Abstract
Tracking the COVID-19 pandemic has been a major challenge for policy makers. Although several efforts are ongoing for accurate forecasting of cases, deaths, and hospitalization at various resolutions, few have been attempted for college campuses despite their potential to become COVID-19 hot-spots. In this paper, we present a real-time effort towards weekly forecasting of campus-level cases during the fall semester for four universities in Virginia, United States. We discuss the challenges related to data curation. A causal model is employed for forecasting with one free time-varying parameter, calibrated against case data. The model is then run forward in time to obtain multiple forecasts. We retrospectively evaluate the performance and, while forecast quality suffers during the campus reopening phase, the model makes reasonable forecasts as the fall semester progresses. We provide sensitivity analysis for the several model parameters. In addition, the forecasts are provided weekly to various state and local agencies.
Introduction
The COVID-19 pandemic necessitated rapid policy decisions at all levels of government and institutional management. College campuses presented a unique challenge: they are high-density environments characterized by communal living and a highly mobile student population, making them potential "hotspots" for rapid viral transmission. While national and state-level forecasting became commonplace, university-level modeling was hindered by data fragmentation and the extreme volatility of the campus reopening phase.
Methodology
To address these challenges, we developed a forecasting framework centered on a Susceptible-Exposed-Infectious-Recovered (SEIR) compartmental model. Unlike purely statistical models, the SEIR approach is mechanistic, simulating the biological spread of the disease.
Model Configuration and Scenarios
A critical component of our model is the time-varying transmissibility parameter ($\beta$), which we calibrate daily against observed case data using the Broyden-Fletcher-Goldfarb-Shanno (BFGS) optimization algorithm. Given the uncertainty in human behavior, we generate three distinct scenarios for each weekly forecast:
- Stable (1.0): Assumes contact rates remain at the 7-day average.
- Best-Case (0.5): Assumes a 50% reduction in contact rates due to mitigation efforts.
- Worst-Case (2.0): Assumes a doubling of contact rates, simulating "COVID fatigue" or large social gatherings.
Data Curation
We focused on four major public universities in Virginia: the University of Virginia (UVA), Virginia Tech (VT), Virginia Commonwealth University (VCU), and James Madison University (JMU). Data was manually curated from public dashboards, which presented significant hurdles due to inconsistent reporting frequencies and retroactive data "restatements." To mitigate these biases, we utilized a rolling 7-day average for model calibration.
Results and Discussion
Our retrospective evaluation reveals that the model's performance was most challenged during the initial "reopening phase" of the semester. The sudden influx of students from diverse geographic regions created a spike in cases that lacked historical grounding for the model.
However, as the semester progressed and the model "learned" the specific dynamics of each campus, forecast accuracy improved significantly. The combined three-scenario approach captured the ground truth case counts between 60% and 90% of the time for one-week-ahead forecasts.
The value of this framework lies in providing administrators with actionable bounds rather than single point estimates. Knowing the "worst-case" scenario allows for proactive resource allocation, such as managing quarantine bed capacity. Furthermore, the mechanistic nature of the SEIR model allows for "what-if" analyses regarding testing frequency and social distancing, which statistical models like ARIMA cannot provide.
Conclusion
This real-time forecasting effort demonstrates the utility of simple compartmental models in informing institutional policy during a public health crisis. While manual data collection remains a bottleneck, the ability to provide weekly, scenario-based insights proved vital for university and state-level decision-makers. Future work will focus on automating data ingestion and scaling the framework to a national level.
Summary updated