In this paper, anomaly correlation coefficient (ACC), time correlation coefficient (TCC), symbol consistency rate (SCR) and trend anomaly comprehensive score (PS) are used to evaluate the prediction skills of ten models in the S2S plan for pentad precipitation anomaly percentage (PAP) in Sichuan province during rainy season. Based on model bias analysis, a "correction of positive and negative probability anomaly" scheme was proposed to improve the model forecast. The results show the prediction skills of most models decrease rapidly, and the gap of prediction skills among these models narrows with increasing forecast lead-time. At around the 10th day, all the models enter the low skill period with little changes with lead-time. The models only have a certain ability to predict the abnormal trend of precipitation, and the BoM model has the best skill among these models. All the models except the BoM underestimate the anomaly degree of precipitation trend. The anomaly deviations of predicted PAP by these models are from −33% to −18% and invariant with the lead-time, but they are unevenly distributed in space. After error correction, ACC, SCR and PS of each model are improved, especially PS. The correction effect on the sub-seasonal scale is better than that on the synoptic scale. After modification, the average PS scores of all models on the sub-seasonal scale are higher than 76.8, of which 66.7% are within the range 79.2—80.2, nearly 8 points higher than the business-scoring standard (72.0). This correction effect has also been verified in 4-year independent sample tests.