Predictive Forecasting of Care Load & Placement Demand

Managing the care of children who arrive in a country without their parents is a task that demands precision, foresight, and a deep understanding of human movement. When thousands of young people cross borders and enter government custody, officials must know how many beds, staff members, and resources will be needed tomorrow, next week, or next month. This is not merely a matter of counting heads; it is a complex puzzle of timing and flow. The challenge lies in predicting how many children will be in care at any given moment, a number that shifts daily based on new arrivals, transfers to other facilities, and departures. To solve this, researchers turn to the science of time-series forecasting, a method that looks at past patterns to anticipate future trends. By studying the rhythm of historical data—how numbers rise and fall over time—analysts can build models that act like a weather report for population needs, offering a glimpse into the future to help organizations prepare.

In a recent study, researchers at Chaitanya Bharathi Institute of Technology in Hyderabad tackled this specific problem using data from the United States Department of Health and Human Services. They focused on the "Unaccompanied Alien Children" program, which tracks the daily numbers of minors in government care. The team gathered a dataset containing over a thousand daily records, stretching from early 2023 through the end of 2025. After cleaning the data to remove gaps and inconsistencies, they were left with 720 reliable daily counts of children in care. Their goal was to build a system that could look at this history and predict the next thirty days of demand with high accuracy.

To make these predictions, the researchers did not rely on a single method. Instead, they built a pipeline that combined traditional statistical techniques with modern machine learning. They taught the computer to recognize patterns by feeding it specific clues derived from the past. These clues included how many children were in care one day ago, one week ago, and two weeks ago, as well as the average numbers over the last week and two weeks. They also added measures of how much the numbers were swinging up and down in the short term, and a calculation of the "pressure" on the system, which compared how many children were being transferred into care against how many were being discharged. By including the day of the week and the month, the system could also account for regular weekly or seasonal shifts.

The team tested several different approaches to see which one worked best. They started with simple methods, such as assuming tomorrow would look exactly like today, and moved on to more complex statistical models that account for trends. Finally, they trained powerful machine learning tools known as Random Forests and Gradient Boosting. These tools work by building many decision trees that vote on the answer, allowing them to spot complicated, non-linear relationships in the data that simpler methods might miss. The researchers split their data carefully, using the earlier records to train the models and saving the final thirty days of data to test how well the models performed on unseen information. This approach ensured that the evaluation was fair and that the models were not simply memorizing the past.

The results offered a clear picture of what the future might hold for the next month. The system predicted a gradual, steady increase in the number of children in care, rising from approximately 2,274 at the start of the forecast period to about 2,324 by the end. The average prediction for this thirty-day window was roughly 2,299 children. When the researchers checked how close these predictions were to the actual numbers in the test period, they found the average error was about 15 children per day. In percentage terms, this error rate was very small, standing at roughly 0.64 percent. This level of accuracy suggests the models are capturing the underlying dynamics of the system well enough to be useful for planning.

However, the study also revealed an important nuance in how these tools are used. While the researchers saved a Random Forest model as their final tool for deployment, the specific error numbers they reported were calculated after running a different model, the Gradient Boosting one. This means the saved tool and the reported performance metrics do not perfectly match in the current version of the project. The authors acknowledge this and note that a future version of the work should ensure the best-performing model is the one that is saved and used. Despite this technical detail, the project successfully demonstrated that a combination of historical data, operational clues, and machine learning can produce a reliable forecast.

The researchers also noted that the data showed a significant shift in early 2025, where the number of children in care dropped sharply from over 6,000 in January to around 2,000 by April. This kind of sudden change, or "regime shift," is difficult to predict, and the study highlights the need for models that can adapt to such structural breaks. The final system was packaged into a user-friendly dashboard, allowing officials to view the data, see the forecast on a chart, and download the predictions for their own use. By turning raw numbers into a clear, thirty-day outlook, this work provides a practical tool for capacity planning, helping organizations manage staffing, shelter space, and transportation with greater confidence. The study concludes that while the current model is a strong starting point, future improvements will focus on comparing all models more rigorously and adding tools to measure the uncertainty of the predictions, ensuring that decision-makers have the most robust information possible.