Applying Random Forest to Predict Unemployment in Medan City for Adaptive Labor Policy Formulation

Abstract

Urban unemployment is a multidimensional policy problem because its dynamics are associated not only with economic growth but also with labor-force participation, human capital, poverty, inequality, inflation, entrepreneurship, and the absorptive capacity of the local economy. This study evaluates the preliminary feasibility of a Random Forest Regressor for forecasting unemployment in Medan City as a foundation for an adaptive labor-policy decision-support system. The experiment employs a structured simulated dataset informed by indicators contained in the Medan City Medium-Term Development Plan for 2025–2029. It comprises 20 observations, four subdistricts, the 2020–2024 period, 12 predictors, and one target variable. The dataset is neither official labor-force microdata nor an official subdistrict-level unemployment estimate; accordingly, the study is framed as an exploratory methodological proof of concept. Observations from 2020–2023 were used for training, while 2024 was retained as a temporal test set. The analytical pipeline incorporated data-quality auditing, one-hot encoding, a baseline Random Forest, year-grouped cross-validation, and randomized hyperparameter search over 30 candidate configurations. The optimized model used 100 trees, a maximum depth of 3, a minimum split size of 3, max_features of 0.7, and no bootstrap sampling. Optimization reduced MAE from 0.4587 to 0.3989, RMSE from 0.4821 to 0.4184, and MAPE from 6.9481% to 6.0382%. Nevertheless, the test R² remained negative at -1.4892, indicating inadequate generalization. Labor-force participation and economic growth accounted for a combined importance of 83.99%, but these results must be interpreted as predictive associations rather than causal effects. Policy deployment therefore requires a substantially larger, official, granular, and externally validated longitudinal dataset in github.com.