Relating anomaly correlation to lead time: Clustering analysis of CFSv2 forecasts of summer precipitation in China

Abstract Global climate model (GCM) forecasts are an integral part of long‐range hydroclimatic forecasting. We propose to use clustering to explore anomaly correlation, which indicates the performance of raw GCM forecasts, in the three‐dimensional space of latitude, longitude, and initialization time. Focusing on a certain period of the year, correlations for forecasts initialized at different preceding periods form a vector. The vectors of anomaly correlation across different GCM grid cells are clustered to reveal how GCM forecasts perform as time progresses. Through the case study of Climate Forecast System Version 2 (CFSv2) forecasts of summer precipitation in China, we observe that the correlation at a certain cell oscillates with lead time and can become negative. The use of clustering reveals two meaningful patterns that characterize the relationship between anomaly correlation and lead time. For some grid cells in Central and Southwest China, CFSv2 forecasts exhibit positive correlations with observations and they tend to improve as time progresses. This result suggests that CFSv2 forecasts tend to capture the summer precipitation induced by the East Asian monsoon and the South Asian monsoon. It also indicates that CFSv2 forecasts can potentially be applied to improving hydrological forecasts in these regions. For some other cells, the correlations are generally close to zero at different lead times. This outcome implies that CFSv2 forecasts still have plenty of room for further improvement. The robustness of the patterns has been tested using both hierarchical clustering and k‐means clustering and examined with the Silhouette score.