Fig. 4: Prediction envelopes and comparative forecast performance.
From: Forecasting influenza activity using machine-learned mobility map
The figure shows the comparative performance of the five models in predicting emergency department visits in the five boroughs of New York City. a, b show the calibration envelope (light shaded) and 90% prediction envelope (dark shaded) at weeks 2 and 6 of 2017, respectively. These weeks were chosen to be roughly two weeks pre-peak and two weeks post-peak across boroughs. The solid lines show the ground truth, and colors throughout are representative of boroughs. The plots are arranged horizontally by boroughs and vertically by the mobility network. c MAPE performance. One-to-four-week-ahead median predictions at different data horizons are used to evaluate the mean absolute percentage error (MAPE), for every other week of the flu season. Note that the lower the MAPE, the better the network. The overall MAPE is provided in the legend shows that AMM, COMMUTE, and RADIATION models perform similarly, and do better than the GRAVITY or NO-MOBILITY baselines.
