Accurate probabilistic quantitative precipitation forecasting plays a vital role in effective water resource management, agricultural planning, and flood risk mitigation, especially in regions prone to extreme weather events. Although Ensemble Prediction Systems (EPS) serve as the primary tools for generating these probabilistic forecasts, their raw outputs often contain systematic errors, including inherent bias and ensemble under-dispersion. Such inaccuracies can lead to less reliable forecasts that undermine decision-making for emergency responses and infrastructure planning.



To overcome these significant limitations, statistical post-processing techniques are widely applied to calibrate raw ensemble outputs and improve forecast reliability. In this study, 24-hour cumulative precipitation forecasts were analyzed for lead times of 24, 48, and 72 hours using two prominent post-processing approaches: Bayesian Model Averaging (BMA) and Ensemble Model Output Statistics with a Censored Shifted Gamma distribution (EMOS-CSG).



The dataset consists of precipitation forecasts generated by eight different configurations of the Weather Research and Forecasting (WRF) model over diverse synoptic stations across Iran, spanning the period from September 2015 to February 2016. A key focus of the research is to evaluate and compare two different parameter estimation strategies: a “global” model that uses data from all stations collectively and a clustering-based “semi-local” model that groups stations with similar climatological characteristics.



Results indicate that raw ensemble predictions are poorly calibrated and tend to over-forecast. Post-processing provides substantial added value, particularly at moderate to high precipitation thresholds (> 5 mm to > 25 mm), as evidenced by increasing RSS. Reliability analyses reveal that optimal configurations are highly model-dependent: BMA requires a Semi-local approach to prevent severe under-forecasting at higher accumulations, whereas EMOS-CSG achieves optimal calibration using a Global configuration. In terms of discrimination and overall accuracy (AUC and Brier Score), Global configurations demonstrate a consistent numerical advantage over Semi-local ones. Furthermore, Brier Skill Score (BSS) evaluations highlight that while BMA_Global excels at the lowest threshold (> 0.1 mm), CSG_Global dominates at higher thresholds, maintaining positive skill where Semi-local methods often fail. However, analysis of 95% confidence intervals reveals that the numerical differences between Global and Semi-local approaches, as well as between the BMA and EMOS-CSG models, generally lack statistical significance. Ultimately, CSG_Global and BMA_Semi-local emerge as the most robust frameworks for reliable precipitation forecasting.



It should be noted that the results are based on data from a single autumn–winter season (September 2015–February 2016). Therefore, the findings should not be assumed to generalize to year-round operational forecasting without further evaluation using multi-season datasets.