAuthor":false,"prefix":"","firstName":"Pu","middleName":"","lastName":"Liu","suffix":""},{"id":599832702,"identity":"7e1ff168-20f0-4cdb-9e15-f7d15d551e27","order_by":2,"name":"Wang Jialing","email":"","orcid":"","institution":"Nanjing University of Information Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Wang","middleName":"","lastName":"Jialing","suffix":""},{"id":599832704,"identity":"5654afa4-b934-4413-a89e-bdb3a9749bf2","order_by":3,"name":"Wang Yong","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA8UlEQVRIiWNgGAWjYFAD9gYGhgQwK4FYLTwHQIoNSNEiAVZJhBaD472HXzC22eXJR749JvHgzx8GfvYcA4afO/BoOXMuzYKxLbnY8HZemkRimwGDZM8bA8beM3i03MgxM2Dcxpy4cXaOmURigwFIxICZsY2glvrEjTPPmEkk/DFgsCdCi/EDxm2HE+dL8AC1sAFtkSCgRfLMGTOGxH/HEzfw5BhbJLYZ80iceVZwsBePFr7jPcYfPpypTpzffsbw5o8/cnL87ckbH/zEo0XhAAMbOEYMDkAEeEDEAdwaGBjkGxiYP0AZo2AUjIJRMAqwAwC9J1H73YkULAAAAABJRU5ErkJggg==","orcid":"","institution":"Nanjing University of Information Science and Technology","correspondingAuthor":true,"prefix":"","firstName":"Wang","middleName":"","lastName":"Yong","suffix":""}],"badges":[],"createdAt":"2026-02-15 05:23:24","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8883798/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8883798/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":103988697,"identity":"ae8c82f0-7e8b-488d-b7ae-b3d684598823","added_by":"auto","created_at":"2026-03-05 11:02:24","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":599223,"visible":true,"origin":"","legend":"\u003cp\u003eSpatial Distribution and Elevation in the Study Area.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8883798/v1/9d67524a38362e547747268c.png"},{"id":103988689,"identity":"2aae29f6-b4cd-4c10-8671-cebe5ae5cda7","added_by":"auto","created_at":"2026-03-05 11:02:23","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":1067060,"visible":true,"origin":"","legend":"\u003cp\u003e(a) and (b) show the CRPSS of EMOS and ecPoint with GEPS as the reference, respectively. (c) shows the CRPSS values between EMOS_CSG (with its CRPS as reference) and ecPoint. (d) shows the improvement of ecPoint and EMOS_CSG over GEPS in terms of CRPS across different forecast lead times. For the box plots, the numbers at the bottom represent the mean values (white circles in the box plots), the thick black lines represent the median values (numbers above the box plots represent the median), and the horizontal lines at both ends represent the extreme values. Positive (negative) CRPSS values indicate better (worse) performance than the reference model(mm/12h).\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8883798/v1/6ea72e0eb97722bd1ba9b44b.png"},{"id":103988693,"identity":"3cddb297-4c97-4578-be48-3c7206ebac2f","added_by":"auto","created_at":"2026-03-05 11:02:23","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":423513,"visible":true,"origin":"","legend":"\u003cp\u003eThe BS scores at all forecast leadtimes for thresholds of 0.0 mm, 0.1 mm, and 5.0 mm are represented by a, b, and c respectively, and d represents the BS scores at all forecast lead times for different thresholds. The top number of the box plot is the mean (the white circle in the box plot), the median is the thick black line, and the horizontal lines at both ends are the extreme values (mm/12h).\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8883798/v1/ea0e099ea7ea7ca962407525.png"},{"id":104402210,"identity":"22b63b92-315b-4dfb-b21a-74d1ee8c4964","added_by":"auto","created_at":"2026-03-11 12:14:40","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":535175,"visible":true,"origin":"","legend":"\u003cp\u003eReliability diagrams and Sharpness profiles of the raw model and two post-processing approaches at diverse thresholds (mm/12h).\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8883798/v1/8d377cffaa936d690929bf7f.png"},{"id":104402086,"identity":"cabed031-6730-4e4c-8905-10f28fd60123","added_by":"auto","created_at":"2026-03-11 12:14:14","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":249951,"visible":true,"origin":"","legend":"\u003cp\u003eBox plots of AUC at all forecast lead times (12-84 hours) for different thresholds; the bottom number of the box plot is the mean (indicated by the white circle in the box plot), the median is the bold black line, and the horizontal lines at both ends represent the extreme values (mm/12h).\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8883798/v1/23865874801cf25ca0999af8.png"},{"id":104401814,"identity":"012106d9-b9c7-43a1-9538-0bcdf784c174","added_by":"auto","created_at":"2026-03-11 12:13:37","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":811611,"visible":true,"origin":"","legend":"\u003cp\u003e85th Percentile Precipitation Forecasts and Observations valid for 0-12h(a,h,o,v), 12-24h(b,I,p,w), 24-36h(c,j,q,x), and 36-48h(d,k,r,y) issued at 12:00 UTC on 26 July 2023 (mm/12h).\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-8883798/v1/33ce696edfbe01a9738437a5.png"},{"id":103988694,"identity":"0665be52-062d-4766-a0c5-b40130664d51","added_by":"auto","created_at":"2026-03-05 11:02:23","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":1161540,"visible":true,"origin":"","legend":"\u003cp\u003eIssued at 12:00 UTC on 17 Jun 2023, forecasting 12-hour precipitation for the 50th, 75th, 85th, and 90th percentiles valid for the 36-48h period. Observations are shown at 12:00 on Jun 19, 2023(mm/12h).\u003c/p\u003e","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-8883798/v1/caac67f006ac4c8f384ed4e1.png"},{"id":105035282,"identity":"8b764ebd-804f-4bef-89a5-969ea590ae0f","added_by":"auto","created_at":"2026-03-20 07:25:47","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":5325471,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8883798/v1/50a9217e-d5e6-4c9c-94f6-41c9fc55b11b.pdf"},{"id":104402351,"identity":"e9f688b2-def5-4968-ad65-d378cdf7be50","added_by":"auto","created_at":"2026-03-11 12:15:08","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":1250055,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementforEvaluatingecpointandEMOS.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8883798/v1/37234972304b6191e2d08929.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Comparative Evaluation of ecPoint and Local EMOS for CMA-GEPS Precipitation Forecast over Eastern China","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003ePrecipitation forecasts play a crucial role in weather forecasting, yet improving their accuracy remains a major challenge in both operational meteorology and scientific research. Numerical Weather Prediction (NWP) models are essential for precipitation forecasts. Due to the inherently chaotic nature of the atmosphere and errors in the initial conditions and model physics in NWP, NWP forecasts are subject to uncertainties. Consequently, ensemble forecasting is necessary to provide uncertainty estimates in precipitation forecasts (Lorenz 1963; Astakhova 2021).\u003c/p\u003e \u003cp\u003eOver recent decades, precipitation forecasting has gradually shifted from deterministic, single-model predictions toward ensemble-based probabilistic approaches to quantify uncertainty (Gneiting 2005; Hemri 2014; Dai et al. 2018; Palmer 2002). Nowadays, the most advanced probabilistic forecasts rely on ensemble NWP systems (Hemri et al. 2022), which form the core of modern probabilistic precipitation and hydrometeorological forecasting (Andrade et al. 2024; Liu et al. 2022; Yadav et al. 2022; Li et al. 2017).\u003c/p\u003e \u003cp\u003eWith advances in ensemble forecasting theory and technology, substantial progress has been made in operational ensemble prediction worldwide. Nevertheless, raw ensemble outputs often exhibit systematic biases and insufficient spread, making statistical post-processing necessary to improve forecast accuracy and reliability (Chen et al. 2024; Hamill et al.1998). As a result, post-processing techniques have become an integral component of operational workflows in many national meteorological services, providing more accurate, automated, and seamless forecast products for users and the public (Vannitsem et al. 2021; Whan et al.2018). Numerous post-processing methods have been proposed in recent years and can be broadly classified as parametric or non-parametric (Vannitsem et al. 2021). Key parametric approaches include Bayesian Model Averaging (Sloughter et al. 2010; Baran et al. 2014) and Ensemble Model Output Statistics (EMOS) (Gneiting et al. 2005; Baran and Lerch 2016; Scheuerer and Hamill 2015), while common non-parametric methods include Quantile Mapping (Madadgar et al. 2014) and Quantile Regression Forests (Meinshausen et al. 2006). EMOS and BMA generally show comparable skill for probabilistic precipitation forecasts, with each having distinct advantages (Javanshiri et al. 2021; Han et al. 2018).\u003c/p\u003e \u003cp\u003eSeveral studies have explored EMOS formulations tailored to precipitation. Baran et al. (2016) applied the Censored-Shifted Gamma (CSG) distribution within the EMOS framework and found it outperformed Generalised Extreme Value (GEV)-based EMOS and Gamma-based BMA. More recently, Angus et al. (2024) evaluated EMOS_CSG using ECMWF dual-resolution 24-hour precipitation forecasts across Europe and showed that EMOS_CSG substantially enhanced forecast skill, approaching the performance of more advanced quantile-mapping techniques while requiring no additional historical data. However, EMOS implicitly maps grid-cell average precipitation to point locations (Hemri S 2022), which may limit performance under highly sub-grid variability conditions (Hewson and Pillosu 2021).\u003c/p\u003e \u003cp\u003eTo address the limitations of previous post-processing techniques, the ECMWF developed ecPoint, an operational post-processing system specifically designed for point precipitation forecasting (Hewson and Pillosu 2021). ecPoint is a non-parametric technique that accounts for sub-grid weather variability, grid-scale biases, and regime-dependent relationships. Notably, it requires only one year of global historical data to generate reliable point-wise probabilistic forecasts (Hewson and Pillosu 2021). Hemri and Hewson (2022) compared ecPoint with EMOS for 12-hour precipitation over Switzerland and surrounding areas and found that ecPoint outperformed EMOS at longer lead times and for heavy precipitation events. Gasc\u0026oacute;n et al. (2024) demonstrated that ecPoint improves ECMWF maximum-precipitation forecasts and better identifies severely affected areas across Europe. Pillosu et al. (2024) validated ecPoint against one year of observations in Ecuador, showing strong capability for identifying flash-flood-prone areas associated with small-scale convective systems.\u003c/p\u003e \u003cp\u003eWhile Hemri et al. (2022) demon strated the advantages of ecPoint over Global EMOS, it remains an open question whether these advantages hold against a locally optimized EMOS (Local EMOS) in regions with topographical variance. Local EMOS is capable of correcting stationary local biases; however, we hypothesize that it lacks the physical mechanism to represent the transient, sub-grid convective extremes characteristic of the East Asian Monsoon. Therefore, this study compares ecPoint against a strict Local EMOS benchmark to determine if statistical localization is sufficient, or if the physics-based sub-grid mapping of ecPoint is required for Eastern China. Through comparative evaluation, this study provides a theoretical foundation for the further development of ensemble forecast post-processing techniques and their promotion in practical applications.\u003c/p\u003e \u003cp\u003eThe remainder of this paper is organized as follows: Section 2 describes the study area, data sources, and the principles of the ecPoint and EMOS methods, along with the verification methodology. Section 3 presents a comprehensive evaluation of the baseline GEPS forecasts and the two post-processing techniques, including probabilistic scores, deterministic measures, and representative case studies. Section 3 provides the conclusions and discussion.\u003c/p\u003e"},{"header":"2 Data and Method","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Research Area and datas\u003c/h2\u003e \u003cp\u003eThe study area encompasses eastern China, bounded by latitudes 25.25\u0026deg;\u0026ndash;36.25\u0026deg;N and longitudes 114.0\u0026deg;\u0026ndash;124.0\u0026deg;E (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e), and includes the Yangtze River Delta region. The region\u0026rsquo;s topography is highly varied, including both coastal and inland areas. It experiences pronounced seasonal variations and is characterized by a subtropical monsoon climate, with additional influences from temperate monsoons, and is situated at the intersection of northern and southern climatic zones and at the interface between terrestrial and marine environments (Ding 2013). Precipitation is concentrated in the summer months, while winters may bring substantial snowfall. The core factors influencing summer weather in eastern China\u0026mdash;the Western Pacific Subtropical High (WPSH), the Meiyu front, and frequent typhoon landfalls\u0026mdash;frequently produce severe weather, causing plum-rain floods and typhoon-induced inundations. Prominent administrative provinces within the study area include Jiangsu, Zhejiang, Shanghai, and Anhui, collectively known as the \u0026ldquo;Region of Rivers and Lakes.\u0026rdquo; Beyond its complex climate and geography, this area is one of the most densely populated in China, amplifying the socio-economic consequences of extreme weather events.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe China Meteorological Administration's Global Ensemble Forecast System (CMA-GEPS) is built on the domestically developed GRAPES (Global/Regional Assimilation and Prediction System) model (Huo et al.2020; Gao et al. 2022; Chen et al. 2003; Shen et al. 2020). Initial perturbations are generated using singular vector techniques (Li et al. 2019), and uncertainty in the model's physical tendencies is represented through the Stochastically Perturbed Parameterization Tendencies (SPPT) and Stochastic Kinetic Energy Backscatter (SKEB) schemes (Yuan et al. 2016; Peng et al. 2019).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Description of the dataset\u003c/h2\u003e \u003cp\u003eThis study employs the GEPS system, which reports data starting at 12:00 UTC. This system, with 30 ensemble members and features a horizontal resolution of 50\u0026times;50 km. The training dataset for model calibration spans May 19, 2020, to May 18, 2023 (a total of 3 years). The test dataset for precipitation forecasts covers June 1 to August 31, 2023, with forecast lead times ranging from 12 to 84 hours, utilizing cumulative precipitation forecast products at 12-hour intervals. Hourly precipitation data from 15,670 meteorological stations within the study area were selected as observational data. For the ecPoint method, the following control forecasts were chosen as forecast factors: total precipitation (tp), convective precipitation (cp), the U-component of the 700 hPa wind field (u700), 700 hPa wind field V-component (v700), and convective available potential energy (cape) as forecast factors. The training process employed a 12-hour forecast lead time. The EMOS_CSG technique utilizes ensemble forecasts of TP as forecast factors, with training data covering seven forecast lead times ranging from 12 to 84 hours.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Post-processing method\u003c/h2\u003e \u003cp\u003eIn this study, we explore the ecPoint and EMOS methods for precipitation post-\u003c/p\u003e \u003cp\u003eprocessing. A brief overview of these two techniques is presented below, with detailed formulations provided in the supplementary material.\u003c/p\u003e \u003cp\u003eThe ecPoint method is a station-based statistical post-processing system developed at ECMWF (Pillosu and Hewson 2017; Hewson and Pillosu 2021). It is also a point-scale scheme based on decision-tree modelling that explicitly accounts for sub grid-scale variability in NWP precipitation fields. Rather than treating grid-box mean precipitation as representative of point measurements, ecPoint conditions on a set of flow-dependent predictors\u0026mdash;such as model resolution, orography, convective environment and precipitation regime\u0026mdash;to derive statistical transfer functions between grid-scale precipitation and its conditional distribution at observation sites within each grid cell. These transfer functions are obtained by estimating conditional probability distributions from large training archives of co-located model forecasts and rain-gauge observations, and are then applied in real time to transform raw model output into calibrated, point-scale probabilistic precipitation forecasts (Hewson and Pillosu 2021).\u003c/p\u003e \u003cp\u003eThe Ensemble Model Output Statistics (EMOS) method is an extension of the traditional Model Output Statistics (MOS) approach for quantitative forecasting. Building on MOS, EMOS provides a parameterized post-processing framework that links ensemble forecasts to a chosen predictive probability distribution. The method first specifies an appropriate probability distribution for the forecast variable and then uses a link function to relate the distribution parameters to the ensemble forecasts. The parameters are estimated by an optimization procedure and used for correcting forecast biases (Gneiting et al., 2005; Baran et al., 2016). For continuous variables such as temperature or wind speed, EMOS typically assumes a normal or truncated normal distribution. However, for precipitation\u0026mdash;which is non-negative, highly skewed, and has a point mass at zero\u0026mdash;the normal distribution is unsuitable. Instead, EMOS adopts a left-Censored, Shifted Gamma (CSG) distribution to model precipitation. The CSG distribution accommodates continuous values that can be positive or zero, with left-censoring at zero (Scheuerer and Hamill 2015). This formulation underpins the CSG-based EMOS model proposed by Baran and Nemoda (2016) (Scheuerer 2015).\u003c/p\u003e \u003c/div\u003e"},{"header":"3 Results","content":"\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Verification\u003c/h2\u003e \u003cp\u003eThis study assesses the probabilistic performance of the raw ensemble and its post-processed counterparts using ecPoint and EMOS through a systematic comparative analysis. For continuous variables, overall probabilistic skill is quantified using the Continuous Ranked Probability Score (CRPS) and the Continuous Ranked Probability Skill Score (CRPSS). For event-based forecasts, the Brier Score (BS) and receiver operating characteristic (ROC) curves, summarized by the area under the curve (AUC), are employed to evaluate reliability and discrimination. Additionally, reliability diagrams and sharpness histograms are used to diagnose forecast calibration and spread.\u003c/p\u003e \u003cp\u003eThe training dataset spans three years. For scoring, we used forecasts from the validation dataset initialized at 12:00 UTC during summer 2023 with lead times of 12\u0026ndash;84 hours; post-processed outputs from all stations within the study domain were included.\u003c/p\u003e \u003cp\u003eBoxplots of CRPS show ecPoint and EMOS (Figure \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e) significantly outperform the raw GEPS at all lead times. ecPoint has the lowest mean CRPS, while EMOS yields the most stable (low median) errors; the raw GEPS shows the highest means and extremes, indicating larger biases and ensemble spread. All methods exhibit a 12‑hour periodicity, and CRPS variability increases with lead time.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eTo explore the spatial heterogeneity of correction performance between the two post-processing techniques across stations, the Continuous Ranked Probability Skill Score (CRPSS) was computed for each station using the raw GEPS as the reference. On average, EMOS and ecPoint improve forecast skill by approximately 10% and 26%, respectively (Figs.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e2\u003c/span\u003ea\u0026ndash;b). EMOS outperforms the raw model at nearly all stations except parts of southern China and localized areas in Shandong. ecPoint exhibits positive CRPSS values at nearly all stations, demonstrating robust and widespread improvement.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe relative CRPSS using EMOS as the baseline (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e2\u003c/span\u003ec) highlights clear spatial contrasts: EMOS performs better in central Anhui, southern Zhejiang, and mountainous/coastal regions of Fujian and Jiangxi, while ecPoint is more effective in transition-terrain zones, valleys (e.g., the Fuzhou Basin), plains, and high-elevation areas of southern southern Anhui and Shandong. Lead-time-dependent CRPSS analyses (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e2\u003c/span\u003ed) show positive medians for both methods, with ecPoint exhibiting increasing skill at longer lead times.\u003c/p\u003e \u003cp\u003eFor precipitation thresholds of \u0026gt;\u0026thinsp;0.0 mm, 0.1 mm, and 5.0 mm (Figs.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e3\u003c/span\u003ea\u0026ndash;c), both post-processing techniques yield substantially lower Brier Scores (BS) relative to GEPS, and ecPoint shows the greatest improvement. At higher thresholds (\u0026ge;\u0026thinsp;10.0\u0026ndash;30.0 mm), the performance of ecPoint and EMOS gradually converges. EMOS slightly outperforms ecPoint at isolated lead times. Across all thresholds, the most substantial BS reduction\u0026mdash;indicating maximum improvement\u0026mdash;occurs at the 84-hour lead time, implying that post-processing effectively extends the useful forecast horizon. Aggregated across all lead times (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e3\u003c/span\u003ed), BS decreases monotonically with increasing threshold for all methods, and both ecPoint and EMOS consistently outperform GEPS. However, at extreme thresholds (\u0026gt;\u0026thinsp;25.0\u0026ndash;100.0 mm), improvement diminishes and may approach zero due to limited sample size.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe reliability diagrams (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e4\u003c/span\u003e) show that ecPoint and EMOS greatly improve reliability across thresholds. ecPoint generally lies closest to the diagonal, while GEPS deviates most strongly, especially at high thresholds. EMOS typically performs between the other two. At thresholds\u0026thinsp;\u0026gt;\u0026thinsp;0.1 mm low-information intervals (e.g., 0\u0026ndash;0.2 mm), both corrected forecasts tend to be under-confident because of influence from the raw ensemble. At medium-to-high confidence intervals, EMOS often becomes over-confident. At thresholds\u0026thinsp;\u0026ge;\u0026thinsp;10 mm, both methods exhibit reduced reliability, with EMOS deviating more than ecPoint. GEPS becomes increasingly conservative at thresholds\u0026thinsp;\u0026gt;\u0026thinsp;15.0 m.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eSharpness plots (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e4\u003c/span\u003e, upper left) show a U-shaped distribution for GEPS, while EMOS demonstrates higher sharpness than ecPoint at lower thresholds but poorer reliability, explaining its higher BS. As thresholds increase, sharpness profiles of all models converge; at thresholds\u0026thinsp;\u0026gt;\u0026thinsp;15.0\u0026ndash;70.0 mm, all three cluster near zero. Together with the reliability results, this indicates strong performance of both post-processing methods for high-threshold precipitation (details in Figure S3).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAUC-threshold plots (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e5\u003c/span\u003e) confirm that all methods perform reasonably at low thresholds (\u0026le;\u0026thinsp;5.0 mm/12h). However, GEPS deteriorates rapidly at thresholds above 25.0 mm. EMOS shows noticeable degradation above 50.0 mm, whereas ecPoint maintains the highest AUC across most thresholds, despite reduced curve smoothness beyond 70.0 mm. For distinguishing heavy and extreme precipitation (e.g., torrential rain), ecPoint or EMOS should be prioritized.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eROC curves and AUC values (Figure S2) show that both ecPoint and EMOS enhance event discrimination across thresholds. At thresholds above 0.1 mm, both methods outperform GEPS. As threshold increases, GEPS progressively approaches the no-skill line, while EMOS and ecPoint maintain clear positive curvature. EMOS\u0026rsquo;s ROC smoothness declines beyond 50.0 mm, and ecPoint\u0026rsquo;s smoothness weakens past 70.0 mm, due to sparse samples. AUC values remain relatively stable with lead time, showing limited sensitivity from 24 to 84 hours. Combining AUC and reliability, both post-processing techniques significantly improve prediction of heavy and extreme precipitation.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Case studies\u003c/h2\u003e \u003cp\u003e \u003cb\u003eTyphoon Doksuri (26\u0026ndash;28 July 2023)\u003c/b\u003e \u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e6\u003c/span\u003e compares 12-hourly 85th-percentile forecasts against observations. GEPS reasonably captures precipitation occurrence during 0\u0026ndash;24 hours, but misplaces heavy rainfall cores too far east. Both post-processing techniques markedly improve location and intensity forecasts. ecPoint provides the most accurate representation of extreme precipitation (\u0026gt;\u0026thinsp;200.0 mm/12 h) in Fujian during 36\u0026ndash;48 hours, whereas GEPS and EMOS underestimate peak magnitudes at several stations.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eRainstorm event associated with a quasi-stationary Meiyu front(19 Jun 2023)\u003c/b\u003e \u003c/p\u003e \u003cp\u003eFrom 00:00 to 12:00 on 19 Jun 2023, under a quasi-stationary Meiyu front, moderate to heavy rainfall occurred along the river and in the southern areas of Jiangsu Province and in northern Jiangxi Province, with localized amounts reaching rainstorm intensity or above. The 36\u0026ndash;48 h precipitation forecasts initialized at 12:00 on 17 June 2023 (Fig.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e7\u003c/span\u003e) show that, at the 50th percentile, the raw model predicts an overly broad precipitation area, whereas the ecPoint and EMOS post-processed forecasts produce spatial patterns more consistent with observations. At the 90th percentile, ecPoint and EMOS better reproduce the location and intensity of rainstorms along the Yangtze River in Jiangsu and in the border region between northern Jiangxi and southern Anhui, while GEPS gives more diffuse and indistinct rainstorm forecasts there. Both post-processing methods, however, show a marked positive bias in precipitation amounts over northern Jiangsu and Shandong, which is more pronounced for ecPoint.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAcross diverse precipitation types and events, both post-processing techniques significantly enhance forecast accuracy, spatial continuity, and ensemble dispersion. ecPoint performs particularly well in capturing high-precipitation areas and convective events at high quantiles (e.g., \u0026ge;\u0026thinsp;95th) but may overpredict heavy rainfall in some regions (e.g., Figure S4). EMOS compensates for light-precipitation under-forecasting but is prone to false positives over broader areas. Overall, both methods yield substantial improvement over the raw GEPS ensemble, with complementary strengths depending on threshold, elevation, and weather regime.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"4 Conclusions and Discussion","content":"\u003cp\u003eThis study evaluates two post-processing techniques\u0026mdash;ecPoint and EMOS\u0026mdash;applied to the CMA-GEPS precipitation ensemble forecast system for Eastern China. Overall, both methods substantially improved the reliability and accuracy of the forecasts, while reducing the ensemble forecast\u0026rsquo;s systematic bias and enhancing the dispersion of the ensemble forecasts. ecPoint demonstrates the highest overall skill, particularly in terms of CRPS, and produces the most substantial improvement for light-precipitation forecasts. For heavy precipitation and above, the accuracy of ecPoint and EMOS becomes comparable, as reflected by their Brier Scores. Both methods substantially improve the calibration and resolution of forecast probabilities. ecPoint exhibits consistently stronger reliability across thresholds, while EMOS performs more stable at moderate thresholds but less reliable under heavy-rainfall conditions. Both post-processing methods improve performance for convective cases. ecPoint, in particular, exhibits strong early-warning skill for extreme precipitation, with enhanced discrimination at high quantiles.\u003c/p\u003e \u003cp\u003eSpatial analysis indicates that the EMOS model performs better in certain high-elevation areas\u0026mdash;particularly in the highly influenced terrain zones of southeastern and central East China, as well as most coastal regions. In contrast, the ecPoint model excels in river valleys, sloping terrain, and most plains (excluding Shandong Mountains and southern Anhui). Considering the limited topographical variability within the study area, the validation results pertaining to different geographical altitudes were not statistically significant and are therefore not discussed further in this paper. Moreover, the variations in scoring metrics across different forecast lead times were also marginal. Future research is warranted to explore these aspects in diverse geographical regions and with extended forecast lead times.\u003c/p\u003e \u003cp\u003eThe EMOS formulation used here employs a single forecast factor (total precipitation, tp) over the 12- to 84-hour lead times. This minimal factor set\u0026mdash;combined with EMOS\u0026rsquo;s monotonic relationship between predictors and predictands\u0026mdash;likely limits its performance and leaves room for further optimization. Conversely, ecPoint requires decision-tree training and thus uses only one time step (12-hour forecasts) but incorporates five forecast factors. The reduction in temporal depth may influence temporal sensitivity but enhances its capability to represent spatial variability and extreme-event structure. To ensure comparability with EMOS, ecPoint forecasts are bilinearly interpolated from grid to station. This interpolation increases spatial dispersion and can lead to more apparent false positives\u0026mdash;especially at high thresholds with inherently low event frequency. As a result, ecPoint\u0026rsquo;s probability-based accuracy metrics (e.g., BS) may be degraded despite its strong reliability and resolution.\u003c/p\u003e \u003cp\u003eIn recent years, Artificial Intelligence techniques, have become a key focus in model post-processing technologies, achieving significant advancements in meteorology. These methods automatically learn complex data patterns, offering innovative solutions for ensemble forecast post-processing (Reichstein et al. 2019). To align with practical operational requirements, AI-based approaches are not within the scope of this study; however, future research will consider incorporating AI methods to further expand the research content based on the findings of this paper.\u003c/p\u003e"},{"header":"Declarations","content":" \u003cp\u003e \u003cstrong\u003eCompeting interests\u003c/strong\u003e \u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e \u003c/p\u003e\u003cp\u003e \u003ch2\u003eSupplements\u003c/h2\u003e \u003cp\u003eSupplement: The supplementary materials for this paper consist of three sections: the first part provides a detailed explanation of the post-processing methods and validation procedures; the second part contains supplementary result discussions; and the third part includes supplementary figures and explanations.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eCRediT authorship contribution-statementeSonum Steiik Data-curation, Formal analysis, Investigation, Software, Validation, Visualization,Writing-original-draft, Writing-eview \u0026amp; editing. LIU Pu: Data-curation, Investigation, Software, Writing-review-\u0026amp; editing.Wang-Jialing: Supervision, Writing-review-\u0026amp; editing. WANG Yong: Conceptualization, Data-curation, Funding acquisition, Supervision, Writing-review-\u0026amp; editing, Project administration.\u003c/p\u003e\u003ch2\u003eAcknowledgement\u003c/h2\u003e\u003cp\u003eThe first author special thanks to Pillosu Fatima and Hewson Timothy of the ECWMF for sharing the ecPoint software and offering assistance. We are grateful to Ms Chen Jieyu for her guidance and support in revising this paper. We also extend our appreciation to Zhu Yanwei, Huo Ziqiang, Cao Bufan, Song Qianqian and Wang Mingming for their participation in discussions and valuable suggestions throughout the research process. In addition, the authors are grateful to the reviewers for their patient responses and thoughtful suggestions throughout the review process.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eAndrade FS, Arsenault R, Poulin A, Troin M, Armstrong W. (2024) Application of weather post-processing methods for operational ensemble hydrological forecasting on multiple catchments in Canada. Journal of Hydrology 642: 131861.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAngus M, Widmann M, Orr A, Ashrit R, Leckebusch GC, Mitra A. (2024) A comparison of two statistical postprocessing methods for heavy-precipitation forecasts over India during the summer monsoon. Quarterly Journal of the Royal Meteorological Society 150(761): 1865-83.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAstakhova E, Alferov D, Alferov Y, Bundel A. (2021) Ensemble approach to weather forecasting. In Journal of Physics: Conference Series p012070.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBottazzi M, Rodr\u0026iacute;guez-Mu\u0026ntilde;oz L, Chiavarini B, Caroli C, Trotta G, Dellacasa C, Marras GF, Montanari M, Santini M, Mancini M, D'Anca A. (2024) High performance computing to support land, climate, and user‐oriented services: The HIGHLANDER Data Portal. Meteorological Applications 31(2): e2166.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBaran S, Lerch S. (2016) Mixture EMOS model for calibrating ensemble forecasts of wind speed. Environmetrics 27(2): 116\u0026thinsp;\u0026minus;\u0026thinsp;30.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBaran S, Nemoda D. (2016) Censored and shifted gamma distribution based EMOS model for probabilistic quantitative precipitation forecasting. Environmetrics 27(5): 280\u0026ndash;292.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBaran S, Horanyi A, Nemoda D. (2014) Comparison of the BMA and EMOS statistical methods in calibrating temperature and wind speed forecast ensembles. Idoj\u0026aacute;r\u0026aacute;s 118(3): 217\u0026ndash;241.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen J, Janke T, Steinke F, Lerch S. (2024) Generative machine learning methods for multivariate ensemble postprocessing. The Annals of Applied Statistics 18(1): 159\u0026ndash;183.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDai K, Zhu YJ, Bi BG. (2018) The review of statistical post-process technologies for quantitative precipitation forecast of ensemble prediction system. Acta Meteorologica Sinica 76(4): 493\u0026ndash;510(in Chinese).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGneiting T, Raftery A E. (2005) Weather forecasting with ensemble methods. Science 310(5746): 248\u0026ndash;249.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGasc\u0026oacute;n E, Montani A, Hewson T D. (2024) Post-processing output from ensembles with and without parametrised convection, to create accurate, blended, high‐fidelity rainfall forecasts. Quarterly Journal of the Royal Meteorological Society 150(762): 3117\u0026ndash;3145.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGr\u0026ouml;nquist P, Yao C, Ben-Nun T, et al. (2021) Deep learning for post-processing ensemble weather forecasts. Philosophical Transactions of the Royal Society A 379(2194): 20200092.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHewson T D, Pillosu F M. (2021) A low-cost post-processing technique improves weather forecasts around the world. Communications Earth \u0026amp; Environment 2(1): 132.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHan K, Choi J T, Kim C. (2018) Comparison of statistical post-processing methods for probabilistic wind speed forecasting. Asia-Pacific Journal of Atmospheric Sciences 54(1): 91\u0026ndash;101.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHamill T M, Colucci S J. (1998) Evaluation of Eta\u0026ndash;RSM ensemble probabilistic precipitation forecasts. Monthly Weather Review 126(3): 711\u0026ndash;724.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHemri S, Hewson T, Gasc\u0026oacute;n E, et al. (2022) How do ecPoint precipitation forecasts compare with postprocessed multi-model ensemble predictions over Switzerland? European Centre for Medium Range Weather Forecasts.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHemri S, Scheuerer M, Pappenberger F, et al. (2014) Trends in the predictive performance of raw ensemble weather forecasts. Geophysical Research Letters 41(24): 9197\u0026ndash;9205.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHersbach H, Bell B, Berrisford P, et al. (2020) The ERA5 global reanalysis. Quarterly journal of the royal meteorological society 146(730): 1999\u0026ndash;2049.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJavanshiri Z, Fathi M, Mohammadi S A. (2021) Comparison of the BMA and EMOS statistical methods for probabilistic quantitative precipitation forecasting. Meteorological Applications 28(1): e1974.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu X, Zhang L, She D, et al. (2022) Postprocessing of hydrometeorological ensemble forecasts based on multisource precipitation in Ganjiang River basin, China. Journal of Hydrology 605: 127323.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi W, Duan Q, Miao C, et al. (2017) A review on statistical postprocessing methods for hydrometeorological ensemble forecasting. Wiley Interdisciplinary Reviews: Water 4(6): e1246.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLorenz E N. (1963) The mechanics of vacillation. Journal of Atmospheric Sciences 20(5): 448\u0026ndash;465.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMeinshausen N, Ridgeway G. (2006) Quantile regression forests. Journal of machine learning research 7(6).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMadadgar S, Moradkhani H, Garen D. (2014) Towards improved post-processing of hydrologic forecast ensembles. Hydrological Processes28(1): 104\u0026ndash;122.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePillosu F M, Bucherie A, Kruczkiewicz A, et al. (2024) Can global rainfall forecasts identify areas at flash flood risk? Proof of concept for Ecuador. ECMWF Technical Memo (917).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePillosu F, Hewson T. (2017) New point-rainfall forecasts for flash flood prediction. ECMWF Newsletter 153: 2\u0026ndash;3.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePalmer, T. N. (2002) The economic value of ensemble forecasts as a tool for risk assessment: From days to decades. Quarterly Journal of the Royal Meteorological Society: A journal of the atmospheric sciences, applied meteorology and physical oceanography 128(581): 747\u0026ndash;774.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eReichstein M, Camps-Valls G, Stevens B, et al. (2019) Deep learning and process understanding for data-driven Earth system science. Nature 566(7743): 195\u0026ndash;204.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eScheuerer M, T. M. Hamill. (2015). Statistical Postprocessing of Ensemble Precipitation Forecasts by Fitting Censored, Shifted Gamma Distributions. Mon. Wea. Rev 143: 4578\u0026ndash;4596.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRaftery A E, Gneiting T, Balabdaoui F, et al. (2005) Using Bayesian model averaging to calibrate forecast ensembles. Monthly weather review 133(5): 1155\u0026ndash;1174.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSloughter J M L, Gneiting T, Raftery A E. (2010) Probabilistic wind speed forecasting using ensembles and Bayesian model averaging[J]. Journal of the american statistical association 105(489): 25\u0026ndash;35.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVannitsem S, Bremnes J B, Demaeyer J, et al. (2021) Statistical postprocessing for weather forecasts: Review, challenges, and avenues in a big data world. Bulletin of the American Meteorological Society 102(3): E681-E699.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWilks D S. (2011) Statistical methods in the atmospheric sciences. Academic press, USA\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWhan K, Schmeits M. (2018) Comparing area probability forecasts of (extreme) local precipitation using parametric and machine learning statistical postprocessing methods. Monthly Weather Review 146(11): 3651\u0026ndash;3673.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYadav R, Yadav S M. (2022) Review on Statistical Post-processing of Ensemble Forecasts. In International Conference Innovation in Smart and Sustainable Infrastructure. pp469-476.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShen Xueshun, Wang Jianjie, Li Zechun, Chen Dehui, Gong Jiandong. (2020) China's independent and innovative development of numerical weather prediction. Acta Meteorologica Sinica 78(3): 451\u0026ndash;476(in Chinese).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen Jing, Zhu Yuejian, Duan Wansuo, Zhi Xiefei, Min Jinzhong, Li Xiaoli, Deng Guo, Yuan Huiling, Feng Jie, Du Jun, Li Qiaoping, Gong Jiandong, Shen Xueshun, Mu Mu. (2025) A review on development, challenges, and future perspectives of ensemble forecast. Acta Meteorologica Sinica 83(3): 480\u0026ndash;502(in Chinese).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen Jing, Xue Jishan, Yan Hong. (2003) Uncertainty and Ensemble Forecasting Experiment of Mesoscale Rainstorm Numerical Forecast in South China. Acta Meteorologica Sinica (04): 432\u0026ndash;446(in Chinese).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXiaoli LI, Yongzhu LIU. (2019) The improvement of GRAPES global extratropical singular vectors and experimental study[J]. Acta Meteorologica Sinica 77(3): 552\u0026ndash;562(in Chinese).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYuan Y, Li X L, Chen J, et al. (2016) Stochastic parameterization toward model uncertainty for the GRAPES mesoscale ensemble prediction system. Meteor Mon 42(10): 161\u0026ndash;1175(in Chinese).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePeng F, Li X L, Chen J et al. (2019) A stochastic kinetic energy backscatter scheme for model perturbations in the GRAPES global ensemble prediction system. Acta Meteor Sinica 77(2): 180\u0026ndash;195(in Chinese).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGAO Li, ZHENG Jiawen, ZHAO Zuosen, LUO Yuelin, REN Pengfei, YAO Guohua. (2022) Research, Development, and Application of the Unified Post-Processing System for the CMA-GEPS/REPS Ensemble Prediction. Advances in Earth Science[J] 37(12): 1211\u0026ndash;1222(in Chinese).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHUO Zhenhua, LIU Yongzhu, CHEN Jing, LI Xiaoli, QU Anxiang. (2020) The preliminary appliation of tropical cyclone targeted singular vectors in the GRAPES global ensemble forecasts[J]. Acta Meteorologica Sinica 78(1): 48\u0026ndash;59 (in Chinese).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDing Y H. (2013) Climate of China[M]. Beijing: China Meteorological Press(in Chinese).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"NWP, Post-processing, Ensemble forecast, Probability prediction, ecPoint, EMOS","lastPublishedDoi":"10.21203/rs.3.rs-8883798/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8883798/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003ePost-processing of numerical weather prediction (NWP) models constitutes a pivotal link in enhancing forecast performance. Despite their recognition as cutting-edge point-based post-processing techniques, systematic comparative evaluations of ecPoint (ECWMF for point forecasts) and Ensemble Model Output Statistics (EMOS)\u0026mdash;particularly assessments of their applicability outside Europe and to Chinese ensemble forecasting systems\u0026mdash;remain insufficient. In this study, we evaluate two advanced post-processing techniques\u0026mdash;Ensemble Model Output Statistics (EMOS) and the point-based European Centre for ECMWF statistical ensemble method (ecPoint)\u0026mdash;for calibrating ensemble precipitation forecasts. A comprehensive assessment of the performance of these ensemble post-processing methods is conducted using the China Meteorological Administration\u0026rsquo;s Global Ensemble Forecasting System (CMA-GEPS) forecast over eastern China. The results demonstrate that both methods significantly reduce systematic biases and improve the reliability and dispersion of ensemble forecasts. Notably, improvement in forecast accuracy is observed even under convective weather conditions, and early warnings of extreme precipitation events. Overall, while both methods show comparable performance, they exhibit distinct behaviors across different regions. The ecPoint method slightly outperforms EMOS in terms of Continuous Ranked Probability Score (CRPS) and provides improved resolution and early warning capabilities at various precipitation thresholds.\u003c/p\u003e","manuscriptTitle":"Comparative Evaluation of ecPoint and Local EMOS for CMA-GEPS Precipitation Forecast over Eastern China","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-05 11:02:19","doi":"10.21203/rs.3.rs-8883798/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"3e317d36-dc53-470d-a0ca-600ce9f4312d","owner":[],"postedDate":"March 5th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-03-19T03:24:35+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-05 11:02:19","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8883798","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8883798","identity":"rs-8883798","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}