Luiz Fernando Satolo, Larissa A. Silva, Luan C. S. M. Ozelim, Jaqueline A. J. P. Soares, Matheus Correia, Elton Escobar-Silva, Glauston Lima, Luz A. C. Pineda, Leonardo Santos

In May 2024, the region of the Rio Grande do Sul state experienced one of the worst floods in Brazilian history, affecting millions and causing severe damage to infrastructure. This study applies an agile hydrological forecasting approach using methods from traditional time-series models, such as ARIMA and SARIMA, and machine learning (ML) models, such as ElasticNet and LASSO. Data from streamflow monitoring stations were used to predict water levels at different lead times. The results showed that SARIMA consistently ranked among the top-performing models, while ElasticNet and LASSO demonstrated competitive performance among the ML methods. To enhance interpretability, Permutation Importance and Accumulated Local Effects were applied, highlighting the significance of autoregressive terms and upstream hydrological conditions. These findings underscore the potential of integrating traditional and ML methods in an agile approach to adaptive flood risk forecasting.