Abstract: To investigate the effectiveness of neighborhood ensemble probability forecasting in reducing uncertainties of heavy precipitation ensemble forecasts, a comparative study was conducted for the July 2023 severe torrential rain in North China. Three neighborhood ensemble probability forecasting methods— Neighborhood Ensemble Probability (NEP), Ensemble Mean Neighborhood Probability (EMNP) and Optimized Neighborhood Ensemble Probability (ONEP)—were compared with the Traditional Ensemble Probability (TEP) approach using forecasts from the China Meteorological Administration Regional Ensemble Prediction System (CMA-REPS). These methods were systematically assessed in terms of spatial distribution characteristics, probabilistic forecast scores, forecast skill at different probability thresholds, and sensitivity of above metrics to neighborhood scales. The results show that: (1) At a neighborhood scale of 60 km, the spatial distribution and pattern of heavy rainfall probability in ONEP are closest to observations, while the forecast areas of EMNP and NEP are relatively smaller. As the neighborhood scale increases, the false alarms of ONEP sharply increases, but the probability values in the core rainfall area remain relatively stable. In contrast, the misses of EMNP and NEP gradually decrease, but their central probability values significantly decrease, especially for EMNP. (2) The performance of ONEP is sensitive to the neighborhood scale, achieving best Area under the Relative Operating Characteristic curve (AROC) and Brier Score at scales between 60 and 140 km. The AROC and Brier Score of NEP show less fluctuation with varying neighborhood scale, with its advantages primarily manifested at neighborhood scales larger than 140 km. Overall, EMNP exhibits lower forecast skill compared to ONEP and NEP. (3) The best Equitable Threat Score (ETS) of EMNP, NEP, and ONEP are better than that of TEP. The probability thresholds associated with the best ETS of EMNP and NEP are relatively low, ranging from only 5% to 25%, whereas the threshold of ONEP can reach 60% to 80%. (4) The study demonstrates that neighborhood probability methods effectively reduce the sensitivity of traditional probability forecasting to the model position errors through the statistical processing of the ensemble information within a spatial neighborhood, and provide a valuable technical approach for the refined forecasting of such extreme precipitation events. Nevertheless, additional case studies are required to assess the robustness and general applicability of these conclusions.