Assessment of GFS weather forecast model performance in reproducing the main atmospheric circulation patterns linked to precipitation in western tropical South America

Rainfall events in South America have increased in frequency and intensity over recent decades, causing significant socio-economic impacts. Understanding the large-scale atmospheric circulation patterns (CPs) associated with these events is crucial for improving weather forecasting and risk assessment. Therefore, this study aims to evaluate the skill of the Global Forecast System (GFS) forecasts in reproducing the main CPs and their associated rainfall over western tropical South America.
Daily winds at 200 and 850 hPa from the ERA5 reanalysis and GFS (D0 to D5, where D0 is the initial state and D1–D5 are the 1–5 day forecast) are used. In addition, gridded precipitation data from CHIRPS and GFS are analyzed. All datasets have a spatial resolution of 0.25° and cover the period 2015–2024. A combined principal component analysis (PCA) and k-means clustering approach is applied to identify nine circulation patterns for ERA5 and GFS. Composite analysis is used to relate each CP to its characteristic spatial precipitation patterns. The analysis is structured in two stages: (i) a comparison between ERA5 and the first step of GFS (GFS-D0) and (ii) an evaluation of forecast consistency from GFS-D0 to the subsequent five forecast days (GFS-D1 to GFS-D5). For both stages, the Heidke Skill Score (HSS) is calculated based on the daily occurrence frequency of the CPs.
ERA5 and GFS (D0 to D5) consistently identify the nine CPs, which are classified into three wet, two transitional, and four dry patterns, exhibiting a well-defined seasonal behavior over tropical South America (10°N–30°S, 90°W–30°W). ERA5 and GFS-D0 identify CPs with similar frequency and spatial behavior with a statistically significant association and high seasonal HSS values close to 0.9. When analyzed at the individual CP scale, all patterns exhibit high agreement, although transitional patterns show slightly lower skill. As forecast lead time increases, forecast consistency gradually degrades. HSS values decrease from approximately 0.9 on day 1 to about 0.5 on day 5 during austral winter, autumn, and spring, indicating a predictability limit beyond the third forecast day. Predictability is seasonal, with the highest persistence during austral summer and the lowest during winter. In this context, wet CPs exhibit the greatest stability, while dry patterns show the fastest degradation. Increasing lead time is also associated with growing spatial differences in wind and precipitation fields. Regarding precipitation, CHIRPS and GFS show a consistent spatial behavior, especially for the first forecast day, while these differences become more pronounced by the fifth forecast day. It is important to remark that CHIRPS and GFS present some discrepancies that could be associated with model biases.
These results demonstrate that GFS accurately reproduces dominant circulation patterns at short lead times. However, there is a clear degradation of predictability beyond three days, with important implications for rainfall forecasting and its spatial representation.