This study investigates the predictive power of econometric and Machine Learning (ML) models for forecasting poverty in Ecuador, focusing on the role of remittances. Using annual data from 2007 to 2024, the performance of a traditional time series model (SARIMAX) is compared with ML architectures, including Random Forest (RF) and Long Short-Term Memory (LSTM). The results, validated using a robust rolling-window cross-validation methodology, demonstrate the marked superiority of the Hybrid (SARIMAX + RF) model, which achieves a significantly lower predictive error. Crucially, the analysis of the importance of variables in the nonlinear RF model reveals that structural factors, the Gini index, and the unemployment rate are the most influential predictors of poverty, surpassing the impact of remittances and gross domestic product (GDP) growth. The effectiveness of the hybrid model highlights complex and non-monotonic relationships that linear models fail to capture. Consequently, the findings suggest that poverty alleviation strategies in Ecuador should prioritize policies aimed at reducing inequality and promoting employment, highlighting the value of ML in formulating more robust and evidence-based policies.