PREDICTIVE PUBLIC ADMINISTRATION: USING MACHINE LEARNING TO ANTICIPATE SOCIAL SERVICE DEMAND

Social services are typically funded by governments in a back-end way, causing delays and unevenness. Predictive analytics can help transform administration to forward-planning by predicting demand across health, housing, and jobseeker benefit areas. We used a quantitative predictive modelling from secondary administrative and socio-economic data (demographics, income, unemployment, urban density, and seasonal indicators). After cleaning data, multiple imputation, and normalization (in the case of a neural network), we divided the data into training/testing sets (70/30) with 10-fold cross-validation. We compared two models: Random Forest (RF) with grid-searched hyperparameters and Artificial Neural Network (ANN) (MLP, ReLU, Adam optimizer, early stopping). We used MAE, RMSE, and R² as the performance metrics; fairness and transparency tests employed feature importance (RF) and distributional error scans across population subgroups. Both ML methods significantly outperformed a linear regression baseline, minimizing errors by 25–35%. For healthcare needs, ANN performed best (R² = 0.91; RMSE = 5.4), both picking up non-linear and seasonality effects. For housing subsidies, RF performed best (R² = 0.87; MAE = 3.7), where unemployment rate and median household income were the top predictors. For jobless benefits, ANN slightly outperformed RF (R² = 0.86 vs. 0.84; RMSE = 5.7 vs. 6.2). Error audits revealed no systematic decrease in principal subgroups following mitigation. ML can substantially improve public administration foresight. ANN is better suited to dynamic, multifaceted demand (e.g., healthcare), while RF excels at high accuracy with interpretability for policy defence (e.g., housing). A hybrid deployment ANN for forecasting accuracy and RF for explanation is available to allow proactive, fair resource allocation with governability. Governance results.