Behavior-driven forecasts of neighborhood-level COVID-19 spread in New York City
Fig 4
One-week ahead retrospective forecasts for neighborhood-level COVID-19 cases.
The behavior-driven forecasts are compared with three baseline models: (B1), a metapopulation model without place category-specific mobility but with seasonal forcing; (B2), the behavior-driven model without seasonal forcing; (B3), the behavior driven model with seasonal forcing but static mobility matrices, crowdedness, and dwell time. We present (A) the relative mean absolute percentage error (MAPE) (MAPEs of the behavior-driven model minus those of the baselines, with blue indicating better forecasts), (B) relative log score (log scores of the behavior-driven model minus those of the baselines, with blue indicating better forecasts), and (C) relative weighted interval score (WIS) (the ratio of the WIS scores of the behavior-driven model to those of the baselines, with blue indicating better forecasts) for all 42 neighborhoods from June 8, 2020 to December 13, 2020.