Predictive Forecasting of Care Load & Placement Demand

Operational planning for the U.S. Unaccompanied Alien Children (UAC) Pro gram requires anticipating changes in the number of children in the care system, rather than relying exclusively on retrospective dashboards. This project develops an exploratory forecasting pipeline for the daily HHS care population and a prelim inary baseline for discharges, using a supplied extract of program records. After removing fully blank rows, the dataset contains 720 dated observations from 12 January 2023 through 21 December 2025. The observed dates span 1,075 calendar days, leaving 355 calendar dates without records; Saturday observations are absent. The primary target is the number of children in HHS care, treated as a stock variable. Discharges and transfers from CBP to HHS are treated as reported flow measures. Six forecasting approaches are compared: naïve persistence, a seven-day moving-average baseline, additive damped-trend exponential smoothing with weekly seasonality, SARIMA, random forest regression, and gradient boosting regression. Evaluation uses seven rolling forecast origins, each forecasting 30 calendar days, and scores only dates with observed HHS care-load values. SARIMA achieves the lowest mean absolute error (MAE) in all three lead-time bands: 18.47 children for days 1–7, 42.95 for days 8–14, and 103.73 for days 15–30. The persistence baseline MAEs are 31.70, 62.59, and 119.85, respectively. A final 30-day SARIMA point forecast from the historical data cutoff begins at 2,485 children and reaches approximately 2,557 by 20 January 2026. Because these forecast dates have passed, they are reported only as historical experimental output, not as current estimates. Discharge demand is summarized using a trailing observed-record baseline rather than presented as a validated daily placement forecast. The study highlights how reporting gaps, interpolated stock values, limited rolling origins, uncertain weekly seasonality, and the absence of approved capacity thresholds constrain operational interpretation. The pipeline is therefore a reproducible analytical prototype, not a validated capacity-management system