Modeling extreme precipitation in South America using climate indices: an interpretable linear framework

This study presents a methodological assessment of teleconnection-based modeling of monthly precipitation extremes in South America, quantified as the number of consecutive days above the month-specific 95th percentile and below the 5th percentile. The analysis covers 2002–2024 using the GPM IMERG Late Daily dataset. A multi-step framework is applied: hierarchical clustering and K-means identify regional patterns of extremes; Non-negative Matrix Factorization and Principal Component Analysis reduce dimensionality; and Variable Importance Analysis with 17 algorithms highlights the most relevant climate indices (CIs). Seasonally filtered time series are modeled using linear regression methods (ridge, lasso, and elastic net) to assess their ability to reproduce the general shape of the target series in a qualitative out-of-sample setting. Results show that wet extremes are more consistently reproduced than dry extremes and display stronger associations with Pacific-based ENSO indices, particularly in western South America, whereas dry extremes are reproduced less consistently, show greater dispersion, and appear less systematically linked to the large-scale indices considered here. By explicitly presenting coefficients and using publicly available indices, the study provides a transparent and interpretable framework for comparing teleconnection-based modeling strategies across regions. Main limitations include the weak reproducibility of dry events and the relatively short monthly record, which may affect the relative performance of the regression models and the VIA-based importance rankings. The contribution of the study therefore lies primarily in methodological comparison, stability assessment, and interpretability under limited sample conditions.