Application of Neural Network for Estimating Mean Monthly Rainfall in the State of Ceará, Brazilian Northeast

This research assessed the predictive power of Long Short-Term Memory (LSTM) networks to forecast mean monthly precipitation in Ceará using data from 2002–2025. We implemented a 12-month walk-forward validation focused on high-variability periods associated with La Niña (2020–2023). LSTM performance was compared with XGBoost, SARIMA, and a seasonal persistence baseline. SARIMA achieved superior average performance and greater operational stability (NSE 0.771; RMSE 38.54 mm/month), outperforming LSTM on average. However, Wilcoxon tests indicated no statistically significant differences between the models and the baseline. The main contribution is the validation of LSTM’s conditional utility: despite higher average volatility, LSTM attained peak skill in specific high-predictability windows (NSE up to 0.903), demonstrating capacity to capture non-linear dependencies. We also produced an average monthly precipitation forecast for 2026, noting limitations in predicting values near zero. We recommend ensemble approaches combining SARIMA and LSTM and the inclusion of exogenous variables to improve accuracy for extreme precipitation events.