Working with a team of leading data scientists and epidemiologists, the State has used several dynamic epidemiology and growth models to predict the spread of the virus and the need for hospital beds, intensive care units, ventilators, and personal protective equipment at a hospital, county, and statewide level.

The models took into account New Jersey's aggressive social distancing policy along with what happened in other countries and tailored those findings to the variations observed at a county level and even hospital level in our state. By using a combination of models, the State has been able to achieve accurate predictions for 10-day forecasts, and has also established long-term forecasts.

More specifically, to account for the aggressive social distancing policy in New Jersey, the State developed multiple adaptive growth models using Multivariate Hawkes processes and control theory principles for short-term forecasting -- these models were developed by researchers at New York University and Facebook AI research. For long-term forecasting, the State has used multiple models including dynamic versions of SIR models and SEIR models, and social distancing growth variation modeling using transfer learning algorithms developed by researchers at New York University. The State has also continuously monitored the University of Washington's Institute for Health Metrics and Evaluation (IHME) model as well as the COVID-19 Hospital Impact Model for Epidemics (CHIME), developed by University of Pennsylvania's School of Medicine's Predictive Healthcare Team.

Short-Term Forecasting

• Adaptive Growth Model with Cross-County Clustering: In this model, developed by Lakshmi Subramanian of the NYU Courant Institute and Srikanth Jagabathula of the NYU Stern School of Business, the State directly estimated the growth rate of new confirmed cases at a county level using a learning rate equation commonly used in control theory and gradient descent algorithms. We used exponential smoothing of the parameter inputs to account for county-level spikes, reporting, and testing fluctuations. The State used cross-county clustering to learn early growth rate variations in counties with very limited data and smooth the variations across counties.
• Multivariate Hawkes Process: Another model the State uses for short-term forecasting, developed by Facebook's AI team, trains a multivariate Hawkes process (which is commonly used in financial stock price modeling) to train a county-level model and a collective state-level model.

Both these models have been re-trained on a daily basis and provided 7-14 day forecasts.

Long-Term Forecasting

• Dynamic SIR: The SIR (susceptible-infected-recovered) model is widely used for epidemiology modeling. The CHIME (COVID-19 Hospital Impact Model for Epidemics) tool uses the SIR model, but manually sets the parameters of the SIR model based on doubling time and social distancing. While the CHIME model does leverage a social distancing parameter, the State has also used a dynamic SIR model that adapted the mixing rate (which captures the number of people each infected person comes in contact with) parameter based on the data observed in growth rate variations at a county level to provide an additional long-term model that further captures the impact of the social distancing orders currently in place in New Jersey.
• Dynamic SEIR model: Expanding upon our Dynamic SIR model, the State trained the parameters of an SEIR (susceptible-exposed-infected-recovered) model in a dynamic manner at a county level based on the growth variations observed in the data at a county and state level. The dynamic SEIR model - developed by Subramanian and Jagabathula - used the growth variations to adaptively adjust the SEIR parameters, in essence, varying the R0 value of the epidemiology model over time to capture the impact of the social distancing orders currently in place in New Jersey.
• Social Distancing Growth Variation Modeling using Transfer Learning: Traditional epidemiology models do not capture movement limitation behavior. To account for that, in this model - also developed by Subramanian and Jagabathula - the State studied the growth-rate variations in other countries after social distancing orders have been issued and use transfer learning to determine the appropriate growth variations that best fit the data for New Jersey. This model was especially useful for learning best case, expected case, and worst-case scenarios of peaks.

With the help of this ensemble of forecasts, the State has been better able to make difficult decisions on what critical resources to acquire, where to deploy these critical resources, anticipate potential hot spots, and ensure that those who need care can get it when they need it most.