<![CDATA[The problem addressed in this study is the limited capacity of district-level food-security monitoring in Indonesia to anticipate deterioration in the following year, particularly when prediction relies on short longitudinal histories and must account for repeated observations, class imbalance, temporal change, and regional variation. The method involved constructing 2,056 temporally ordered prediction instances from Food Security Index data covering 514 Indonesian districts and cities during 2019–2024, representing six regional indicators through their current values and annual changes, and evaluating Logistic Regression, Random Forest, and eXtreme Gradient Boosting through district-grouped crossvalidation, alternative imbalance treatments, and an untouched 2023–2024 out-of-time test; temporal ablation, cluster-robust Logistic Regression, SHapley Additive exPlanations, sensitivity analysis, anddirect assessment in East Java were subsequently conducted. The result showed that Logistic Regression achieved the strongest screening-oriented performance, with a recall of 0.690, an F1-score of 0.450, a Receiver Operating Characteristic Area Under the Curve of 0.615, and a Precision–Recall Area Under the Curve of 0.404, while annual-change features improved F1-score and Precision–Recall Area Under the Curve across all three classifiers. However, performance declined in East Java, where two of four deterioration cases were detected, and 22 false-positive warnings were generated. The implication is that parsimonious temporal features provide useful predictive information beyond current regional conditions, although the model is more appropriate for screening and prioritization than for autonomous administrative classification, while operational use requires local calibration, longitudinal data auditing, threshold assessment, and validation across additional provinces and later annual transitions.]]>
