Abstract:This study utilizes the hindcast and real-time prediction data of CMME_S2S, the sub-seasonal to seasonal (S2S) product from the Chinese Multi-model Ensemble prediction system. Five evaluation metrics are adopted to assess the air temperature prediction skill over Southwest China. The prediction skill of CMME_S2S is further compared with that of S2S models from five domestic and international institutions . The results show that CMME_S2S exhibits significantly higher Prediction Correctness (PC) in daily air temperature prediction than the single-model ensemble means from corresponding institutions. However, it presents no distinct advantages in IPS, PCC, AS and other metrics for predicting pentad air temperature anomaly trends. Prediction skills of the models show obvious seasonal variations. For CMME_S2S, PC scores are the highest in summer and lowest in winter, and are higher than those of other models throughout all seasons. Its IPS, PCC and AS scores also peak in summer and bottom in winter, showing no significant differences from other models. Terrain complexity leads to pronounced spatial variations in the PC forecast skill of CMME_S2S. The skill value is the highest in the CT-M (hilly) subregion, followed by the CT-H (plateau) subregion, and the lowest in the CT-L (basin) subregion. By comparison, the IPS, PCC and AS skills of CMME_S2S differ slightly across various terrains, and other models demonstrate consistent spatial features. Regions with prominent errors of CMME_S2S in different seasons are mainly concentrated in the northern hilly areas with large elevation gradients, showing a consistent bias pattern of warm deviation in the west and cold deviation in the east. Within the same season, the largest discrepancies between predicted and observed air temperature occur over the northern plateau and basin areas. Meanwhile, CMME_S2S presents a relatively uniform spatial distribution of warm and cold biases across all seasons, with the smallest overall errors among the five comparative models, suggesting that CMME_S2S effectively suppresses random errors from individual models, which explains its superior prediction skill for air temperature anomalies. Prediction skills of the CMME_S2S for air temperature anomalies and temporal trends are generally higher during La Ni?a years but lower during years of ENSO phase transition, which provides a pathway for future model calibration and bias correction.