<![CDATA[Although national surveys report rising financial literacy in Indonesia, losses from investment fraud and illegal peer-to-peer lending continue to increase, suggesting that aggregate indices obscure differences across knowledge, behavior, and attitude. This study developed a four-class literacy classification framework using gradient boosting models and SHAP-based explanations, drawing on an online survey of 606 respondents in urban North Sumatra. The instrument, adapted from OJK/SNLIK, consisted of nine knowledge, fourteen behavior, and four attitude items. SMOTE-NC was applied only within cross-validation training folds, so the validation and test data contained no synthetic observations. CatBoost obtained the highest cross-validated Macro F1 (0.8950) and achieved a test accuracy of 87.91% and a Macro F1 of 0.8587, comparable to XGBoost and Logistic Regression. Because the classes were derived from the same items used as features, these models largely approximate the scoring rule. A model based only on demographic and contextual variables, which avoids this dependence, reached a Macro F1 of 0.5347 against a majority-class baseline of 0.1579. Attitude items showed the highest per-feature SHAP contribution (0.321), partly owing to their larger scoring weight, although this dimension was less reliable (α = 0.6167). Gold price awareness, decision confidence, and risk preference were the most discriminative items.]]>
Copyrights © 2026
