Flood occurrence prediction using Monte Carlo methods and machine learning for mitigating Climate Impact in Northwestern Argentina

Flash floods in Catamarca, Córdoba, and La Madrid (northwestern Argentina) threaten lives and infrastructure due to complex orography and highly variable precipitation. Linear time-series models often underperform under heavy-tailed, non-stationary rainfall. We evaluate a hybrid framework that (i) quantifies distributional shifts in daily rainfall using divergence metrics—Kullback–Leibler, Bhattacharyya, and Wasserstein/Earth Mover’s distances—computed for each hydrological year and season, (ii) learns early-warning signals with machine-learning regressors (AutoARIMA, multilayer perceptron, random forest, XGBoost), and (iii) characterizes forecast uncertainty via Monte Carlo ensembles. Using 1981–2024 data from NASA-POWER and three in-situ stations, XGBoost attains the lowest error (MSE 0.021, MASE 0.612, nRMSE 7.2%), and, under residual-bootstrap simulation, produces narrow and stable 5–20 year forecast bands. The framework couples divergence analysis with ensemble learning to improve flood-relevant rainfall prediction and provides actionable uncertainty quantification for risk management in orographically complex regions.