Abstract
This study uses machine learning (ML) to classify binary satisfaction with democracy among 17,902 observations from available Ecuadorian Latinobarómetro waves between 1996 and 2024. The analysis considers three conceptual temporal periods and evaluates Logistic Regression, Decision Tree, and Random Forest candidates under Class Weight Grid, RandomOverSampler, and SMOTENC strategies. Within each period, one nominal configuration was selected using training-only repeated nested grouped cross-validation, with survey year as the grouping variable. The frozen test partition was accessed only after the configuration, hyperparameters, and classification threshold were fixed. Selected-versus-runner-up advantages were small and uncertain across the nine repeated grouped outer-fold evaluations. On the untouched test sets, macro F1 was 0.5809, 0.6741, and 0.6922 for Periods 1–3, respectively. In terms of macro F1, the selected configurations exceeded both dummy benchmarks in every period, although fixed unweighted Logistic Regression achieved slightly higher macro F1 in Periods 2 and 3. Boundary and leave-one-wave-out analyses provided descriptive robustness evidence but did not establish unrestricted temporal generalization. Direct joint grouped permutation importance identified institutional trust as the leading group in Periods 1 and 2 and economic perceptions in Period 3; these were model- and specification-dependent predictive score drops, not causal effects. Overall, the analysis identifies predictive associations within the available analytical extract without establishing stable model-family dominance or causal relationships.